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.0000357 | Taste, A New Incentive to Switch to (R)-Praziquantel in Schistosomiasis Treatment | Praziquantel (PZQ) is the drug compound of choice in the control and treatment of schistosomiasis. PZQ is administered as a racemate, i. e. 1∶1 mixture of enantiomers. The schistosomicidal activity arises from one PZQ-enantiomer, whereas the other enantiomer does not contribute to the activity. The WHO's Special Programme for Research and Training in Tropical Diseases (TDR) has assigned the low-cost preparation of pure schistosomicidal (−)-PZQ a key priority for future R&D on PZQ, but so far this transition has not happened. PZQ has two major administration drawbacks, the first being the high dose needed, and its well documented bitter and disgusting taste. Attempts of taste-masking by low-cost means have not been successful. We hypothesized that the non-schistosomicidal component in PZQ would be the main contributor to the unpleasant taste of the drug. If the hypothesis was confirmed, the two major administration drawbacks of PZQ, the high dose needed and its bitter taste, could be addressed in one go by removing the component contributing to the bitter taste.
PZQ was separated into its schistosomicidal and the non-schistosomicidal component, the absolute stereochemical configuration of (−)-PZQ was determined to be (R)-PZQ by X-ray crystallography, and the extent of bitterness was determined for regular racemic PZQ and the schistosomicidal component in a taste study in humans. Finding: The schistosomicidal component alone is significantly less bitter than regular, racemic PZQ.
Our hypothesis is confirmed. We propose to use only the pure schistosomicidal component of PZQ, offering the advantage of halving the dose and expectedly improving the compliance due to the removal of the bitter taste. Therefore, (R)-PZQ should be specifically suitable for the treatment of school-age children against schistosomiasis. With this finding, we would like to offer an additional incentive to the TDR's recommendation to switch to the pure schistosomicidal (R)-PZQ.
| Schistosomiasis, or Bilharzia, is a parasitic disease caused by flatworms, which affects about 200 million people worldwide. Praziquantel (PZQ) is the drug compound of choice in the control and treatment of this disease. Only half of the drug dose currently administered actually has activity against schistosomiasis, whereas the other half has no activity. Therefore, the WHO has assigned the low-cost preparation of the pure active component a key priority for future PZQ research and development. PZQ has two major administration drawbacks, the first being the high dose needed, the second its well documented bitter taste. Attempts of masking the unpleasant taste have not been successful. We hypothesized that the non-active component in PZQ would be the main contributor to the unpleasant taste of the drug. We determined the extent of bitterness for regular PZQ compared to the pure active component in a taste study in humans. We found that the pure active component alone is significantly less bitter than regular PZQ. This new finding should serve as an additional incentive for the PZQ research and development community to provide a low-cost, large-scale preparation route to the pure active component of PZQ.
| Praziquantel [1] (PZQ) is the drug compound of choice in the control and treatment of schistosomiasis [2], in fact, it is the only commercially readily available drug. So far, no backup compound for PZQ of comparable efficacy and breadth of application is available. Clinically relevant resistance has not been observed, however differences in responses of PZQ-resistant and -susceptible Schistosoma mansoni to PZQ in vitro have been described [3]. PZQ is included in the WHO Model List of Essential Drugs [4] and is at the core of numerous schistosomiasis control programmes. The WHO's strategy for schistosomiasis control [5] aims at reducing morbidity through treatment with PZQ, with a focus on periodic treatment of school-age children and adults considered to be at risk. School-age children are seen as a high-risk group for schistosome infections because they are more susceptible to infection in cases where their increased nutritional needs are not adequately met, might be compromised by helminth infections in their cognitive development, and are continuously exposed to contaminated soil and water but probably less aware of the need for good personal hygiene [6].
While the safety and efficacy against all schistosoma species are outstanding, PZQ has two major administration drawbacks, the first being the high dose needed, 40 mg PZQ/kg bodyweight: Dosages in children are determined by measurement of children's heights using tablet poles, and range from one to five 600 mg-tablets for one treatment. Especially young children have been reported not to be able to swallow these 600 mg tablets [7]. The second drawback is PZQ's well documented bitter and disgusting taste, which can lead to gagging or vomiting if tablets are chewed contrary to recommendation [8]. In veterinary medicine, the oral delivery of PZQ to taste-sensitive companion animals like cats is known to be a challenge. Traditional methods of taste-masking, like the addition of aromas or sugar, are ineffective for PZQ. The bitterness of PZQ even led to PZQ's use as a bitter model drug compound in the effectiveness testing of sophisticated and expensive taste-masking techniques like micro-encapsulation [9] or drug active coating [10]. Apart from anecdotal evidence [2], we are not aware of reports of low compliance among children treated within schistosoma programmes due to the bitter taste. However, we have to assume that the unpleasant taste of PZQ does not lead to a treatment situation which school-age children would enjoy.
PZQ is administered as a racemate, i. e. 1∶1, mixture of two compounds of identical constitution but non-superimposable mirror-image configuration, so called enantiomers. The straightforward and low-cost chemical synthesis has to be assumed as the reason for the use of the racemate, although it has been known for years that the schistosomicidal activity mainly relies in one PZQ-enantiomer, designated (−)-PZQ (alternatively termed levo-PZQ, l-PZQ, sometimes L-PZQ), whereas the other enantiomer, designated (+)-PZQ (alternatively termed dextro-PZQ, d-PZQ), does not contribute to the activity [11]–[13] (Figure 1). From this perspective, only half of the drug compound administered is in fact the drug active, whereas the other half must be considered molecular ballast, which has to be metabolized and excreted while not contributing to the schistosomicidal activity. To the best of our knowledge, no clinical studies in humans exist if and how non-schistosomicidal (+)-PZQ alone contributes to the side effects known of racemic PZQ, but this may be assumed: Upon incubation of PZQ and both enantiomers with isolated rat hepatocytes, additional metabolites were detected resulting from the non-contributing (+)-PZQ [14]. Various methods of producing the pure schistosomicidal component (−)-PZQ exist, which are considerably more expensive than the production of racemic PZQ itself. So far, the potential alone to administer half the current dose by replacing racemic PZQ by (−)-PZQ did not lead to a production process for (−)-PZQ comparable in costs for racemic PZQ.
In the context of the WHO's Global Plan to combat NTDs [15], the Special Programme for Research and Training in Tropical Diseases (TDR) set up an incentive for further R&D work by emphasizing the low-cost preparation of pure schistosomicidal (−)-PZQ (see also the schistosomiasis research collaborative community within The Synaptic Leap [16]) as a key priority for future R&D on PZQ [17]. Three pharmacological goals for the development were stated: (1) same dose of (−)-PZQ as currently in regular, racemic PZQ, with smaller tablet size and less frequent/severe adverse events, (2) higher dose of (−)-PZQ with similar tablet size and possibly similar adverse event profile as current treatment which could reduce the probability of or delay the development of resistance, or (3) a combination of these two objectives. As we already mentioned, a smaller tablet size would be more suitable for the treatment of children. Taking into account that the WHO's strategy specifically aims at school-age children, we were intrigued by the question whether the taste disadvantage of PZQ could be turned into an additional incentive to introduce (−)-PZQ against schistosomiasis as the drug active of choice. Background to our consideration was the well-documented fact that in most cases taste experiences depend on the stereochemical configuration of the agent [18], i. e. the taste buds react enantioselectively–like all natural receptors which are composed of chiral constituents like L-amino acids. We hypothesized that (−)-PZQ and (+)-PZQ would contribute to the bitter taste to a different extent, and that the non-schistosomicidal (+)-PZQ would be the main or sole contributor to the disgusting taste. Surprisingly, no public knowledge exists on the tastes of the two enantiomers. We prepared schistosomicidal (−)-PZQ, assigned the stereochemical configuration by X-ray crystallography, and determined the extent of bitterness for regular racemic PZQ versus the schistosomicidal component (−)-PZQ in a taste study in humans. We chose this comparison over the comparison of (−)-PZQ to non-schistosomicidal (+)-PZQ because the latter alone does not have any role in a treatment situation. Also the pharmacological studies by others had compared racemic PZQ to (−)-PZQ, and not (+)-PZQ to (−)-PZQ [19].
Although effective synthetic methods for the enantioselective preparation of PZQ have been reported [20], we opted for the direct enantioseparation of the racemate yielding gram quantities of both optical forms. The preparative scale chromatography was performed on microcrystalline cellulose triacetate using methanol as the mobile phase, conditions under which the enantiomer having the negative optical rotation emerged first from the column [21]. After crystallisation from methanol/water, (−)-PZQ was obtained in enantiomeric excess >99%, as determined by HPLC (column used Chiralcel OD-H). No residual other enantiomer (+)-PZQ was detected in this sample. X-ray structural analysis, using Cu-Kα radiation, of a monoclinic crystal in hemi-hydrate form obtained from said fraction by crystallization from methanol/water unequivocally proved the R-configuration of the molecule by measuring Friedel pairs and the Flack parameter (x = −0.1(3)) (Figure 2). Further details of the crystal structure analysis are available on request from the CCDC (www.ccdc.cam.ac.uk) quoting the names of the authors and journal citation.
The bitterness values of racemic PZQ and its schistosomicidal component (R)-PZQ were determined according to the European Pharmacopoeia [22] by comparison with quinine hydrochloride, the bitterness value of which is set at 2×105. The bitterness value is defined by the European Pharmacopoeia as the reciprocal of the concentration of a solution in a dilution series of a compound, a liquid or an extract that still has a bitter taste. Concentrations of solutions used in the tests ranged from 1.69×10−8 to 1.0×10−4 g/mL. A test panel consisting of sixteen members was assembled. Although children comprise the treatment target group no children were included in the test panel. All panel members were adults completely untrained in performing sensory tests. To correct for individual differences in tasting bitterness amongst the panel members a correction factor was determined for each panel member by preparing dilutions of quinine hydrochloride. The mouth was rinsed with water before tasting. The dilution with the lowest concentration having a bitter taste was determined by taking 10 mL of the weakest solution into the mouth and passing it from side to side over the back of the tongue for 30 seconds. If the solution was not found to be bitter, the panellist had to spit out and wait for one minute before the mouth was rinsed again with water. After 10 minutes, the next dilution in order of increasing concentration was tasted. The correction factor k for each panel member was calculated according to the European Pharmacopoeia by k = n/5, where n is the number of millilitres of the stock solution in the dilution of the lowest concentration that is judged to be bitter. One panel member detected bitterness already in pure water, and was therefore excluded from the test panel. Dilutions of the test compounds racemic PZQ and (R)-PZQ were prepared and tasted by the remaining fifteen members of the test panel in the same manner as described for quinine hydrochloride. The bitterness value as experienced by each member was calculated according to the European Pharmacopoeia taking the individual-related correction factor into account by Y×k/X×0.1, where Y is the dilution factor of the dilution, and X is the number of millilitres of the respective dilution which, when diluted to 10 mL with water, still has a bitter taste. The bitterness value of the test compounds resulted from calculating the average of the individual values.
Requested statement: Informed written consent was obtained from all panelists to participate in this taste study. As a taste study, and not a medical study in the sense of the WMA Declaration of Helsinki, it did not require approval of an independent review board (highly diluted preparations were tasted and spat out–they were not ingested). Nevertheless, it was conducted according to the principles of the WMA Declaration of Helsinki where applicable.
The results of the determination of bitterness values are shown in Table 1. Remarkable is the variation of the individuals' results as indicated by the relative standard deviation and the dispersion of the results in the box-and-whisker diagram (Figure 3). In contrast to the average, the medians of the results, as depicted in the box-and-whisker diagram, are different from each other. The observed variation was probably provoked by the test panel consisting of untrained members only [23]. Thirteen out of fifteen panel members found (R)-PZQ to taste less bitter than racemic PZQ. Although no statistical test is required or proposed by the European Pharmacopoeia, a statistical test (using SAS software, release 9.1.3, SAS Institute Inc., Cary, NC, USA) was conducted to investigate the observed difference between the compounds. Considering the small sample size and the nature of the data which does not justify the assumption of a normal distribution, a nonparametric, distribution-free method was chosen. On the 5% level of significance, Wilcoxon's Signed Rank Test (two-sided) resulted in a significant difference between the taste of racemic PZQ and (R)-PZQ (p = 0.0107). This result was confirmed by the Sign Test (two-sided, p = 0.0018).
In addition to the quantitative determination of the bitterness values, qualitative taste sensations were noted by the members of the test panel for each compound. For racemic PZQ, all panel members commonly observed the sensation of an unpleasant chemical or metallic taste or a taste circumscribed best by old rubber. On the other hand, for (R)-PZQ the panellists commonly described the sensation of a moderate chemical taste, comparable to that of a polyethylene or a rubber pipe. Although the tastes were not recognized alike across the test panel, for the majority of the test panel we can state that (R)-PZQ had a less unpleasant taste compared to racemic PZQ.
The schistosomicidal component of regular PZQ, (R)-PZQ has a less unpleasant taste compared to racemic PZQ, which was found to be comparably bitter or unpleasant. It can be assumed that the disgusting taste of racemic PZQ stems from the non-schistosomicidal component, (S)-PZQ. Removing the latter from currently used racemic PZQ therefore not only offers the chance to halve the dose, with the potential to decrease the number or size of the tablets, but also addresses the second disadvantage of regular, racemic PZQ-its unpleasant taste. With this finding and its publication we would like to offer an additional incentive to focus work of the PZQ R&D community on further decreasing the cost of production of (R)-PZQ with the goal to switch to pure (R)-PZQ as a replacement for racemic PZQ for the treatment of school-age children against schistosomiasis.
|
10.1371/journal.pbio.0050212 | The p75 Neurotrophin Receptor Is a Central Regulator of Glioma Invasion | The invasive nature of cancers in general, and malignant gliomas in particular, is a major clinical problem rendering tumors incurable by conventional therapies. Using a novel invasive glioma mouse model established by serial in vivo selection, we identified the p75 neurotrophin receptor (p75NTR) as a critical regulator of glioma invasion. Through a series of functional, biochemical, and clinical studies, we found that p75NTR dramatically enhanced migration and invasion of genetically distinct glioma and frequently exhibited robust expression in highly invasive glioblastoma patient specimens. Moreover, we found that p75NTR-mediated invasion was neurotrophin dependent, resulting in the activation of downstream pathways and producing striking cytoskeletal changes of the invading cells. These results provide the first evidence for p75NTR as a major contributor to the highly invasive nature of malignant gliomas and identify a novel therapeutic target.
| Gliomas are highly malignant and invasive tumors with tendrils that extend far from the primary tumor site, rendering conventional therapies ineffective and leading to an invariably poor prognosis. To understand the molecular mechanisms underlying this invasive behavior, we injected immunocompromised mice with human gliomas and compared invasive cells, which left the primary tumor site, to noninvasive cells, which remained at the site of injection. We identified the neurotrophin receptor p75NTR—which normally functions during development to induce neurite outgrowth and promote neuronal cell death—as an important regulator of glioma invasion. We present the first evidence that this neurotrophin receptor can also be a potent mediator of glioma invasion, and we show that the expression of this receptor is sufficient to impart a dramatic invasive behavior on genetically distinct tumors. These data highlight a previously unknown function of this receptor and suggest it may be a novel therapeutic target in the treatment of this devastating cancer.
| Malignant gliomas are diffuse, highly invasive, and often multifocal tumors that have a dismal prognosis, with a median survival of only 1 y and “long-term survivors” (i.e., surviving ≥3 y) are rare [1,2]. A major barrier to effective malignant glioma treatment is their highly invasive nature; they extend tendrils of tumor several centimeters away from the main tumor mass, which render these tumors incurable by local therapies such as surgery or radiotherapy [3]. Ninety-five percent of gliomas recur within 2.5 cm of the resection margin, which contains invasive cells that act as a “disease reservoir” and elude current treatments [4]. Glioma cells do this by becoming distinct from their noninvasive counterparts. Specifically, they activate a number of coordinate cellular programs, which include those necessary for migration (e.g., motility) and invasion (e.g., extracellular matrix degradation) [5] and also a number of pathways (e.g., reduced proliferation, marked resistance to apoptosis) [4,6,7] which render the invasive cells highly resistant to conventional treatments. A detailed understanding of the mechanisms underlying this invasive behavior is essential for the development of effective therapies.
Although in their infancy, attempts to identify genes involved in glioma invasion have used a number of techniques, including the isolation of invasive cells from human cell lines in vitro [6,7], the use of organotypic brain slice cultures [8], and the collection of tumor and invasive cells from frozen glioblastoma patient specimens using laser capture microdissection [9–11]. Although each method has been successful in its own right, none of these models have been ideal or comprehensive for discovering the underlying mechanisms of invasion. New models or alternative strategies are needed. We have therefore undertaken the approach of serial in vivo selection to identify genes important for the invasive behavior of malignant glioma. Similar strategies have been used to effectively identify mechanisms underlying the metastatic behavior of both melanoma and breast tumors [12,13]. Using this approach, we isolated highly invasive glioma cells from a relatively noninvasive human malignant glioma. Gene expression profiles comparing these two tumor cell populations identified the p75 neurotrophin receptor (p75NTR) as an important and potent mediator of invasion in human glioma.
p75NTR is a transmembrane glycoprotein and a member of the tumor necrosis factor (TNF) superfamily that was originally isolated as a nerve growth factor (NGF) receptor, but has since been shown to bind both the mature and precursor forms of the neurotrophin family of ligands (brain-derived neurotrophin factor [BDNF], neurotrophin-3 [NT-3], and neurotrophin-4/5 [NT-4/5]) [14–18]. In neurons, p75NTR is coexpressed with a second group of neurotrophin receptors, the tropomysin receptor kinases (Trks). It has become increasingly clear that the dogma in neuroscience that Trks mediate neuronal survival and p75NTR causes neuronal cell death is too narrow a view [19–22]. Rather, there is a growing appreciation that p75NTR, like other members in the TNF superfamily, mediates a very broad range of cellular functions, depending on the cell context and the repertoire of co-receptors that exist (e.g., Trks [23], Nogo receptor [24], and sortilin [25]). In neurons, p75NTR has been shown to increase [26,27] or inhibit [28] axon growth, reduce [29] or promote [30,31] neuronal cell death, and is either necessary [32] or not required [33] for inhibition of neuronal regeneration. These apparent discrepancies are not confined to neurons; p75NTR has also been shown to both inhibit [34] and promote [26,35] Schwann cell migration during development. Even though p75NTR does not contain any catalytic activity, it interacts with several proteins that help transmit signals required for its various functions. Neurotrophin engagement of p75NTR controls the activity of the small GTPase RhoA, providing a direct link from the receptor to modulating cellular architecture. As is the case for phenotypic responses, RhoA has been shown to be activated or inhibited depending on cellular context [27,28,34,36–38]. Reports have hinted at roles for p75NTR in growth [39] and apoptosis [40] of glioma cells; however, data presented here support a much different role for p75NTR—that of mediating glioma cell invasion.
One of the problems in xenotransplanting human glioma cells into the brains of immunocompromised mice is that the resulting tumors are circumscribed, with very little cell infiltration into the brain parenchyma [41]. To generate an orthotopic model that more closely mimics the human disease and allows for the identification of molecular determinants of glioma invasion in a global and unbiased manner, we used an in vivo–selection procedure to select for highly invasive human glioma cells (Figure 1A). We isolated highly invasive glioma cells from the noninvasive human malignant glioma cell line U87 expressing green fluorescent protein (GFP) (U87GFP) and a neomycin resistance gene. Expression of these genes afforded us the ability to isolate the rare glioma cell that migrated away from the primary tumor site. These “invasive” cells were grown and expanded in tissue culture, and reintroduced into the brains of immunocompromised mice where they formed highly infiltrative tumors with poorly defined edges (Figure 1B). These extremely invasive cells were found vast distances from the main tumor mass, with GFP-positive tumor cells readily identifiable in the contralateral hemisphere. In clear contrast, reimplantation of the noninvasive “tumor” cells led to the formation of large tumors with sharply demarcated edges (Figure 1B). Using this model, we identified gene expression differences between the noninvasive and highly invasive in vivo–selected glioma cells. RNA extracted from tumor and invasive populations was used to prepare labeled cDNA that was hybridized to 14,000-gene human oligonucleotide microarrays (produced by the Southern Alberta Microarray Facility, University of Calgary). Genes up- or down-regulated in the invasive population were compared to the tumor population, and genes that showed consistent gene expression changes of 2-fold or greater are outlined in Figure 2A. To ensure the integrity of the microarray data, we chose seven arbitrary genes for validation, the expression of five of which are shown in Figure 2B and 2C. Semiquantitative real-time polymerase chain reaction (RT-PCR) confirmed the expression of all seven genes, including granulocyte colony-stimulating factor (G-CSF), interleukin-8 (IL-8), DZFKp434B204 (unknown hypothetical protein), tissue inhibitor of metalloproteinases-3 (TIMP-3), and p75NTR (Figure 2B and 2C). The semiquantitative RT-PCR indicates that our microarray data is an under-representation of the fold changes in RNA expression. Based on the reproducibility of the data, previous implication in tumorigenesis in other cancers (e.g., melanoma and prostate) [42–47], and the novelty of the finding in brain tumors, we chose p75NTR for further study. Importantly, we confirmed the up-regulation of p75NTR was not only at the mRNA level, but that a dramatic alteration in abundance of p75NTR was seen in the invading cells (Figure 2C). A number of invasive lines were generated by serial in vivo selection and microarray analysis using a second independent U87 invasive line validated the presence of p75NTR by microarray that was confirmed by RT-PCR and Western blot (unpublished data). In addition, using the in vivo–selection paradigm outlined in Figure 1, we isolated both tumor and invasive cells from a second human glioma cell line, U251N. These in vivo–selected invasive U251N cells also expressed high levels of endogenous p75NTR (Figure S1).
Although p75NTR and its ligands, the neurotrophins, are expressed throughout the nervous system, particularly during development, a role for p75NTR in central nervous system tumors has not been described previously. We therefore assessed whether the up-regulation of p75NTR found in the invasive glioma cells had a functional consequence (i.e., increased their migration and invasion). The noninvasive and highly invasive cells were treated with the p75NTR ligand NGF, and migration and invasion were measured. The addition of NGF to invasive cells significantly increased the number of cells able to invade through matrigel, but had no effect on the invasive ability of the tumor cells (which had no detectable p75NTR; Figure 2D). Because neurotrophins are also ligands for the Trk receptors, RT-PCR and immunoprecipitation experiments were performed. No detectable mRNA or protein for the Trk receptors was found in the invading glioma cells (unpublished data). In addition, we tested the effect of the unprocessed or proform of NGF (pro-NGF), a high-affinity ligand for p75NTR [33,39] that is unable to activate Trk [16]. Accordingly, treatment of the invasive cells with cleavage-resistant pro-NGF enhanced their migration at concentrations as low as 1 ng/ml while having no effect on the tumor cells (Figure S2). Although we found that neurotrophin could enhance invasion of the p75NTR-positive invasive cells (Figure 2D), we also observed a significant increase in the absence of ligand. Signals from p75NTR can arise both in the absence and presence of ligand; however, these signals often evoke opposing biological responses. Because the outcome of both neurotrophin-dependent and neurotrophin-independent signaling was the same, we considered the possibility that the glioma cells were producing and secreting neurotrophin(s), thus activating an autocrine loop. We assessed the expression of several neurotrophins and found that BDNF was present in both the conditioned media and the cell lysate of all glioma cells tested (unpublished data). Furthermore, we found that the presence of p75NTR shifted the localization of BDNF from the conditioned media to the cell membrane (Figure S3), supporting the notion of autocrine/paracrine activation of the p75NTR receptor.
To directly test the hypothesis that elevated expression of p75NTR is necessary for neurotrophin-induced glioma migration and invasion, we surveyed a panel of human glioma cell lines for p75NTR protein expression. We found that the human glioma cell line SF767 endogenously expressed high levels of p75NTR, as detected by Western blot (Figure 3A) and fluorescence-activated cell-sorting (FACS) analysis (unpublished data). Using RNA interference (RNAi), we down-regulated p75NTR in the SF767 cell line using an expression vector containing a p75NTR-specific small interfering ribonucleic acid (siRNA) and confirmed the down-regulation by RT-PCR and Western blot (Figure 3A). A random, nonspecific siRNA sequence was used as a control. Down-regulation of p75NTR levels in SF767 was sufficient to reduce its migration in vitro and rendered the cells nonresponsive to addition of NGF in both migration and invasion assays (Figure 3B and 3C). Similarly, down-regulation of p75NTR by siRNA in the original in vivo–selected U87 invasive cells significantly blocked migration and invasion (Figure S4).
Since down-regulation of p75NTR in SF767 cells and U87 in vivo–selected invasive cells inhibited glioma invasion, we assessed whether ectopic expression of p75NTR alone was sufficient to increase glioma migration and invasion in a cell line without detectable p75NTR (the original U87 cell line). To this end, we stably transfected the full-length cDNA of human p75NTR into the U87 glioma cell line, using stable transfection of the empty pcDNA vector as a control. Expression levels of p75NTR in these cells were confirmed by RT-PCR and Western blot (Figure 3D). Expression of p75NTR caused a significant increase in migration and invasion in vitro (Figure 3E and 3F). Treatment of these cells with NGF had no further enhancement on their migration or invasion consistent with the idea that when p75NTR is expressed, an autocrine loop is completed, leading to enhanced migration and invasion.
Malignant gliomas clinically show extensive infiltration away from the main tumor and into the surrounding normal brain tissue. To determine whether the expression of p75NTR was important for glioma cell invasion in vivo, we implanted the U87 human glioma cell line ectopically expressing p75NTR into the brains of severe combined immunodeficiency (SCID) mice. U87 cells stably transfected with the empty pcDNA vector were implanted for comparison as a control. Twenty-eight days after implantation, the mice were sacrificed and the brains prepared for immunohistochemical staining using antibodies directed against human nuclei and p75NTR. Implantation of U87 glioma cells stably transfected with pcDNA led to the formation of well-circumscribed tumors that were p75NTR negative (Figure 4A). In sharp contrast, implantation of U87 glioma cells stably expressing p75NTR resulted in the formation of tumors with highly infiltrative edges (Figure 4B). Isolated p75NTR-positive human glioma cells could be detected in regions vastly distant from the main tumor mass (Figure S5).
Because malignant gliomas are an extremely heterogeneous group of tumors and the in vivo–selected U251N cells also expressed high levels of p75 (Figure S1), we determined whether the sole expression of p75NTR was sufficient to impart an invasive phenotype, not only on the U87 cells, but also on the genetically distinct U251N cells. U251N cells ectopically expressing p75NTR (U251Np75), along with empty vector–transfected cells as a control (U251NpcDNA), were implanted into the brains of SCID mice as described above. As we have observed previously, the U251N control cell line (U251NpcDNA) was inherently more invasive than U87pcDNA in vivo, with finger-like projections extending from the main tumor mass into the surrounding normal brain (compare Figure 4A and 4C). Nevertheless, ectopic expression of p75NTR (U251Np75) dramatically enhanced the inherent invasive ability, with p75NTR-positive cells being found at locations distinct from the main tumor mass (compare Figure 4C and 4D). Thus, up-regulation of p75NTR is sufficient to allow glioma cells of diverse genetic backgrounds to invade into the surrounding normal brain. Because p75NTR can have effects on several physiological responses, we also evaluated the effect of p75NTR expression on cell cycle, proliferation, and survival, and observed no significant change (unpublished data).
In order to test whether neurotrophin was important in the invasive behavior of these cells, we constructed two p75NTR mutants, p75CRD105 and p75CRD130, containing a four–amino acid insertion in the cysteine-rich domain (CRD) following amino acids 105 and 130 (CRD 105 and CRD 130), respectively. Insertions at these locations disrupt the normal spacing of the cysteine residues within the164 ligand-binding domain and create p75NTR proteins that are unable to bind to mature neurotrophin [48]. These constructs were stably transfected into U87 glioma cells, and cell surface expression for the mutant p75NTR proteins was confirmed by FACS analysis (Figure 5A). To verify that the mutant p75NTR do not bind neurotrophin, BDNF expression in the conditioned medium and total cell lysates of U87 cells expressing CRD105 and CRD130 were performed. Unlike the wild-type p75NTR-expressing glioma cells in which expression of p75NTR causes a shift in BDNF localization from the medium to the cell lysate, cells expressing the mutant alleles (CRD105 and CRD130) did not result in a shift of BDNF localization, confirming that these mutants do not bind endogenous BDNF (Figure 5B). These cells were implanted into the brains of SCID mice and allowed to grow for 21 d. The mice were sacrificed, and frozen brain sections were stained with antibodies against human nuclei (Figure 5C; brown color, top row) and human p75NTR (Figure 5C; brown color, bottom row). Disruption of the neurotrophin binding capacity of p75NTR results in tumors with well-defined borders similar to tumors formed by the parental U87 glioma cells that do not express p75NTR. These data suggest that neurotrophin binding is required for p75NTR-mediated glioma invasion.
The highly invasive nature of malignant gliomas has been a substantial barrier in the treatment of patients with this disease. Data presented here strongly suggest that p75NTR, in a neurotrophin-dependent manner, is an important regulator of glioma invasion. To clinically validate p75NTR's role in glioma migration and invasion, and demonstrate its relevance in malignant glioma patient specimens, we analyzed the expression of p75NTR in a panel of surgically resected tumor specimens and normal human brain using immunohistochemical staining (Figure 6A), RT-PCR, and Western blot (Figure 6B). Expression of p75NTR protein was detected in 20 of 40 human glioma patient specimens (50%) (one of 11 low-grade astrocytomas [8%], two of nine mid-grade astrocytomas [22%], and 17 of 20 glioblastoma multiforme (GBM) specimens [85%]) and was undetectable in normal human brain (zero of five). Thus, expression of p75NTR is a common event in GBM. To demonstrate that the presence of p75NTR in these patient specimens confers an increased migratory ability, short-term cultures of these samples were analyzed in transwell motility assays. The percentage of cells positive for p75NTR in the original population was determined by immunostaining and compared to the percentage of p75NTR-positive cells in the migratory population (i.e., those cells that migrated to the underside of the transwell chamber during the assay). As a positive control for this assay, a mixture of 25% U87p75 cells and 75% U87pcDNA cells were used as input. At completion of the control assay, the migratory population contained approximately 50% p75NTR-positive cells (Figure 6C), as expected from initial experiments that demonstrated that p75NTR-positive cells migrate at a greater rate than the p75NTR-negative cells (Figure 3E and 3F). Similar effects were observed with the glioma patient specimens. The percentage of p75NTR-positive cells in the migratory population compared to the original population was increased by 40%–100% (Figure 6C), demonstrating that the p75NTR-positive cells within the glioma patient samples are more migratory than the p75NTR-negative glioma cells.
During the in vitro growth stage of the serial in vivo–selection procedure, we observed that the invasive glioma cells had striking morphological differences to the “tumor” cells. To examine the morphology of these cells, fluorescent staining of the actin cytoskeleton was performed. Staining of the actin cytoskeleton using rhodamine phalloidin revealed cells with numerous filamentous protrusions present only in the invading population (Figure 7A). Similarly, we found that expression of p75NTR alone induced structural rearrangement of the actin cytoskeleton similar to that of the in vivo–selected invasive cells (Figure 7B). Because the small molecular weight GTPase RhoA is a potential downstream readout from p75NTR that may help contribute to the distinct phenotype, we examined the effect of RhoA. Expression studies in HEK293 cells demonstrated that in the absence of ligand, p75NTR constitutively activated Rho, whereas ligand binding leads to a decrease in the levels of active Rho [27]. In addition, Gehler et al. [49] have shown that neurotrophin-bound p75NTR induces growth cone filopodia through the modulation of RhoA and that neurotrophin binding is necessary and sufficient to regulate filopodia dynamics. We found that concomitant with the changes in actin cytoskeleton, cells expressing p75NTR had reduced RhoA activity (Figure 7C and 7D).
Human malignant gliomas are highly invasive tumors. This highly invasive nature associates theses tumors with an extremely poor prognosis owing to recurrence of the tumor outside the margin of therapeutic resection [50]. Invasion of glioma cells into the normal surrounding brain requires changes that make these cells distinct from their noninvasive counterparts. Specifically, these glioma cells activate a number of coordinate cellular programs that involve the regulation of many molecules, including adhesion molecules, extracellular matrix constituents, proteases, cytoskeleton components, and signaling molecules. Altered regulation of any of these constituents may lead to changes in glioma cell migration and invasion. Although numerous molecules have been implicated in the migration and invasion of gliomas, what triggers glioma cells to leave the main tumor mass and invade into the normal brain is not well understood. To this end, we have developed a serial in vivo–selection paradigm to isolate highly invasive glioma cells from a human glioma cell line that is noninvasive in xenotransplantation models. A similar approach has been used to successfully assess the global gene expression profile of both melanoma and breast cancer metastasis [7,8]. Using this strategy, we identified and verified genes that were up-regulated in the invading glioma cells. One of the most differentially expressed genes encodes the neurotrophin receptor p75NTR that is the focus of this study. We provide the first evidence both in vitro and in vivo that p75NTR is a major mediator of glioma migration and invasion.
In recent years, there has been a growing importance of the neurotrophin signaling axis in cancer. Specifically, there is increasing evidence that the neurotrophic receptor tyrosine kinase TrkB, sometimes in conjunction with its primary ligand BDNF, is over-expressed in a variety of human cancers, ranging from neuroblastomas to pancreatic ductal adenocarcinomas [51–55]. Here, we present data that the pan-neurotrophin receptor p75NTR is expressed in malignant glioma and is a major contributor to their highly invasive nature. Although a universal role for p75NTR in cancer has not been established, recent studies implicate p75NTR in the metastatic progression of melanoma, and specifically in those tumors that metastasize to the brain [43,46,56]. Conversely, p75NTR expression has been linked to the progression of prostate cancer, but in this cancer, p75NTR, which is expressed in normal prostate epithelia, is lost upon transformation [45]. The divergence observed in the tumor progression of these two distinct tumors can likely be explained by the presence of Trk. In prostrate tumor cells, Trk expression is retained and mediates proliferation [42,57], whereas p75NTR-induced invasion in melanoma is independent of Trk expression [58]. Thus the recurring theme emerges that p75NTR function is cell-type specific (even in cancer) and must be independently determined for each cellular context. Here, we have shown that p75NTR-induced glioma invasion is also Trk independent with neither mRNA nor protein for the Trk receptors expressed by the invading glioma cells. Further supporting the Trk independence of p75NTR-mediated glioma invasion is the finding that treatment of the invasive cells with cleavage-resistant pro-NGF (which cannot bind Trk; Figure S2) also enhanced the migration of invading glioma cells.
Tumor cells can survive by means of an autostimulatory (autocrine) signaling loop, such as that mediated by TrkB and BDNF, or through a paracrine cross-communication with their environment. In brain metastatic melanoma, normal brain tissue adjacent to the melanoma displays increased neurotrophin expression [56], making it tempting to speculate that the metastatic melanoma uses the neurotrophin-rich nervous system as a paracrine mediator of invasion. It has similarly not escaped our attention that the neurotrophin environment of the brain may provide an extremely advantageous milieu for an invading glioma cell. Our data show that the p75NTR-expressing glioma cells are ligand responsive and may therefore use neurotrophins available in the brain environment to their advantage. In addition, we show that the invasive nature of glioma cells expressing p75NTR is negated when these cells express mutant p75NTR receptors that no longer bind to neurotrophin.
The concept of p75NTR playing a role in migration is not unprecedented. Neural crest cells, the most extensively studied population of migrating cells in the nervous system, express p75NTR even before they commit to any cell differentiation lineage [59]. In addition, Anton et al. [35] showed that stimulation of the p75NTR by NGF allowed Schwann cells to migrate on peripheral nerves, and examination of p75−/− mice showed severe impairment of Schwann cell migration, with no response to NGF [26]. More recently, the Hempstead laboratory [46] has shown that activation of p75NTR with NGF or pro-NGF (the unprocessed, precursor form of NGF) caused migration of melanoma cells and increased expression of p75NTR correlated with advanced stages and invasive potential of melanoma brain metastasis [60]. At present, the underlying mechanism of p75-induced migration of melanoma cells is not understood; however, p75NTR has been shown to interact with the actin cytoskeleton [46]. The small GTPase RhoA is a downstream effector of p75NTR [27,28]. The capability of p75NTR to modulate the activity of RhoA provides a reasonable explanation as to how p75NTR regulation might result in changes in cellular architecture of glioma cells. We found that concomitant with increased glioma invasion, glioma cells expressing p75NTR showed reduced RhoA activity and striking actin rearrangement.
Previous molecular characterization has defined genetic changes between low-grade and high-grade glioma [61–67]. In addition, molecular signatures of glioblastoma subtypes have been identified, including profiles of primary and secondary glioblastoma subgroups [68–70]. On the other hand, very little is known with respect to the transcriptional profiles of invading glioma cells. Studies have been performed using laser capture microdissection in patient specimens to collect the invasive cells and the cells from the main tumor mass. Although this approach has been used successfully to identify invasion-related genes [9], these experiments make the assumption that the invasive cells at the leading edge of the tumor have distinct profiles from the main tumor mass and that only tumor cells at the invading edge express genes important for migration and invasion. Yet, within the highly heterogeneous environment of a glioblastoma, in which there are many hypoxic and necrotic regions, it would be easy to envision that tumor cells experiencing oxidative stress would activate mechanisms enabling them to move to a more favorable environment. As such, some genes may not be identified using such an approach. Indeed, our data show that in addition to p75NTR-expressing glioma cells at the invasive edge of patient tumors, histological analysis identified p75NTR-positive glioma cells in regions of the tumor not adjacent to normal brain parenchyma. An alternative explanation for the appearance of p75NTR-positive glioma cells is that p75NTR promotes survival of glioma cells in vivo, though we did not find that p75NTR conferred a survival advantage in vitro. Additionally, reports of “stem-like” cells in brain tumors suggest that brain tumors arise from the transformation of neural stem cells [71–73], and when implanted into the brains of SCID mice, these cells form highly invasive tumors [16,74,75]. Whether these brain tumor stem cells express p75NTR is an important question for future studies, especially given that nestin-positive, p75NTR-positive cells have been identified in the subventricular zone of the adult brain [76].
Identification of key regulatory proteins of glioma invasion is extremely important clinically because this will be used to provide therapeutically relevant targets to prevent malignant glioma recurrence at the invasive margin of gliomas [4]. Herein, we present the first evidence that p75NTR is important in glioma migration, and the mere expression of p75NTR is sufficient to impart a dramatic invasive behavior on genetically distinct glioblastomas. Because p75NTR has also been implicated in the progression of melanoma, and specifically in those tumors that metastasize to the brain [33,46,77], therapies that target p75NTR, p75NTR downstream effectors, or their ligands may not only be beneficial for malignant glioma, but may target other metastatic diseases.
The human glioma cell line U87 was obtained from the American Type Culture Collection (http://www.atcc.org). The human glioma cell lines U251N and SF767 were kind gifts from V. W. Yong (University of Calgary, Calgary, Alberta, Canada) and M. Berens (Barrow Neurological Institute, Phoenix, Arizona, United States), respectively. All cells were maintained in complete medium (Dulbecco's Modified Eagle's Medium [DMEM] F12 supplemented with 10% heat-inactivated fetal bovine serum, 1% antibiotic/antimycotic, 0.1 mM nonessential amino acids, 2 mM L-glutamine and 1 mM sodium pyruvate (GIBCO BRL, http://www.invitrogen.com)] at 37 °C in a humidified 5% CO2 incubator. Cells were passaged by harvesting with trypsin when they reached 80%–90% confluence. Stable transfectants of U87 and SF767 cells were maintained in the same medium, with the addition of 400 μg/ml G418 or 200 μg/ml hygromycin, respectively (Invitrogen, http://www.invitrogen.com).
The GFP expression vector was pGFP-N1 from Clontech (http://www.clontech.com). The human p75NTR expression vector was constructed as described previously [46]. The expression plasmids containing the p75NTR mutants were constructed either by subcloning of PCR fragments containing the desired p75NTR sequences (for p75CRD130 construct) or by PCR-based site-directed mutagenesis (for the p75CRD105 constructs). Primers used for the construction of the mutants were: p75CRD105 primers (sense: 5′-CGG GCT CGG GCC GCT CGA GCG GCC TCG TGT TC–3′; antisense: 5′-GAA CAC GAG GCC GCT CGA GCG GCC CGA GCC CG–3′) and template p75WT [46]; p75CRD130 primers (sense: 5′-GAA GAT CTC CAA GGA GGC ATG CCC CAC AGG CC–3′; antisense: 5′-CTC ACT ATA GGT CGA CCG GAA TTC G– 3′) and template pT3/T7-p75. The original templates were from B. Hempstead (p75WT; Cornell University Medical College) and M. Chao (pT3/T7-p75; New York University School of Medicine, New York, New York, United States). The sequences of all the mutant expression plasmids were confirmed prior to stable transfection. The p75NTR-specific siRNA expression vector was constructed by ligating a double-stranded hairpin oligonucleotide: 5′-GAT CCG AGG ATC GGA GGC TTG TCA TTC AAG AGA TGA CAA GCC TCC GAT CCT CTT TTT TGG AAA-3′, containing a p75NTR-specific siRNA sequence (underlined), into the pSilencer 2.1-U6 hygro vector (Ambion, http://www.ambion.com). The negative control pSilencer vector, containing a random siRNA with limited homology to any known human, mouse, or rat sequences, was obtained from Ambion.
Cells to be transfected were seeded at 2 × 105 cells/well of a six-well plate and incubated at 37 °C overnight in complete media. Vector DNA was introduced to the cells using FuGENE 6 transfection reagent (Roche Diagnostic, http://www.roche.com) according to the manufacturer's instructions. The cells were then incubated at 37 °C overnight; and the following day, the medium was changed to fresh complete medium containing an antibiotic (concentration determined by toxicity curve for cell line) to select for those cells that had taken up the vector. The cells were then grown under antibiotic selection until the wells were confluent. For GFP transfection, transfected cells were identified by fluorescent microscopy and GFP expression of greater than 95% was obtained by fluorescence-activated cell sorting. For p75NTR, p75CRD105, p75CRD130, and p75NTR-siRNA transfection, transfected cells were identified by RT-PCR and Western blot.
Six- to 8-wk-old female SCID mice were purchased from Charles River Laboratories (http://www.criver.com). The animals were housed in groups of three to five, maintained on a 12-h light/dark schedule with a temperature of 22 °C ± 1 °C and a relative humidity of 50% ± 5%. Food and water were available ad libitum. All procedures were reviewed and approved by the University of Calgary Animal Care Committee.
Actively growing U87 cells expressing GFP and neomycin resistance genes (U87GFP) were harvested by trypsinization, washed, and resuspended in sterile PBS (137 mM NaCl, 8.1 mM Na2HPO4, 2.68 mM KCl, and 1.47 mM KH2PO4 [pH 7.5]). These cells were implanted intracerebrally into the right putamen of SCID mice (1 × 105 cells/mouse) at a depth of 3 mm through a scalp incision and a 0.5-mm burr hole made 1.5–2 mm right of the midline and 0.5–1 mm posterior to the coronal suture. Stereotactic techniques were described previously [77]. Tumor formation was allowed to proceed for 21+ days, depending on the health of the mouse and the type of cells injected. The mice were then sacrificed and the brain examined using fluorescence. The brain was divided in half coronally; one half was used for frozen sections and the other used for tissue culture. For tissue culture, the hemisphere containing the main tumor mass was separated from the contralateral hemisphere, and the two pieces were treated individually. The tissue was minced into small pieces and dissociated with trypsin and DNase I at 37 °C. The tissue suspension was then forced through a 100-μm mesh, and the resulting cell suspension was centrifuged and resuspended in complete medium containing 400 μg/ml G418 to select for the GFP-transfected tumor cells. Cells obtained from the tumor mass were labeled as “tumor” cells, and those from the contralateral hemisphere were labeled as “invasive” cells. Tumor and invasive cells were then reinoculated into SCID mice, and the procedure was repeated.
Total cellular RNA was extracted from subconfluent cells using Trizol Reagent (Invitrogen) and DNase-treated using DNA-free (Ambion). The reverse transcription reaction took place in a buffer of 10 mM Tris-HCl (pH 9.0), 50 mM KCl, and 1.5 mM MgCl2, and contained 3 μg of total RNA, 25 units of RNAguard RNase inhibitor, 1 mM each of deoxynucleoside triphosphates, 100 ng of pd(N)6 random hexanucleotide primers (Amersham Biosciences, http://www.amersham.com), and 200 units of Superscript II reverse transcriptase (Invitrogen). The PCR amplification reaction was carried out in the same buffer and contained 1 μl of the cDNA synthesis reaction, 80 μM each of deoxynucleoside triphosphates, 1 unit of Taq DNA polymerase (Amersham Biosciences), and 0.1 μM each of p75NTR-specific primers (forward: 5′-CGT ATT CCG ACG AGG CCA ACC-3′; reverse: 5′-CCA CAA GGC CCA CAA CCA CAG C-3′), p75CRD105-specific primers (forward: 5′-CGG GCT CGG GCC GCT CGA GCG GCC TCG TGT TC-3′; reverse primer: 5′-GAA CAC GAG GCC GCT CGA GCG GCC CGA GCC CG-3′), p75CRD130-specific primers (forward: 5′-GAA GAT CTC CAA GGA GGC ATG CCC CAC AGG CC-3′; reverse primer: 5′-CTC ACT ATA GGT CGA CCG GAA TTC G-3′), or 0.2 μM each of GAPDH-specific primers (forward primer: 5′-CGG AGT CAA CGG ATT TGG TCG TAT-3′; reverse primer: 5′-AGC CTT CTC CAT GGT GGT GAA GAC-3′). The amplification consisted of 35 cycles of 45 s at 94 °C, 30 s at 63 °C, and 45 s at 72 °C, followed by a 7-min extension at 72 °C after the last cycle. The reaction products were then resolved on a 1% agarose gel containing ethidium bromide.
Total cellular lysates were obtained by gentle rocking in lysis buffer (20 mM Tris [pH 8.0], 137.5 mM NaCl, 10% glycerol, 1% Nonidet P-40, 25 μg/ml aprotinin, 10 μg/ml leupeptin, 3 mM sodium orthovanadate, 1 mM PMSF) at 4 °C. Protein extracts of human glioma biopsies were obtained by immersing the samples in ice-cold extraction buffer (50 mM Tris [pH 7.6], 200 mM NaCl, 10 mM CaCl2, 1% Triton X-100) followed by homogenization on ice. Cellular debris was removed by centrifugation, and protein quantification was performed using the bicinchoninic acid (BCA) assay (Pierce Biotechnology, http://www.piercenet.com). Proteins were resolved on 10% SDS-PAGE gels, and Western blots were performed using the following primary antibodies: rabbit polyclonal anti-human p75NTR (Promega, http://www.promega.com), goat polyclonal anti-pyruvate kinase (Chemicon, http://www.chemicon.com), mouse monoclonal anti-neurofilament, 70 kDa (Chemicon), mouse monoclonal anti-glial fibrillary acid protein (ChemiconA), or mouse monoclonal anti-β-tubulin (Sigma-Aldrich, http://www.sigmaaldrich.com). The appropriate HRP-conjugated secondary antibody (Pierce Biotechnology) was used and visualized using enhanced chemiluminescence (Amersham Biosciences).
Cells were collected using Puck's EDTA at 37 °C and then washed in PBS containing 1 mM EDTA (PBS/EDTA). Cells were then treated with monoclonal anti-p75NTR, clone ME20.4 (which recognizes the extracellular domain; Upstate Biotechnology, http://www.upstate.com), diluted 1:250 in PBS/EDTA for 30 min on ice. The negative control sample was incubated in only PBS/EDTA. After washing with PBS/EDTA, cells were treated with Alexa 488–conjugated goat anti-mouse IgG (Invitrogen Molecular Probes, http://probes.invitrogen.com) diluted 1:500 in PBS/EDTA for 30 min on ice. Cells were then washed with PBS/EDTA, resuspended in PBS/EDTA, and analyzed on a FACScan flow cytometer (Becton, Dickinson and Company, http://www.bdbiosciences.com).
Cells were allowed to condition the medium for 5 d. The conditioned medium was then collected, centrifuged, and filtered through a 0.2-μm syringe filter (VWR International, http://www.vwr.com). The remaining cells were washed with ice-cold PBS, and total cellular lysates were extracted as described for Western blot. Protein quantification was performed using the BCA assay (Pierce Biotechnology) and BDNF enzyme-linked immunosorbent assays (ELISAs) (R&D Systems, http://www.rndsystems.com) were performed as per the company protocol. Briefly, MaxiSorp ELISA plates (Nalge Nunc International, http://www.nalgenunc.com) were coated with monoclonal anti-human BDNF (R&D Systems), nonspecific binding was blocked, and serial dilutions of recombinant human BDNF (Sigma-Aldrich), equal volumes of conditioned medium, or equal quantities of lysate were added. Bound antigen was detected using the corresponding biotinylated antibody, streptavidin HRP, and a tetramethylbenzidine substrate (R&D Systems).
Migration assays were performed using a microliter-scale radial monolayer migration assay as described by Berens et al. [78]. Briefly, ten-well Teflon-masked microscope slides were coated with 20 μg/ml laminin. Cells were seeded through a cell sedimentation manifold (Creative Scientific Methods, http://www.cre8ive-sci.com) at 2,000 cells/well to establish a circular 1-mm diameter confluent monolayer. Once the sedimentation manifolds were removed, cells were given complete medium containing the appropriate treatment. A digital image of the cells was taken (before migration = 0 h) using a Zeiss Axiovert 200M inverted fluorescent microscope (Carl Zeiss, http://www.zeiss.com). The cells were then incubated in a humidified chamber at 37 °C and 5% CO2, and a second digital image was taken 48 h later. Best-fit circles were drawn around the area covered by the cells at the 0-h and 48-h time points and the actual cell area determined using Axiovision 4.2 imaging software (Carl Zeiss). Quantitative migration scores were calculated as the increase in the area covered by the cells beyond the initial area of the cells.
Matrigel (BD Bioscience, Mississauga, Ontario, Canada) was diluted with two parts of cold serum-free medium, layered onto an 8-μm pore-size transwell chamber (BD Bioscience, http://www.bdbiosciences.com), and incubated at room temperature for 1 h. The wells were then rinsed with serum-free medium. The coated chambers were placed into the wells of a 24-well tissue culture plate containing 500 μl of media with or without the desired treatment. Serum-starved cells (2.5 × 104) were seeded into each chamber, in a volume of 500 μl of the same medium contained in the bottom of the well, and incubated at 37 °C for 48 h. The medium was then removed from the chambers and cells scraped off the top of the membrane using a PBS-soaked cotton-tipped swab. Cells were fixed to the bottom of the chamber with methanol, stained in hematoxylin, and mounted on slides. Invasion was quantified by counting the stained cells adherent to the lower side of the membranes in ten fields (at 10× magnification) for each of three chambers for each condition.
Actively growing U87pcDNA, U87p75, U87CRD105, U87CRD130, U251NpcDNA, and U251Np75 cells were implanted intracerebrally into SCID mice as described previously [77]. Mice were sacrificed weekly from day 14–42. At each time point, the brains were removed, frozen in OCT compound (Tissue-Tek; Electron Microscopy Sciences, http://www.emsdiasum.com), and cryosectioned for examination by immunohistochemistry.
Frozen sections were air dried at room temperature, fixed with cold acetone, and then rinsed with PBS. Paraffin sections were dewaxed and rehydrated using a xylene/ethanol series followed by rinsing with PBS. Endogenous peroxidases in the sections were inactivated with 0.075% H2O2/methanol, and nonspecific binding was blocked with 10% normal goat serum in PBS. The sections were then incubated with rabbit polyclonal anti-human p75NTR (Promega, http://www.promega.com) or mouse monoclonal anti-human nuclei (Chemicon) in blocking buffer overnight at 4 °C. Following washing with PBS, the appropriate biotinylated secondary antibody (Vector Laboratories, http://www.vectorlabs.com) was applied. Avidin-biotin peroxidase complexes were then formed using the VECTASTAIN Elite ABC kit (Vector Laboratories) and detected by addition of SIGMAFAST DAB (3,3′-diaminobenzidine tetrahydrochloride) (Sigma-Aldrich). The SIGMAFAST DAB was converted to a brown reaction product by the peroxidase. Hematoxylin (for paraffin sections) and toluidine blue (for frozen sections) were used as nuclear counterstains. Sections were then dehydrated in an ethanol/xylene series and mounted with Entellan (Electron Microscopy Sciences).
Coverslips were coated with a Collagen I (3 mg/ml; Vitrogen 100; Cohesion Technologies, http://www.cohesiontech.com) and incubated overnight at 37 °C. Excess collagen solution was aspirated, and cells were plated at 2 × 105/mL in DMEM culture medium (DMEM with 10% FBS, 6 mM L-glutamine, 100 μM nonessential amino acids, 1 mM sodium pyruvate, 400 μg/ml G418) and allowed to equilibrate overnight at 37 °C, 5% CO2. Coverslips were then rinsed twice with PBS, fixed in 3.7% formaldehyde diluted in PBS for 10 min, and rinsed twice with PBS. Unpolymerized actin was extracted for 3 min in CSK buffer (10 mM MES [pH 6.1], 138 mM KCl, 3 mM MgCl2, 2 mM EGTA, 320 mM sucrose, 0.1% Triton X-100) followed by two rinses with PBS. Alexa Fluor 568 phalloidin (Invitrogen) was diluted 1:40 in 1% BSA/PBS and 200 μl of this solution was added to each coverslip for 20 min at room temperature. Coverslips were rinsed twice with PBS, counterstained with a 500 nM solution of DAPI for 3 min, mounted in glycerol, and imaged with an Olympus IX70 Delta Vision RT Microscope (http://www.olympus.co.jp/en/) and the SoftWoRx software package.
Tumor and normal tissues were obtained from the Canadian Brain Tumor Tissue Bank in London, Ontario, and Foothills Hospital, Calgary, Alberta. Briefly, tissue was taken during surgery while patients were under a general anesthetic, and was placed immediately in liquid nitrogen and stored at −80 °C or placed in culture medium for establishment of short-term cultures. An institutional ethics board approved the collection and use of all of the surgical tissue used, and all of the patients gave signed informed consent. The following tissues were studied: 20 GBMs, eight anaplastic astrocytomas, one anaplastic oligodendroglioma, five astrocytomas, five mixed oligoastrocytomas, one oligodendroglioma, and five controls obtained during nontumor brain surgery.
Operative samples of human gliomas were obtained during brain tumor surgery and transported to the laboratory in culture medium. Short-term cultures were then established. Briefly, necrotic and connective tissue and any blood clots were removed using forceps, and the remaining tissue was washed in PBS and cut into pieces of approximately 1 mm2. The tissue was then incubated for 30 min at 37 °C in an enzyme cocktail of trypsin (0.25%) and DNase I (10 μg/ml) in PBS. The digested tissue was strained through a 100-μm mesh and washed with PBS. The cells were then pelleted and washed with DMEM-F12 media. Following lysis of red blood cells, the remaining cells were washed with PBS, pelleted, resuspended in complete media containing 20% FBS, and plated.
Primary human glioma cells cultured for less than 3 wk were “serum-starved” by incubating them in medium containing only 1% FBS for 2 h at 37 °C and 5% CO2. Cells were then released from the culture dish using Puck's EDTA (1 mM EDTA, 10 mM HEPES, 5 mM KCl, 140 mM NaCl, 4 mM NaHCO3, and 6 mM dextrose [pH 7.3]) at 37 °C. Cells (2.5 × 104) suspended in 1% FBS medium were plated in eight-well chamber slides (Nunc) and transwell chambers (Costar) coated with 20 μg/ml laminin. Medium containing 1% FBS was placed below the chamber. The cells were incubated at 37 °C and 5% CO2 for 6 h. The medium was then removed from the chambers, and cells were fixed with 4% paraformaldehyde. The cells were then stained for human p75NTR, as described for immunohistochemical staining, and counterstained with hematoxylin. In the transwell chamber, cells that did not migrate were scraped off the top of the membrane using a cotton-tipped swab. Migration was quantified by counting the p75NTR-positive (brown) and p75NTR-negative (blue) cells in the original population (on the slide) or in the migratory population (adherent to the under side of the transwell membrane) in five fields (at 20× magnification) for each of four chambers.
|
10.1371/journal.ppat.1006506 | KIR3DL01 upregulation on gut natural killer cells in response to SIV infection of KIR- and MHC class I-defined rhesus macaques | Natural killer cells provide an important early defense against viral pathogens and are regulated in part by interactions between highly polymorphic killer-cell immunoglobulin-like receptors (KIRs) on NK cells and their MHC class I ligands on target cells. We previously identified MHC class I ligands for two rhesus macaque KIRs: KIR3DL01 recognizes Mamu-Bw4 molecules and KIR3DL05 recognizes Mamu-A1*002. To determine how these interactions influence NK cell responses, we infected KIR3DL01+ and KIR3DL05+ macaques with and without defined ligands for these receptors with SIVmac239, and monitored NK cell responses in peripheral blood and lymphoid tissues. NK cell responses in blood were broadly stimulated, as indicated by rapid increases in the CD16+ population during acute infection and sustained increases in the CD16+ and CD16-CD56- populations during chronic infection. Markers of proliferation (Ki-67), activation (CD69 & HLA-DR) and antiviral activity (CD107a & TNFα) were also widely expressed, but began to diverge during chronic infection, as reflected by sustained CD107a and TNFα upregulation by KIR3DL01+, but not by KIR3DL05+ NK cells. Significant increases in the frequency of KIR3DL01+ (but not KIR3DL05+) NK cells were also observed in tissues, particularly in the gut-associated lymphoid tissues, where this receptor was preferentially upregulated on CD56+ and CD16-CD56- subsets. These results reveal broad NK cell activation and dynamic changes in the phenotypic properties of NK cells in response to SIV infection, including the enrichment of KIR3DL01+ NK cells in tissues that support high levels of virus replication.
| Natural killer (NK) cells are an important cellular defense against viral pathogens, and are regulated in part by interactions between killer-cell immunoglobulin-like receptors (KIRs) on NK cells and MHC class I ligands on target cells. Using multi-parameter flow cytometry, we report the first longitudinal study of changes in the phenotypic and functional properties of NK cells in KIR- and MHC class I-defined rhesus macaques infected with simian immunodeficiency virus (SIV). Our findings reveal broad NK cell activation and highly dynamic changes in the phenotypic properties of NK cells in response to SIV infection, including an enrichment of NK cells expressing KIR3DL01 in tissues that represent sites of high levels of virus replication.
| Natural killer cells provide a critical early defense against viral pathogens by directly responding to infected cells without prior antigenic stimulation. This is accomplished through the integration of signals from activating and inhibitory receptors, which in primates include the highly polymorphic killer-cell immunoglobulin-like receptors (KIRs) [1,2]. KIRs contain two or three extracellular immunoglobulin-like domains (2D or 3D), and depending on whether they have long (L) or short (S) cytoplasmic tails, transduce either inhibitory or activating signals [1,2]. MHC class I molecules serve as ligands for the inhibitory KIRs [1,2], and although the ligands for the activating KIRs are not as well defined, there is evidence that these receptors also recognize MHC class I molecules [3–5]. In the case of inhibitory KIRs, engagement of ligands on the surface of healthy cells normally suppresses NK cell activation; however, if these interactions are disrupted, for instance as a consequence of MHC class I downregulation by the HIV-1 Nef protein [6–8], this inhibition is lost, triggering NK cell degranulation and the cytolysis of infected cells.
The specificity of inhibitory KIRs is primarily determined by contacts with the α1 and α2 domains of their ligands. All HLA-B molecules and some HLA-A molecules can be categorized as either Bw4 or Bw6 allotypes depending on residues 77–83 of their α1 domains [9]. Whereas KIR3DL1 selectively binds to HLA-Bw4 ligands, no human KIRs are known to recognize HLA-Bw6 molecules. HLA-C molecules can likewise be classified as C1 or C2 allotypes on the basis of polymorphisms at positions 77 and 80, which are recognized respectively by KIR2DL2 and KIR2DL3 or KIR2DL1 depending on the amino acid residues at these positions [10,11]. Consistent with crystal structures showing that KIRs contact HLA class I surfaces over C-terminal peptide residues [12–14], peptides bound by MHC class I ligands can also influence these interactions [15,16].
KIR and HLA class I polymorphisms are associated with differences in the course of HIV-1 infection [17–19]. In HIV-1 infected individuals, KIR3DS1 and highly expressed KIR3DL1 alleles in combination with HLA-Bw4 alleles encoding isoleucine at position 80 (HLA-Bw4-80I) are associated with lower viral loads and slower courses of disease progression [17,20]. Accordingly, KIR3DS1+ and KIR3DL1+ NK cells preferentially expand in response to HIV-1 infection in HLA-Bw4-80I+ individuals [21]. In vitro studies have also shown that KIR3DS1+ NK cells can suppress HIV-1 replication in lymphocytes from HLA-Bw4-80I+ donors, but not from HLA-Bw6 homozygous donors [3], and that KIR3DL1+ NK cells respond to HIV-1-infected cells that have downmodulated HLA-Bw4 ligands in a manner that reflects hierarchical differences in their education [22]. Additional studies have identified HIV-1 polymorphisms that suppress KIR2DL2+ and KIR2DL3+ NK cell responses to virus-infected or peptide-pulsed cells in vitro, suggesting that HIV-1 is under selective pressure in certain individuals to acquire changes in epitopes that stabilize HLA-C interactions with inhibitory KIRs as a mechanism of immune evasion [23–25].
Simian immunodeficiency virus (SIV) infection of the rhesus macaque is an important animal model for HIV-1 pathogenesis and AIDS vaccine development [26]; however, studies to address the role of NK cells in this system have been limited by immunogenetic differences between humans and macaques and a lack of defined ligands for macaque KIRs. Unlike humans, which have HLA-A, -B and–C genes, macaques and other Old world monkeys do not have a C locus [27,28]. Instead, these species have an expanded repertoire of A and B genes [27–30]. There are up to four Macaca mulatta (Mamu)-A genes and a highly variable number of Mamu-B genes on any given haplotype in the rhesus macaque [31,32]. Macaques accordingly lack KIR2DL/S genes that encode receptors for HLA-C, but have an expanded complement of highly polymorphic KIR3DL/S genes [33–37]. Phylogenetic and segregation analyses support the existence of 22 KIR genes in macaques [35,36,38]; however, as a consequence of the rapid pace of KIR evolution, only two of these genes (Mamu-KIR2DL04 and -KIR3DL20) have recognizable human orthologs [1,2,39–41]. Thus, it is not possible to predict the ligands for macaque KIRs based on sequence similarity with their human counterparts.
MHC class I ligands have nevertheless been identified experimentally for a few rhesus macaque KIRs [29,30,42]. Mamu-A1*002, a molecule with a canonical Bw6 motif, was identified as a ligand for Mamu-KIR3DL05 (KIR3DL05) [30]. Mamu-A1*002 and KIR3DL05 are respectively expressed by approximately 20% and 40% of Indian-origin rhesus macaques, and the binding of KIR3DL05 to Mamu-A1*002 is strongly influenced by SIV peptides [16,30]. Functional assays with primary NK cells also identified multiple Bw4 molecules as ligands for Mamu-KIR3DL01 (KIR3DL01). KIR3DL01 is the most polymorphic KIR in rhesus macaques and is expressed by 85–95% of animals of Indian origin [29,40]. Most rhesus macaques also have one or more Mamu-Bw4 alleles predicted to encode ligands for this receptor. Despite their coincidental similarity in nomenclature, rhesus KIR3DL01 and human KIR3DL1 are not orthologous gene products; however, their shared specificity for Bw4 ligands suggests that they may serve similar functions.
In the present study, we investigated NK cell responses to SIV infection of KIR- and MHC class I-defined macaques. Twelve KIR3DL05+ macaques, of which half were Mamu-A1*002+, eleven were KIR3DL01+, and all but one were Mamu-Bw4+, were infected with SIVmac239, and longitudinal changes in NK cell subsets were monitored in peripheral blood and tissues. Infection with SIV broadly stimulated NK cell responses, resulting in significant increases in the number of NK cells in blood expressing markers of activation, proliferation and antiviral activity. Significant increases were also observed in the frequency of KIR3DL01+ NK cells in lymph nodes and gut-associated lymphoid tissues. These results reveal dynamic changes in the phenotypic and functional properties of NK cells in response to SIV infection and an enrichment of KIR3DL01+ NK cells at sites of early virus replication and CD4+ T cell turnover.
We previously identified MHC class I ligands for two rhesus macaque KIRs. We found that KIR3DL05 binds to Mamu-A1*002, a common MHC class I molecule in the rhesus macaque with a Bw6 motif [30], and that KIR3DL01, which is among the most polymorphic and commonly expressed KIRs in rhesus macaques, recognizes MHC class I ligands with a Bw4 motif [29]. We further demonstrated that nearly a third of the SIV peptides bound by Mamu-A1*002 suppress the cytolytic activity of KIR3DL05+ NK cells by stabilizing this interaction [16]. To determine how these receptor-ligand interactions influence NK cell responses and the outcome of immunodeficiency virus infection, twelve KIR- and MHC class I-defined rhesus macaques were infected intravenously with SIVmac239, and longitudinal changes in NK cells and viral loads were monitored in peripheral blood and lymphoid tissues.
KIR3DL05+ macaques were initially identified by staining PBMCs with Mamu-A1*002 Gag GY9 tetramers as previously described [30]. Six Mamu-A1*002+ and six–A1*002- animals were then selected from this group on the basis of MHC class I genotyping (Table 1). Five of the Mamu-A1*002+ animals and three of Mamu-A1*002- animals were also positive for Mamu-A3*13, which encodes another molecule identified as a ligand for KIR3DL05 (Table 1) [42]. As a reflection of the high prevalence of KIR3DL01 in rhesus macaques, eleven of these animals expressed KIR3DL01 allotypes that could be detected by staining with the NKVFS1 antibody (Table 2) [29]. Eleven of the animals were also positive for one or more Mamu-Bw4 alleles predicted to encode ligands for KIR3DL01 (Table 1). Complete KIR genotyping by next generation sequencing corroborated the presence or absence of KIR3DL01 and KIR3DL05, and identified twenty-three novel KIR alleles in these animals (Table 2 & S1 Table).
SIV loads in plasma and lymphocyte counts in peripheral blood were monitored at weekly to monthly intervals after SIV inoculation. Absolute counts for NK and T cell subsets were determined by staining whole blood with antibodies to lineage-specific markers in a bead-based assay to adjust for sample volume (S1 Fig). NK cells were defined as CD8+CD3- lymphocytes and verified by staining with an antibody to human NKG2A that cross-reacts with multiple NKG2 family members in macaques [43]. PBMCs were also stained in parallel with a separate panel that included antibodies (or tetramer) to KIR3DL01, KIR3DL05 (Mamu-A1*002 Gag GY9 tetramer), CD16 and CD56. Absolute counts for NK cell subsets defined by the expression of KIR3DL01, KIR3DL05, CD16 and CD56 were calculated as a percentage of total NK cell counts at each time point. Longitudinal changes in lymphocyte counts for individual animals are shown in Fig 1 and summarized as mean cell counts for Mamu-A1*002+ versus–A1*002- animals in S2 Fig.
Consistent with the variegated expression of KIRs, the majority of NK cells were KIR3DL01/05 double-negative (KIR3DL01-05-) and there was considerable animal-to-animal variation in the frequency of KIR3DL01+ and KIR3DL05+ NK cells. Prior to SIV inoculation, KIR3DL01+ cells constituted 12.8% ±4.7 (22.5±13.2 cells/μl), KIR3DL05+ cells constituted 8.7% ±1.4 (17.3±6.3 cells/μl) and double-positive (KIR3DL01+05+) cells constituted 0.58% ±0.21 (0.84±0.34 cells/μl) of circulating NK cells. In response to SIV infection, sharp increases were observed in total NK and CD8+ T cell counts (Fig 1A and 1B). These responses were reflected by significant increases in each of the KIR-defined NK cell subsets during acute infection (weeks 1–4), and were sustained by the KIR3DL05+, KIR3DL01+05+ and KIR3DL01-05- subsets during chronic infection (weeks 6–24) (Fig 1D–1G).
Phenotypic analyses previously defined CD16+CD56- and CD16-CD56+ NK cell populations in rhesus macaques that correspond to CD16+CD56dim and CD16-CD56bright populations in humans [44,45]. Similar to their human counterparts, CD16+CD56- (CD16+) NK cells represent the predominant NK cell population in blood and have a higher capacity for cytolytic activity than the less frequent and less mature CD16-CD56+ (CD56+) subset [44]. Macaques also have a CD16-CD56- NK cell population that is not found in humans, which appears to represent an intermediate in the differentiation of CD56+ NK cells into CD16+ NK cells [44–46]. Consistent with a previous cross-sectional study [44], we observed significant increases in cell counts for the CD16+ subset during acute and chronic infection (Fig 1H), and for the CD16-CD56- subset during chronic infection (Fig 1J), whereas the CD56+ population remained unchanged (Fig 1I).
Unlike humans, which only express KIRs on functionally mature CD16+CD56dim NK cells, macaques express KIRs on CD16+, CD56+ and CD16-CD56- NK cells [47]. Longitudinal comparisons revealed differences in the distribution of KIR3DL01 and KIR3DL05 on these subsets. Prior to SIV inoculation, KIR3DL01 and KIR3DL05 were both expressed at a higher frequency on CD16+ NK cells than on CD56+ or CD16-CD56- NK cells (Fig 2). Following SIV infection, the frequency of CD56+ and CD16-CD56- NK cells expressing KIR3DL05 rapidly increased, approaching similar percentages as the CD16+ subset during chronic infection (Fig 2B). In contrast, the distribution of KIR3DL01 did not change during acute infection; however, the percentage of KIR3DL01+ CD16+ NK cells gradually declined coincident with an increase in the frequency of CD56+ and CD16-CD56- NK cells expressing this KIR during chronic infection (Fig 2A). These changes are reflected by a decrease in the frequency of CD16-CD56- NK cells and an increase in the frequency of CD16+ NK cells lacking both KIR3DL01 and KIR3DL05 (KIR3DL01-05-) (Fig 2C). These observations reveal differential changes in the proportion of CD16+, CD56+ and CD16-CD56- NK cells expressing KIR3DL01 versus KIR3DL05 in response to SIV infection.
NK cell education in humans increases the frequency of cells bearing KIRs that recognize HLA class I ligands, while decreasing cognate receptor levels on the cell surface [48–50]. We therefore analyzed the percentage NK cells expressing KIR3DL01 and KIR3DL05, and the mean fluorescence intensity of surface staining for these receptors, with respect to the presence or absence of their MHC class I ligands. Neither the frequency nor the intensity of KIR3DL01 staining correlated with the predicted number of Mamu-Bw4 alleles (Fig 3A and 3B). The frequency of KIR3DL05+ NK cells also did not correlate with the number of alleles encoding ligands for this receptor (Mamu-A1*002 and/or–A3*13) (Fig 3C); however, KIR3DL05 staining did correlate inversely with the presence of these alleles (Fig 3D). This correlation appears to be primarily driven by Mamu-A1*002, since independent comparisons of KIR3DL05 staining for animals with or without these alleles revealed significantly lower KIR3DL05 levels in association with Mamu-A1*002 (Fig 3F), but not Mamu-A3*13 (S3A Fig). Additional analyses indicated that these patterns of KIR3DL01 and KIR3DL05 staining did not change in response to SIV infection. Comparisons of the frequency and intensity of KIR3DL01 staining at several time points before and after SIV infection did not reveal significant associations with the number of Mamu-Bw4 alleles (S3B Fig). Similarly, whereas there was no difference in the frequency of KIR3DL05+ NK cells before or after SIV infection (Fig 3E), KIR3DL05 staining was significantly lower for Mamu-A1*002+ animals prior to SIV inoculation and at multiple time points during acute and chronic infection (Fig 3F).
The absence of a correlation between KIR3DL01 staining and the number of Bw4 alleles in these animals may reflect incomplete knowledge of the ligands for this receptor, since several of the Mamu-Bw4 alleles listed in Table 1 were predicted to encode ligands for KIR3DL01 based on sequences in their α1 and α2 domains [29], rather than on experimental verification. Furthermore, because next generation sequencing methods used for KIR and MHC class I genotyping cannot differentiate two or more alleles with the same sequence, these analyses do not account for possible differences in the copy number of KIR3DL01 or Mamu-Bw4 genes that may influence KIR3DL01 expression in some animals. In the case of KIR3DL05, a dominant effect of Mamu-A1*002 on the expression of this receptor is consistent with the unusually high avidity of Mamu-A1*002 for KIR3DL05 [30] and with higher expression levels for Mamu-A1*002 than for ‘minor’ alleles of the Mamu-A3*13 locus [51]. Thus, the lower levels of KIR3DL05 staining detected on NK cells from Mamu-A1*002 animals suggest that this molecule may have a particularly strong effect on the education of KIR3DL05+ NK cells.
Despite differences in KIR3DL05+ staining in Mamu-A1*002+ versus–A1*002- macaques, and changes in the frequency of NK cells expressing this KIR during acute infection, viral loads and CD4+ T cell counts did not differ for Mamu-A1*002+ versus -A1*002- animals (Fig 1C, 1K and 1L). These comparisons therefore did not reveal an advantage to SIV replication in Mamu-A1*002+ macaques as a consequence of the presentation of inhibitory peptides to KIR3DL05+ NK cells, as suggested by cell culture experiments with sorted KIR3DL05+ NK cells [16].
Changes were observed in the expression of proliferation and activation markers on circulating NK cells in response to SIV infection that correspond to changes in absolute NK cell counts. The percentages of NK cells expressing Ki-67 as a marker for proliferation, and CD69 or HLA-DR as activation markers, were determined for CD16+, CD56+, CD16-CD56-, KIR3DL01+, KIR3DL05+ and -KIR3DL01-05- NK cells (S4 Fig), and used to calculate absolute counts for each subset. Since differences in NK cell counts were not observed for Mamu-A1*002+ versus–A1*002- animals, data from all eleven KIR3DL01+ animals and all twelve KIR3DL05+ animals were analyzed together. Significant increases were detected in the number of cells expressing Ki-67, CD69 and HLA-DR during acute and chronic infection (Fig 4). The most significant increases were observed for CD16+ NK cells (Fig 4A, 4C and 4E), which represent the largest and most functionally mature NK cell population in peripheral blood. Increases in the number of cells expressing Ki-67 were particularly evident during acute infection (Fig 4A and 4B), indicating that changes in NK cell counts during the first three weeks of SIV infection are probably due, at least in part, to cell proliferation. Higher numbers of KIR3DL01+, KIR3DL05+ and KIR3DL01-05- NK cells expressing CD69 and HLA-DR further indicate broad NK cell stimulation (Fig 4D and 4F). Thus, changes in NK cell activation and proliferation parallel increases in cell numbers during acute and chronic SIV infection.
To assess the potential of NK cells to home to tissues, PBMCs were stained with a panel that included antibodies to the mucosal homing receptor α4β7 as an indicator of trafficking to the intestinal mucosa, the chemokine receptor CCR7 as a marker for trafficking to peripheral lymphoid tissues and the chemokine receptor CXCR3 as a marker for trafficking to sites of inflammation. The numbers of CD16+, CD56+, CD16-CD56-, KIR3DL01+, KIR3DL05+ and KIR3DL01-05- NK cells expressing each of these markers were calculated from absolute counts as described above. Relatively few circulating NK cells expressed CCR7 or CXCR3, and with the exception of a modest increase in CXCR3+ KIR3DL01-05- NK cells during acute infection, significant changes for these markers were not detected (S5 Fig). Significant increases were, however, observed in the frequency of CD16+ and CD16-CD56- NK cells expressing α4β7, particularly during chronic infection (Fig 4G). Among the KIR-defined subsets, the KIR3DL01-05- population exhibited the greatest increase in the percentage of α4β7+ cells, with significant increases in the frequency of α4β7 also detectable for KIR3DL01+ cells during acute infection and KIR3DL05+ cells during chronic infection (Fig 4H).
The potential antiviral activity of circulating NK cells was assessed by staining for functional markers of degranulation and cytokine release. PBMCs were incubated overnight, with and without stimulation with MHC class I-deficient 721.221 cells, and in the presence of an antibody to CD107a as a marker for degranulation. The cells were then stained the following day with reagents to differentiate KIR3DL01+ and KIR3DL05+ NK cells and for intracellular accumulation of TNFα (S6 Fig). Because of variations in cell viability as a result of overnight incubation, these markers were analyzed as percentages of their respective NK cell populations rather than calculating absolute cell counts.
In accordance with the especially broad and potent “missing self” stimulus provided by the 721.221 cell line, the overall magnitude of CD107a and TNFα upregulation was much higher in response to incubation with 721.221 cells than in the absence of these cells; however, differences in the expression of these markers were not detectable among KIR3DL01+, KIR3DL05+ and KIR3DL01-05- NK cell subsets (Fig 5A and 5B). We therefore focused on NK cell responses without 721.221 cells, occurring as a result of in vivo and/or ex vivo activation by SIV-infected cells, as a more physiological reflection of antiviral activity. In the absence of 721.221 cells, CD107a and TNFα were strongly upregulated on KIR3DL01-05- NK cells during acute and chronic infection (Fig 5C and 5D). Although CD107a expression on KIR3DL01+ and KIR3DL05+ NK cells was delayed relative to the KIR3DL01-05- population, significant increases in both CD107a and TNFα were also detected for these subsets during acute infection (Fig 5C and 5D); however, while these responses were sustained for KIR3DL01+ NK cells, the percentage of KIR3DL05+ NK cells expressing CD107a and TNFα declined to baseline levels during chronic infection (Fig 5C and 5D). Hence, these results reveal functional differences in degranulation and cytokine release that suggest greater antiviral activity for KIR3DL01+ NK cells than for KIR3DL05+ NK cells.
Lymphocytes were isolated from lymph node and colorectal biopsies prior to SIV infection, and at two- and eight-weeks post-infection, to assess changes in the frequency of NK cells in these tissues. Compared to peripheral blood, NK cells constituted a relatively small percentage of lymphocytes in these tissue compartments. Before SIV infection, average NK cell frequencies in lymph nodes and gut-associated lymphoid tissues (GALT) were 1.2%±0.086 and 0.52%±0.065, respectively, compared to 7.0%±0.96 in PBMCs. By eight-weeks post-infection, a small but highly significant increase in the percentage of total NK cells was detectable in lymph nodes (Fig 6A). This increase was reflected by changes in the CD16+, CD56+ and CD16-CD56- subsets. While the majority of NK cells in macaque lymph nodes are CD16-CD56- [44], a significant increase in the frequency of CD16+ NK cells, and a corresponding decrease in the frequency of CD56+ cells, was observed in response to SIV infection (Fig 6B). In contrast, the percentage of total NK cells in GALT did not change following SIV infection (Fig 6A), and aside from a transient increase in the CD56+ population during acute infection, the proportions of CD16+, CD56+ and CD16-CD56- cells also remained unchanged (Fig 6C).
To assess changes in the frequency of NK cells expressing KIR3DL01 and KIR3DL05, the percentages of these cells were compared in lymph node and colorectal biopsies. Whereas SIV infection did not alter the frequency of KIR3DL05+ NK cells (Fig 7A), significant increases in the frequency of KIR3DL01+ NK cells were observed in GALT during acute and chronic infection (weeks 2 and 8) and in lymph nodes during chronic infection (week 8) (Fig 7B), which were mirrored by corresponding reductions in the frequency of KIR3DL01-05- NK cells in these tissues (Fig 7C). Further analysis of KIR expression revealed that, similar to peripheral blood, KIR3DL01 and KIR3DL05 are expressed by a higher percentage of CD16+ NK cells than CD56+ or CD16-CD56- NK cells in lymph nodes (Fig 7D and 7E); however, the increased frequency of KIR3DL01+ NK cells appeared to reflect the upregulation of this receptor on CD56+ and CD16-CD56- NK cells (Fig 7E). In the GALT, increases in the expression of KIR3DL01 (Fig 7G), but not KIR3DL05 (Fig 7F), on CD56+ and CD16-CD56- NK cells were more dramatic. Indeed, increases in the expression of KIR3DL01 on CD56+ and CD16-CD56- NK cells account almost entirely for the higher frequency of cells expressing this KIR after SIV infection (39.2%±7.6 and 24.9%±8.2, respectively) (Fig 7G). Representative data showing the preferential upregulation of KIR3DL01, but not KIR3DL05, on NK cells of the GALT at week 2 post-infection, and that the majority of these KIR3DL01+ cells are either CD56+ or CD16-CD56-, is provided in Fig 8.
KIR and HLA class I polymorphisms play a central role in the regulation of NK cell responses and can have a significant impact on the course of HIV-1 infection [17,19,23]. Studies to address the functional significance of KIR-MHC class I interactions in SIV-infected macaques have, however, been hampered by the lack of defined ligands for KIRs in non-human primates. To address this limitation, we and others have identified ligands for macaque KIRs [29,30,40,42,52]. We identified Mamu-A1*002, a common MHC class I molecule expressed in approximately 20% of Indian-origin rhesus macaques, as a ligand for KIR3DL05 [30], and found that nearly a third of the SIV peptides bound by Mamu-A1*002 suppress the cytolytic activity of KIR3DL05+ NK cells [16]. We further identified MHC class I molecules with a Bw4 motif as ligands for KIR3DL01, which is the most polymorphic and commonly expressed KIR in rhesus macaques [29]. To determine how these receptor-ligand interactions influence NK cell responses and the ability to contain virus replication, we infected twelve KIR- and MHC class I-defined rhesus macaques with SIV and monitored NK cell responses and viral loads in peripheral blood and lymphoid tissues.
As in humans, macaque KIRs are expressed in a variegated and stochastic fashion, which accounts for the presence of KIR3DL01 and KIR3DL05 on a fraction of total NK cells [29,30]. Unlike humans, however, which only express KIRs on functionally mature CD16+CD56dim NK cells, KIRs are expressed on all peripheral blood NK cell subsets in macaques [47]. KIR3DL01 and KIR3DL05 were accordingly detected on CD16+, CD56+ and CD16-CD56- NK cells. Nevertheless, these KIRs are expressed on a higher percentage of CD16+ cells than CD56+ or CD16-CD56- cells, further supporting a functional correspondence between the CD16+ NK population in macaques and the dominant CD16+CD56dim population in human blood.
The education of human NK cells is associated with an increase in the frequency of cells bearing KIRs that recognize self HLA class I ligands and a corresponding decrease in surface staining for those receptors [48–50]. Although neither the frequency of KIR3DL01+ or KIR3DL05+ NK cells, nor the intensity of KIR3DL01 staining, was associated with differences the number of MHC class I alleles predicted to encode ligands for these receptors, the level of KIR3DL05 staining correlated inversely with the presence of Mamu-A1*002 and -A3*13. This correlation appears to be driven primarily by Mamu-A1*002, since independent comparisons revealed significant differences in KIR3DL05 staining for Mamu-A1*002+ versus -A1*002- animals, but not for Mamu-A3*13+ versus -A3*13- animals; however, since KIR3DL05 staining was lower for Mamu-A3*13+ animals than for animals lacking both Mamu-A1*002 and -A3*13, and most of the Mamu-A1*002+ animals were also -A3*13+ as a consequence of linkage disequilibrium between these alleles [53], it is possible that Mamu-A3*13 may have a modest influence on KIR3DL05 levels that could be additive in combination with Mamu-A1*002. Nevertheless, a more dominant effect of Mamu-A1*002 on KIR3DL05 staining would be consistent with the unusually high avidity of Mamu-A1*002 for KIR3DL05 [30] and with higher levels of Mamu-A1*002 expression than for gene products of the ‘minor’ Mamu-A3*13 locus [51]. Thus, lower levels of KIR3DL05 staining for Mamu-A1*002+ macaques suggest a pronounced effect of Mamu-A1*002 on the education of KIR3DL05+ NK cells.
It should be noted, however, that the nature of NK cell education, or licensing, is not fully understood. It is possible that reduced levels of KIR3DL05 on NK cells from Mamu-A1*002+ animals reflect the binding of Mamu-A1*002 to KIR3DL05 on the same cells, similar to cis interactions previously described for MHC class I ligands of murine Ly49A [54] and human leukocyte Ig-like receptors (LILRs) [55,56]. Although to our knowledge cis interactions have not been reported for KIRs, it is tempting to speculate that high avidity interactions between Mamu-A1*002 and KIR3DL05 may sequester KIR3DL05, thereby reducing the accessibility of this receptor for staining on the cell surface. Such interactions would not necessarily be inconsistent with the functional effects of KIR ligands on NK cell licensing, since the sequestration of inhibitory KIRs may reduce the threshold required for NK cell activation through these receptors.
Although SIV peptides bound by Mamu-A1*002 were recently shown to suppress the cytolytic activity of Mamu-KIR3DL05+ NK cells, neither viral loads nor NK cell responses differed for Mamu-A1*002+ versus–A1*002- animals. Hence, these results did not reveal an advantage to SIV replication in Mamu-A1*002+ animals as a result of the presentation of inhibitory peptides to KIR3DL05+ NK cells, as in vitro assays with sorted primary NK cells might suggest [16]. However, these findings do not necessarily preclude a contribution of peptides to the evasion of NK cell responses. Rhesus macaques typically express six to thirteen different KIRs [36,40], and the ligands for most of these receptors remain undefined. Macaque NK cells also express other more conserved inhibitory and activating receptors, such as CD94/NKG2 heterodimers and natural cytotoxicity receptors (NCRs) that may influence responses to viral infection. Thus, it is possible that the effects of Mamu-A1*002-bound viral peptides on KIR3DL05+ NK cells may have been obscured by other receptor-ligand interactions. Furthermore, while lower levels of KIR3DL05 on NK cells from Mamu-A1*002+ animals suggest a dominant effect of Mamu-A1*002 on the education of KIR3DL05+ NK cells, other MHC class I molecules may also serve as ligands for this receptor. Therefore, given the complexity of KIR and MHC class I immunogenetics in rhesus macaques, and our limited knowledge of receptor-ligand interactions in this species, it is perhaps not surprising that we did not detect gross differences in viral loads or NK cell responses as a result of peptide-dependent modulation of a single KIR ligand.
The onset of adaptive immunity may also have complicated NK cell responses during chronic infection. Since previous studies have shown that the selection of CD8+ T cell escape variants in most of the SIV epitopes bound by Mamu-A1*002 generally occurs only after months of chronic infection [57–59], CD8+ T cell escape is unlikely to have affected KIR3DL05+ NK cell responses or viral loads during acute infection; however, we cannot exclude the possibility that the emerge of escape variants may have contributed to variability in KIR3DL05+ NK responses during chronic infection.
SIV infection nevertheless broadly stimulated NK cell responses. Within the first four weeks of infection, rapid increases in total NK cell counts, as well as cell counts for the CD16+ and CD16-CD56- populations, were observed in peripheral blood. In accordance with cross-sectional comparisons, these increases were maintained during chronic infection [44]. Similar changes were observed in NK cell populations defined by KIR3DL01 and KIR3DL05. Increases in KIR3DL01+, KIR3DL05+ and KIR3DL01+05+ NK cells during acute infection paralleled increases in total NK cells counts; however, while the number of KIR3DL05+ NK cells remained elevated during chronic infection, these increases were not sustained for the KIR3DL01+ subset. Longitudinal analyses revealed additional differences in the percentage of CD16+ NK cells in blood expressing KIR3DL01 versus KIR3DL05. Whereas the percentage of CD16+ NK cells expressing KIR3DL05 was relatively unchanged, there was a gradual decrease in the frequency of KIR3DL01+ CD16+ NK cells after eight weeks of infection. Although this decrease was partially offset by increases in the frequency of CD56+ and CD16-CD56- cells expressing this KIR, the decline of the larger KIR3DL01+CD16+ population probably accounts for lower overall KIR3DL01+ NK cell counts in blood during chronic infection.
Phenotypic analyses further revealed rapid increases in NK cell activation in response to SIV infection. Consistent with increases in circulating NK cell counts, the proliferation marker Ki-67 was highly upregulated during acute and chronic infection on KIR3DL01+, KIR3DL05+ and KIR3DL01-05- NK cells. Similar increases were also observed in the number of NK cells expressing the activation markers CD69 and HLA-DR, and the mucosal homing receptor α4β7. Markers of antiviral activity, including CD107a and TNFα, were likewise broadly upregulated during acute infection; however, the expression of these markers on KIR3DL01+ versus KIR3DL05+ NK cells diverged during chronic infection. Whereas CD107a and TNFα continued to be upregulated on KIR3DL01+ and KIR3DL01-05- cells, the percentage of KIR3DL05+ NK cells expressing these antigens declined to baseline levels by week six post-infection. These results reveal differential antiviral responses for KIR3DL05+ versus KIR3DL01+ NK cells, perhaps reflecting a greater role for the KIR3DL01+ subset in controlling SIV replication.
Characterization of NK cell subsets in lymph nodes and the gastrointestinal mucosa revealed additional changes in response to SIV infection. Although NK cells constitute a relatively minor percentage of lymphocytes in these tissues, a highly significant increase in the frequency of total NK cells was detected in lymph nodes during chronic infection. Consistent with previous studies [60], we found that the majority of NK cells in lymph nodes are CD16-CD56-, and of the smaller populations of CD16+ and CD56+ NK cells, the CD56+ subset is more prevalent in naïve animals. Following SIV infection, however, there was an inversion in the frequency of CD16+ versus CD56+ NK cells, with an accumulation of CD16+ NK cells and a corresponding decrease in the percentage of CD56+ NK cells. Significant increases were also detected in lymph node frequencies of KIR3DL05+ and KIR3DL01+ NK cells during acute and chronic infection, respectively.
In gut-associated lymphoid tissues, the percentage of total NK cells did not change in response to SIV infection, and with the exception of a transient increase in the frequency of the CD56+ subset, the relative proportions of CD16+, CD56+ and CD16-CD56- NK cells in this compartment also remained unchanged. However, dramatic changes were observed in the frequency of KIR3DL01+ NK cells. Significant increases in the percentage of KIR3DL01+ NK cells in GALT were detected during both acute and chronic infection. Surprisingly, most of these cells were CD56+ or CD16-CD56-, suggesting that this increase reflects the upregulation of KIR3DL01 on less mature NK cells rather than the accumulation of CD16+ NK cells expressing this KIR. Moreover, this response appears to be specific to KIR3DL01, since similar increases were not observed for KIR3DL05+ cells.
The enrichment of KIR3DL01+ NK cells in the gastrointestinal mucosa is intriguing, since the gut-associated lymphoid tissues are known to be a major source of HIV-1 and SIV replication and CD4+ T cell turnover [61–64]. Although rhesus KIR3DL01 and human KIR3DL1 are not orthologous, they may have similar functions; both recognize Bw4 ligands and are the most polymorphic KIRs of their respective species [29]. Moreover, the lysis of HIV-infected cells by KIR3DL1+ NK cells is primarily triggered by downmodulation of HLA-Bw4 ligands from the cell surface by the viral Nef protein [8,22], and we previously demonstrated that Mamu-Bw4 molecules are efficiently downmodulated by SIV Nef [65]. Thus, increases in the frequency of KIR3DL01+ NK cells in the GALT may reflect an innate cellular response to especially high levels of SIV replication and Bw4 downmodulation in these tissues.
To our knowledge, this study represents the first longitudinal analysis of NK cell responses in KIR- and MHC class I-defined macaques. Responses to SIV infection in peripheral blood were characterized by rapid increases in the CD16+ and CD16-CD56- populations, including KIR3DL01+ and KIR3DL05+ subsets. Markers of proliferation, activation and antiviral activity were widely expressed during acute infection, but began to diverge after four weeks, as indicated by sustained CD107a and TNFα upregulation by KIR3DL01+ NK cells, but not by KIR3DL05+ NK cells. Differential responses for KIR3DL01+ versus KIR3DL05+ NK cells were also evident in tissues. Whereas the percentages of KIR3DL05+ NK cells in lymph nodes and the gastrointestinal mucosa did not change, significant increases were observed in the frequency of KIR3DL01+ NK cells, especially in gut-associated lymphoid tissues. Thus, our results reveal broad NK cell activation and dynamic changes in multiple subsets in response to SIV infection, including an enrichment of KIR3DL01+ NK cells in mucosal tissues that represent major sites of ongoing virus replication and CD4+ lymphocyte depletion.
Twelve rhesus macaques (Macaca mulatta) of Indian origin, including six male and six female animals, were used in this study. Housing and care of the animals at the Wisconsin National Primate Research Center (WNPRC) were in compliance with the standards of the American Association for the Accreditation of Laboratory Animal Care and the University of Wisconsin Research Animal Resources Center (UWRARC). Animal experiments were approved by the UWRARC (protocol number G005496) and conducted in accordance with the principles described in the Guide for the Care and Use of Laboratory Animals [66]. Steps to improve animal welfare included environmental enrichment, such as foraging opportunities and manipulatable devices. Water was continuously available, commercial monkey chow was provided twice daily and fresh produce was supplied three times per week. Animals were sedated with ketamine HCl prior to the collection of blood and biopsy samples to minimize pain and distress associated with experimental procedures and were monitored twice daily by animal care and veterinary staff.
Twelve KIR3DL05+ rhesus macaques, including six Mamu-A1*002+ and six–A1*002- animals, were selected for this study. These included eleven animals expressing allotypes of KIR3DL01 with aspartic acid at position 233 (KIR3DL01 D233) [29] and excluded animals expressing the MHC class I alleles Mamu-A1*001, -B*008 and–B*17 associated with spontaneous control of SIV replication [58,67–69]. KIR3DL05+ and KIR3DL01 D233+ macaques were identified by staining PBMCs with Mamu-A1*002 Gag GY9 tetramer and the NKVFS1 monoclonal antibody as previously described [29,30]. All animals were also genotyped by sequencing full-length KIR transcripts as recently described [70]. Briefly, RNA was isolated from PBMCs, and full-length KIR cDNA was sequenced using a PacBio RS II instrument with P6-C4 sequencing reagents. Sequences were identified and novel alleles were classified by comparison to previously reported alleles of rhesus macaque KIRs. GenBank accession numbers for newly identified KIR alleles are listed in S1 Table. MHC class I genotyping was performed using genomic DNA isolated from PBMC by sequencing a 150 bp region of exon 2 (Illumina MiSeq system). Sequences were analyzed by comparison to an in-house database as previously described [71]. The MHC class I- and KIR-genotypes of each of the animals in this study are summarized in Tables 1 and 2.
Animals were infected intravenously with SIVmac239. A vial of SIVmac239 challenge stock prepared in activated rhesus macaque PBMC was provided by Dr. Ronald Desrosiers, Miller School of Medicine, University of Miami. On the day of inoculation, the vial was thawed and diluted to 50 animal infectious doses per ml in sterile, serum-free RPMI. Within 30 minutes of preparation, a one ml dose of the virus dilution (7.8 pg p27) was administered to each animal under ketamine anesthesia through a 22 g catheter placed aseptically in the saphenous vein.
Plasma was collected from blood drawn in tubes with EDTA as an anticoagulant and cryopreserved at -80°C. Virus was pelleted from 0.5 to 1.0 ml of plasma by ultracentrifugation for one hour at 20,000 x g. Viral RNA was extracted, reverse transcribed into cDNA, and quantified by real-time PCR using an assay based on amplification of a conserved SIV gag sequence as previously described [72].
Unless specified otherwise, all antibodies were purchased from BD Biosciences. All flow cytometry data were collected using a BD LSRII SORP and analyses were done with FlowJo 9.9 software (TreeStar Inc.).
A linear mixed-effect model was used for the analysis of longitudinal data. An individual macaque was included as a random-effect to account for correlation within subjects. The presence or absence of Mamu-A1*002 and the time points after infection were included as fixed-effects in the models. The three phases of the infection were naïve or pre-infection (week 0), acute infection (weeks 1–4) and chronic infection (weeks 5–24). Two-sided p-values less than 0.05 were considered statistically significant. For the comparison of discrete data, Mann-Whitney U tests were performed using GraphPad Prism V6g.
|
10.1371/journal.pgen.1002240 | Evidence for Hitchhiking of Deleterious Mutations within the Human Genome | Deleterious mutations present a significant obstacle to adaptive evolution. Deleterious mutations can inhibit the spread of linked adaptive mutations through a population; conversely, adaptive substitutions can increase the frequency of linked deleterious mutations and even result in their fixation. To assess the impact of adaptive mutations on linked deleterious mutations, we examined the distribution of deleterious and neutral amino acid polymorphism in the human genome. Within genomic regions that show evidence of recent hitchhiking, we find fewer neutral but a similar number of deleterious SNPs compared to other genomic regions. The higher ratio of deleterious to neutral SNPs is consistent with simulated hitchhiking events and implies that positive selection eliminates some deleterious alleles and increases the frequency of others. The distribution of disease-associated alleles is also altered in hitchhiking regions. Disease alleles within hitchhiking regions have been associated with auto-immune disorders, metabolic diseases, cancers, and mental disorders. Our results suggest that positive selection has had a significant impact on deleterious polymorphism and may be partly responsible for the high frequency of certain human disease alleles.
| Deleterious mutations reduce fitness within natural populations and must be continually removed by natural selection. However, some deleterious mutations reach unexpectedly high frequencies. There are a number of mechanisms by which this could occur, including changes in genetic or environmental constraints. Here, we investigate the hypothesis that some deleterious mutations have hitchhiked to high frequency due to linkage to sites that have been under positive selection. Using a collated set of regions likely to have been influenced by positive selection, we find that the number of deleterious polymorphisms in hitchhiking and non-hitchhiking regions is similar, but that the ratio of deleterious to neutral polymorphism is higher in hitchhiking compared to non-hitchhiking regions. Both computer simulations and empirical data indicate that while hitchhiking eliminates many deleterious mutations, some are increased in frequency. The distribution of human disease-associated mutations is also altered in hitchhiking compared to non-hitchhiking regions. Together, our results provide evidence that hitchhiking has influenced the frequency of linked deleterious mutations in humans, implying that the evolutionary dynamics of advantageous and deleterious mutations may often depend on one another.
| The continuous removal of deleterious mutations is essential to maintaining a species' reproductive output and even its existence. While deleterious mutations incur a considerable fitness cost [1], they are not always effectively removed from a population. Deleterious mutations are more difficult to remove from small populations and their accumulation can lead to further reductions in population size and eventually to extinction, a process called mutational meltdown [2]–[4]. Sexual recombination facilitates the elimination of deleterious mutations [5] and the lack of recombination on the Y sex chromosome may have contributed to its degeneration through the accumulation of deleterious mutations [6].
In humans, many deleterious mutations have reached high population frequencies. Each human is estimated to carry on the order of 1,000 deleterious mutations in their genome [7]–[9]. Although most deleterious mutations are rare, a significant fraction is common in the population. For example, 19% of deleterious mutations identified in three human genomes are common enough to be shared among them [9]. However, the cause and consequence of common deleterious mutations have been difficult to determine.
A number of factors may contribute to the large number of common deleterious mutations in humans. Most genome-wide methods used to identify deleterious mutations are based on the alteration of sites that are significantly conserved across species [10], [11]. As such, lineage-specific changes in selective constraint provide one explanation for common alleles that alter highly conserved sites.
Changes in selective constraint can be caused by changes in population size, the environment, or other genetic changes [12]. Because the efficacy of selection is a function of effective population size, a reduction in population size can result in reduced constraints on sites that are conserved in other species [13]. Many common deleterious mutations in humans can be attributed to the small effective population size of humans and recent human population bottlenecks [14], [15]. However, changes in constraint can also be mediated by genetic or environmental changes. For example, the thrifty gene hypothesis posits that the high frequency of diabetes risk alleles is a consequence of their being previously advantageous during periods of food scarcity [16]. Relaxed constraints may also arise due to certain types of genetic changes, such as gene duplication or compensatory mutations. The observation that human disease alleles are often present in mouse supports the notion that the selective constraints on a site are not always static but can change with the genetic or environmental background [17]. However, not all common deleterious mutations may result from species-specific differences in selective constraint.
Positive selection can influence the frequency of deleterious mutations directly, through genetic hitchhiking, or indirectly, through a reduction in effective populations size mediated by an increase in the variance of reproductive success [18]. As a consequence, positive selection can increase the rate at which deleterious mutations accumulate, particularly when the effect of the advantageous mutation outweighs the effects of linked deleterious mutations [19]–[22]. Hitchhiking of deleterious mutations along with advantageous mutations may have contributed to the degeneration of the Y sex chromosome [20], [23] and the increased number of deleterious mutations present in domesticated species [24], [25].
In this study, we examined the effect of positive selection on linked deleterious polymorphism in the human genome. We compared the abundance of deleterious and neutral nonsynonymous single nucleotide polymorphisms (SNPs) in regions showing evidence of hitchhiking to other genomic regions. While hitchhiking is expected to remove neutral variation from a population [26], we find that the rate of deleterious SNPs is not reduced, resulting in an enrichment of deleterious relative to neutral SNPs in hitchhiking regions. Our results imply that positively selected mutations may often influence the frequency of linked deleterious mutations.
To characterize the effect on positive selection on linked deleterious mutations we conducted simulations under a Wright-Fisher model. Subsequent to a single hitchhiking event, the rate of neutral and deleterious polymorphism was reduced as a function of the rate of recombination (Figure 1A). Despite the overall reduction in the number of deleterious polymorphisms, at intermediate rates of recombination, hitchhiking caused an increase in the number of high frequency deleterious polymorphisms, as measured by θH (Figure 1A), similar to its effect on neutral polymorphism [27]. Compared to deleterious polymorphism, hitchhiking caused a greater reduction in neutral polymorphism, resulting in an enrichment of deleterious relative to neutral polymorphism. The enrichment was greatest for high compared to intermediate and low frequency polymorphism, as measured by θH, θπ, and θW, respectively (Figure 1D). Because the reduction in fitness due to deleterious polymorphism remained constant during hitchhiking, the cost of increasing the frequency of some deleterious alleles to high frequency must be offset by the elimination of other deleterious alleles.
To examine the average effect of multiple hitchhiking events we also simulated populations under a continuous influx of advantageous mutations. Similar to single hitchhiking events, recurrent hitchhiking reduced the rate of neutral and deleterious polymorphism (Figure 1C), and increased the ratio of deleterious to neutral polymorphism (Figure 1B). While the degree to which hitchhiking caused an enrichment of deleterious polymorphism depended on the strength of positive and negative selection and the rate of advantageous and deleterious mutation (Figure S1), our simulations indicate that hitchhiking may often have a measurable impact on the ratio of deleterious to neutral polymorphism segregating in natural populations.
To examine the impact of positive selection on deleterious polymorphism in humans we classified nonsynonymous SNPs from the 1000 Genomes Project [28] as neutral or deleterious using a likelihood ratio test based on cross-species conservation (Materials and Methods). Although not all classifications may be correct, the likelihood ratio test classifies 72% of human disease mutations as deleterious and only 6.7% of nonsynonymous substitutions between species as deleterious [9]. Out of 48,558 autosomal nonsynonymous SNPs tested, 14,094 (29.0%) were predicted to be deleterious, of which 2,263 (16.1%) have a derived allele frequency of over 10%. Using a cutoff of 10%, the fraction of SNPs called deleterious is 17.8% for common alleles compared to 33.0% for rare alleles, consistent with the expected effects of negative selection.
Hitchhiking is expected to have a stronger effect on linked variation in regions of low recombination [26]. While the spread of a positively selected allele through a population causes a reduction in the amount of linked neutral variation, it may interfere with the elimination of linked deleterious mutations. Consistent with this hypothesis, the rate of synonymous and neutral nonsynonymous SNPs decreases in regions of low recombination, whereas the rate of deleterious SNPs remains nearly constant (Figure 2A). As a consequence, the ratio of deleterious to neutral and deleterious to synonymous SNPs is significantly correlated with the rate of recombination (P = 3.1×10−15 and P<2.0×10−16, respectively, Figure 2B). The association remains significant when accounting for the frequency of conserved codons and biased gene conversion (P = 2.1×10−7 and P = 9.3×10−6, respectively, Figure S2), which are also correlated with the rate of recombination. However, this correlation is also expected due to background selection, which reduces the efficacy of selection against deleterious mutations [29], [30].
In contrast to background selection, which exerts more uniform effects across the genome [31], hitchhiking can generate strong local effects. Furthermore, hitchhiking can have large effects in regions of both low and high recombination whereas background selection is expected to have much smaller effects in regions of high recombination [32].
To determine whether deleterious SNPs have been influenced by recent episodes of positive selection, we examined genomic regions showing evidence of hitchhiking based on multiple tests of selection [33]. In hitchhiking regions defined by two or more tests of selection, we found a significantly higher ratio of deleterious to neutral SNPs compared to other genomic regions (Figure 3 and Table S1). The elevated ratio of deleterious to neutral SNPs within hitchhiking regions cannot be explained by a reduced rate of recombination or a higher density of conserved sites; the difference between hitchhiking and non-hitchhiking regions remained significant using a logistic regression model with these factors as covariates (P = 6.3×10−5, Figure S3). The increase in the ratio of deleterious to neutral SNPs in hitchhiking relative to non-hitchhiking regions is 1.09-fold for regions identified by two or more tests of selection and increases to 1.87-fold for regions identified by all nine tests of selection. The increase in the ratio of deleterious to neutral SNPs in hitchhiking regions is due to a decrease in the number of neutral SNPs rather than an increase in the number of deleterious SNPs (Figure 3B and 3C). With the exception of the composite likelihood ratio test (CLR) [34], all of the methods used to detect hitchhiking identify regions with a higher ratio of deleterious to neutral SNPs (Figure 3D). Thus, the increase in the relative abundance of deleterious SNPs in hitchhiking regions does not appear to be associated with any specific test of selection.
The effects of hitchhiking are expected to decline as a function of recombinational distance from the site under selection [26]. To examine the decay in the number of deleterious SNPs associated with hitchhiking, we used iHS [35] and Rsb [36] defined hitchhiking regions. iHS is better at detecting incomplete hitchhiking events [35], where the advantageous mutations is still segregating in the population, whereas Rsb is better at detecting complete or nearly-complete episodes of selection [36]. The frequency of deleterious SNPs decreases as a function of distance from iHS defined hitchhiking region (P = 2×10−7, Figure 4A). Compared to iHS regions, the frequency of deleterious SNPs shows a more modest decline with distance from the Rsb defined hitchhiking regions (P = 0.018, Figure 4B). This difference could result from Rsb detecting older hitchhiking events providing additional time for negative selection to eliminate linked deleterious mutations or due to a weaker influence of hitchhiking outside of Rsb defined regions, which are twice as large as iHS defined regions (Table S1).
As the rate of recombination decreases, hitchhiking causes a larger increase in the ratio of deleterious to neutral SNPs for common compared to low frequency SNPs (Figure 1). To determine whether hitchhiking regions show a similar pattern, we compared the ratio of deleterious to neutral SNPs as a function of allele frequency. Similar to the simulation results, the ratio of deleterious to neutral SNPs declines with increasing allele frequency. However, the ratio of deleterious to neutral SNPs in hitchhiking regions is not significantly different among three frequency classes (Table 1). We observed the same pattern using HapMap SNPs (data not shown) indicating that low coverage sequencing errors in the 1000 Genomes Project is unlikely to explain this result. Although the absence of a difference in the ratio of deleterious to neutral SNPs across allele frequencies is somewhat surprising, it is consistent with simulations that have a high rate of recombination or strong negative selection (Figure 1 and Figure S1).
Many of the methods used to detect hitchhiking were independently applied to populations of different ancestry. Although some hitchhiking events may be specific to European, African, or Asian populations, e.g. [35], the power to detect hitchhiking is expected to differ among populations even when an adaptive mutation is fixed in all populations [37], [38]. We examined the enrichment of deleterious SNPs in iHS defined hitchhiking regions in the European, African, and Asian samples. Surprisingly, we found no enrichment of deleterious SNPs in African and Asian defined hitchhiking regions (Table S2). Despite these population-specific differences revealed by iHS, the ratio of deleterious to neutral SNPs is elevated in hitchhiking regions defined by multiple methods in the African, European and Asian samples (Table S3).
For most hitchhiking regions the target of selection is not known. We identified ten hitchhiking regions from the literature for which there is evidence for the target of selection. The putative targets are LCT [39], [40], SLC45A2 [41], TYRP1 [42], HERC2 [42], KITLG [42], SLC24A5 [43], TYR [44], EDAR [41], PCDH15 [42] and LEPR [42]. Within these genes the ratio of deleterious to neutral SNPs (1.83) is higher than in non-hitchhiking regions (0.41) (Fisher's Exact Test P = 0.0023, Table 2). The deleterious SNPs include 5/6 nonsynonymous SNPs that are putative targets of selection. Within the 1 Mbp regions flanking these genes, there is also a higher ratio of deleterious to neutral SNPs (0.69) relative to that in non-hitchhiking regions (0.41) (Fisher's Exact Test P = 0.034).
Positive selection at SLC45A2 and TYR is particularly interesting since linked deleterious SNPs have been associated with human disease. The putative target of selection on TYR is a nonsynonymous SNP (S192Y) that has an allele frequency of 42% in the European sample (CEU) and is associated with the absence of freckles in Europeans [44]. Another nonsynonymous SNP in TYR (R402Q), 106 kb away, is classified as deleterious, has a frequency of 21% in CEU and is associated with mild ocular albinism and risk for cutaneous melanoma and basal cell carcinoma [45], [46]. The putative target of selection on SLC24A5 is a nonsynonymous SNP (A111T) that is associated with skin pigmentation and is nearly fixed in European populations but is at low frequency in African and Asian populations [43]. Positive selection on this allele may have influenced the frequency of deleterious SNPs in FBN1, 265 kb downstream of SLC24A5. FBN1 has five deleterious SNPs in HapMap CEU, all of which are present at low frequency in CEU, 0.5–1.4%, but are absent from both the African or Asian HapMap samples. Three of these deleterious SNPs cause Marfan syndrome [47], [48] and one has been found in patients with Marfan syndrome or related phenotypes [49].
Hitchhiking may have also influenced SNPs that are associated with human disease. This might occur by increasing the frequency of rare, disease-causing mutations or by increasing the frequency of more common, disease-risk alleles. To investigate this possibility we compared the abundance of disease-associated alleles in hitchhiking and non-hitchhiking regions.
Within known disease genes in OMIM, there are 9,481 mutations that have been associated with human disease, of which 1,722 were common enough to be typed in the HapMap project and can be considered SNPs. The ratio of all OMIM variants in hitchhiking relative to non-hitchhiking regions (0.053) is lower than that of the number of OMIM morbid genes (0.071), consistent with the elimination of variation within hitchhiking regions (Table S4). However, the ratio of common OMIM variants in hitchhiking to non-hitchhiking regions, 0.079, is significantly higher than that of rare variants, 0.047 (Fisher's Exact Test, P<10−5, Figure 5). This difference is opposite to that found for neutral HapMap SNPs, which are skewed towards rare alleles in hitchhiking relative to non-hitchhiking regions. Furthermore, the minor allele frequencies of OMIM SNPs is slightly higher in hitchhiking compared to non-hitchhiking regions (Wilcoxon Rank Sum Test, P = 0.03). Similar to OMIM SNPs, the ratio of disease-associated SNPs in hitchhiking relative to non-hitchhiking is higher for common compared to rare alleles identified in the 1000 Genomes Project, although the difference is not significant (Figure 5, Fisher's Exact Test, P = 0.20). For the 1000 Genomes Project data, the mean frequency of common disease alleles in hitchhiking regions (0.25) is higher than that in non-hitchhiking regions (0.20), although the difference is not significant (Wilcoxon Rank Sum Test, P = 0.80). Thus, hitchhiking regions appear to be characterized by an increase in the number common disease-associated SNPs rather than by an increase in the number of rare, disease-associated variants.
To examine the abundance of common, risk-associated alleles within hitchhiking regions, we used alleles that have been associated with human disease from genome-wide association studies (GWAS) [50] and from a literature survey (see Materials and Methods). Consistent with a previous study [50], the ratio of risk-alleles identified by GWAS in hitchhiking to non-hitchhiking regions, 0.059, is not greater than that expected based on the number of genes, 0.068 (Table S4). However, nonsynonymous risk alleles, which are likely enriched for functional variants, have a higher hitchhiking to non-hitchhiking ratio than that of other risk-alleles (Figure 5, Fisher's Exact Test, P = 0.02). Although risk alleles in hitchhiking regions do not have significantly higher allele frequencies than those in non-hitchhiking regions (Wilcoxon Rank Sum Test, P = 0.63), the proportion of risk alleles with odds ratios over 2.0 in hitchhiking regions (18.9%) is significantly higher than that in non-hitchhiking regions (11.5%) (Fisher's Exact Test, P = 0.03). For disease-associated nonsynonymous SNPs identified in a literature survey, the ratio of SNPs in hitchhiking to non-hitchhiking regions is lower than that of neutral SNPs (Table S4).
To identify which types of diseases hitchhiking may have influenced, we examined disease-associated SNPs and genes with deleterious SNPs within hitchhiking regions. Classification of the 126 OMIM SNPs within hitchhiking regions by phenotype (Table S5) revealed a number of SNPs involved auto-immune disorders (21 SNPs), energy metabolism (16 SNPs), and a variety of mental, neurological, and neurodevelopmental disorders (25 SNPs). Classification of the 461 genes (Table S6) within hitchhiking regions that contain deleterious SNPs by their disease association revealed a number that have been associated with cardiovascular (N = 21), immune (N = 19), metabolic (N = 18), neurological (N = 12) and psychiatric disease (N = 10), and cancer (N = 17), according to the Genetic Association Database classification [51]. Classification of the 12 nonsynonymous SNPs identified by GWAS and the three nonsynonymous SNPs identified from the literature revealed five associated with auto-immune disease, three associated with metabolic disease, and two associated with cancer. However, none of these disease classifications are significantly different from those outside of hitchhiking regions.
Most deleterious SNPs lie outside of currently defined hitchhiking regions. However, this does not exclude the possibility that they were influenced by positive selection. The overlap among methods used to detect hitchhiking is low [33], and some hitchhiking events may not be detected by any of the methods. For example, a beneficial mutation may initially spread slowly through a population while it becomes disentangled from linked deleterious mutations. In this scenario, patterns of hitchhiking may be weak or absent, similar to those that occur when positive selection acts on standing genetic variation [52]. To characterize genomic regions enriched for deleterious SNPs, we split the genome into 1 Mbp windows and selected the top 2% of windows with the highest rate of deleterious SNPs per kb of coding sequence.
Regions enriched for deleterious SNPs have a high ratio of deleterious to neutral nonsynonymous SNPs, 0.66, much higher than the genome average, 0.41. Together, these 43 regions contain 7.4% of deleterious SNPs (Table S7). 17 of these regions show evidence of hitchhiking, ten with evidence from three or more tests of selection. In addition, one region may have been influenced by positive selection on DARC [53], even though it does not overlap with hitchhiking regions defined by multiple tests of selection [33]. Ten of the regions contain deleterious SNPs in multiple duplicated olfactory receptor or keratin genes. Of the remaining 21 regions, 16 have deleterious SNPs in more than two genes. While loss of constraint may explain the accumulation of deleterious SNPs in some genes, particularly those that are duplicated, it is less likely to explain deleterious SNPs in multiple linked genes with disparate functions.
Deleterious mutations have a significant impact on a species' ability to survive, reproduce and adapt to new environments [2]–[4]. In humans, there is an abundance of common nonsynonymous SNPs that disrupt sites highly conserved across species and likely to be deleterious [9]. By examining the genome distribution of nonsynonymous SNPs classified as either neutral or deleterious, we found a greater reduction in neutral compared to deleterious polymorphism within genomic regions likely to have been influenced by hitchhiking. This observation combined with hitchhiking simulations suggests that while many deleterious SNPs are eliminated due to hitchhiking, a substantial number of rare deleterious mutations must also increase to frequencies common enough to be considered polymorphic. Our results imply that positive selection is not responsible for the abundance of common deleterious SNPs across the human genome but is relevant to understanding the distribution and dynamics of deleterious mutations as well as certain disease alleles.
Despite evidence for a hitchhiking effect, most common deleterious SNPs are unlikely to have been influenced by positive selection and are better explained by a change in selective constraint, mediated by a population bottleneck [15] or environment change [54]. Only 11.5% of deleterious SNPs occur in regions showing evidence of hitchhiking (Table S1). However, this does not exclude the possibility that positive selection has influenced the frequency of some deleterious SNPs outside of hitchhiking regions. Hitchhiking regions were defined by the overlap of two or more methods of detecting selection and are unlikely to include all regions influenced by hitchhiking [33]. In support of this possibility, we identified a number of genomic regions that contain an exceptionally high ratio of deleterious to neutral SNPs. Although some of these regions include multiple duplicated genes, which could explain the large number of SNPs predicted to be deleterious, one of the regions includes a gene thought to have been under selection, DARC [53], and many of the regions contain deleterious SNPs in genes with disparate functions.
Within hitchhiking regions, we found an elevated ratio of deleterious to neutral SNPs caused by a reduction in the number of neutral SNPs. The elevated ratio of deleterious to neutral SNPs is consistent with simulations of both single and recurrent hitchhiking events across a range of parameters (Figure 1 and Figure S1) and can be explained by the difference in the frequency distribution of deleterious and neutral SNPs prior to hitchhiking. During a hitchhiking event neutral and deleterious alleles increase or decrease in frequency depending on their original configuration with the advantageous mutation. However, rare alleles are more likely to be deleterious and common alleles are more likely to be neutral. Thus, positive selection removes many common alleles, which tend to be neutral, and increases the frequency of many rare alleles, which tend to be deleterious, resulting in an increase in the ratio of deleterious to neutral SNPs. However, the simulated hitchhiking events showed two patterns that were not observed in the human data. First, hitchhiking caused a reduction in the number of deleterious SNPs. Second, hitchhiking caused a much larger increase in the ratio of deleterious to neutral SNPs at high frequencies relative to that at low frequencies. The significance of these differences is hard to evaluate since many factors known to influence hitchhiking were not examined, e.g. dominance, population structure, changes in population size and selection on new mutations versus standing genetic variation. Furthermore, hitchhiking simulations with high rates of recombination or strong selection against deleterious mutations tended to show patterns that are more consistent with those observed in humans (Figure 1 and Figure S1). Although some theoretical results have recently been obtained [22], further work will be needed to understand the effects of hitchhiking on deleterious mutations in humans.
A number of factors besides hitchhiking may contribute to the increased ratio of deleterious to neutral SNPs. Background selection is expected to increase the ratio of deleterious to neutral SNPs, particularly within regions of low recombination (Figure 1). While the rate of recombination can explain some of the difference between hitchhiking and non-hitchhiking regions, the ratio of deleterious to neutral SNPs is significantly higher in hitchhiking regions even after controlling for differences in recombination rate between hitchhiking and non-hitchhiking regions. Given the slightly lower rates of recombination in hitchhiking regions, the logistic regression model predicts hitchhiking regions should have a ratio of deleterious to neutral SNPs of 0.46, which is only slightly higher than that in non-hitchhiking regions, 0.44, and less than that observed, 0.53. It is conceivable that background selection may exert much weaker effects over shorter intervals that are not related to regional rates of recombination. However, weak background selection would have to exert a stronger influence within hitchhiking compared to non-hitchhiking regions, making it difficult to attribute the increased ratio of deleterious to neutral SNPs within these regions to background selection alone.
Another factor that complicates the analysis of differences between hitchhiking and non-hitchhiking regions is how hitchhiking regions were defined. Hitchhiking regions were defined by genome scans for patterns of variation expected to occur as a result of positive selection. However, some regions identified in genome scans for selection are likely neutral outliers that by chance show patterns of variation similar to those created by hitchhiking. This was one of our main motivations for using hitchhiking regions defined by two or more genome scans for selection. Although a contribution from neutral outliers cannot be excluded, the observation that the ratio of deleterious to neutral SNPs is 1.87-fold higher in regions identified by all nine genome scans and 1.68-fold higher in regions containing genes known to have been under positive selection suggests that hitchhiking makes a significant contribution to the elevated ratio of deleterious to neutral SNPs.
Similar to deleterious SNPs, common, disease-associated SNPs are enriched in hitchhiking compared to non-hitchhiking regions. In contrast, the number of rare, disease-associated mutations in hitchhiking relative to non-hitchhiking regions is lower than that of OMIM morbid genes. This difference can be explained by hitchhiking. Since most rare disease mutations occur on different chromosomes, hitchhiking will increase the frequency of one or a small number of disease mutations but decrease or eliminate the majority of rare disease mutations. However, the difference between rare and common disease-associated alleles is complicated by the heterogeneous evidence used to define disease-associated mutations in OMIM and the fact that common mutations are more likely to be associated with disease than rare mutations. The effect of hitchhiking on GWAS SNPs is more complex since most GWAS SNPs may be neutral. The ratio of GWAS SNPs in hitchhiking to non-hitchhiking regions is lower than that of all genes or neutral SNPs (Table S4). The lower frequency of GWAS SNPs in hitchhiking to non-hitchhiking regions is consistent with a previous study [50] and may be caused by the removal of common SNPs and reduced power of linkage disequilibrium-based tests of association. Consistent with this possibility, the hitchhiking to non-hitchhiking ratio of GWAS SNPs that are nonsynonymous, and thus more likely to be causative, is higher than that of all GWAS SNPs.
Our results also bear on the incidence of certain human diseases [55], [56] and disease alleles [57], which in some cases are higher than what one might expect based on disease severity. While genetic drift and population bottlenecks are likely to contribute to common disease alleles, balancing selection has also been invoked in some instances. For example, the high frequency of the delta F508 mutation in CFTR has been hypothesized to be the result of a heterozygote advantage due to cholera resistance [58], [59]. Mutations in G6PD and Beta-globin have been hypothesized to provide a heterozygote advantage due to malaria resistance [57]. Another explanation for why some disease alleles are so common is the ancestral-susceptibility hypothesis, under which derived alleles associated with human disease were advantageous to ancestral lifestyles and environmental conditions [54]. Similarly, under the less is more model, loss of function mutations that were previously disadvantageous can become advantageous [60]. In support of this model, we found five out of six nonsynonymous SNPs that are putative targets of positive selection are highly conserved across species and so classified as deleterious.
However, our results also provide evidence for an alternative explanation for the frequency of common disease-associated alleles: the frequency of certain disease alleles is increased due to hitchhiking with linked advantageous mutations. A number of previous observations support this explanation. The MHC locus has been associated with over 40 human genetic diseases [61], and multiple lines of evidence suggest long-term balancing selection [62]. A mutation in HFE that causes hemochromatosis is 150 kb away from a hitchhiking region and may have increased in frequency due to hitchhiking [63]–[65]. Hitchhiking has also been implicated in the increased frequency of a common risk haplotype for diabetes, hypertension and celiac disease [66] and another risk haplotype for Crohn's disease [67]. Intriguingly, the delta F508 mutation in CFTR is one of the most common disease-causing alleles in Caucasians, with an estimated allele frequency of 1.4% [68], and CFTR occurs within a hitchhiking region. Four of the HapMap nonsynonymous SNPs within CFTR are classified as deleterious, one of which has been associated with infertility [69]. One of the regions with the strongest evidence for hitchhiking (7 tests) also has one of the highest ratios of deleterious to neutral SNPs (16/22, Table S7). Within this region, 8/16 deleterious SNPs occur in BLK, NEIL2, and CTSB, and there are three disease alleles in the Human Gene Mutation Database [70], with frequencies of 0.8%, 5.3% and 44% based on the 1000 Genomes Project. The frequency of these deleterious/disease alleles may have been influenced by positive selection in this region.
The interaction between positive and negative selection makes it difficult to isolate and understand the effects of each individually. In the presence of deleterious mutations, the effect of hitchhiking on linked neutral variation may be reduced compared to that which would occur in the absence of deleterious mutations, similar to patterns created by soft sweeps [52]. Conversely, hitchhiking increases the frequency of some deleterious mutations and decreases the frequency of others such that the distribution of deleterious mutations is significant different from that expected in the absence of hitchhiking. Furthermore, the recent expansion in human population size combined with population subdivision may amplify or reduce the influence of hitchhiking on deleterious SNPs. This will make it valuable to examine the extent to which deleterious alleles are enriched in hitchhiking regions in other species, particularly domesticated species where the strength of selection was likely strong and for which targets of selection are in some cases known.
The effects of hitchhiking on deleterious and neutral polymorphism were simulated using a Wright-Fisher model [71]. Simulated populations had a size, N, of 1000 diploid individuals. Mutations were distributed into the population assuming an infinite sites model with a Poisson rate of 2Nu, where u is the mutation rate per chromosome. A Poisson number of recombination events was generated in the population with a rate of Nr, where r is the rate of recombination per individual. Chromosomes in the next generation were sampled based on the fitness of the individual from which they were derived. Fitness was calculated by the multiplicative effects of each non-neutral allele, 1+hs for heterozygous sites and 1+s for homozygous sites, where s is the selection coefficient and h is the degree of dominance. The dominance coefficient was 0.5 for all simulations. For each set of parameters, simulations were run for 20N generations before sampling. For a single hitchhiking event, an advantageous mutation was generated in the center of the chromosome and sampled at the end of hitchhiking conditional on its fixation. For multiple hitchhiking events, advantageous mutations were generated at a constant rate uniformly across the chromosome and samples were taken in intervals of N generations. θW, θπ and θH were estimated using a sample size of 100 chromosomes as described in [27].
Low-coverage SNP calls for CEU, CHB+JPT, and YRI samples were downloaded from the 1000 Genomes Project (release 2010_07) [28], and all tri-allelic sites were filtered out. Coding SNPs were identified based on their genomic coordinates in the NCBI reference genome (build 36) and Ensembl known genes (release #49). After eliminating SNPs on the sex chromosomes, SNPs in known pseudogenes or gene fragments, and sites monomorphic across CEU, CHB+JPT and YRI samples, there were 47,730 synonymous and 48,558 nonsynonymous SNPs within coding regions with multi-species alignments used by the likelihood ratio test (see below).
Nonsynonymous SNPs were classified as neutral or deleterious using a previously implemented likelihood ratio test (LRT) for conservation across multiple species [9]. The LRT is based on 18,993 multiple sequence alignments from 32 vertebrate species. Positions with less than 10 aligned eutherian mammals were excluded from the analysis due to low power of the LRT. At each codon in the alignment, the LRT calculates the likelihood of the data under a neutral model, where the nonsynonymous substitution rate (dN) equals the synonymous substitution rate (dS), relative to a conserved model, where dN can deviate from dS. For these calculations, dS is set to an average rate of 12.2 substitutions per site across the entire tree based on an estimate from gap-free concatenated alignments of 1,227 genes (54 kb) with data from all species. Nonsynonymous SNPs were predicted to be deleterious if: 1) the codon is significantly conserved by the LRT (P<0.001), 2) dN is less than dS, and 3) the derived amino acid is not present at orthologous positions in other eutherian mammals.
The density of SNPs was measured as a function of local recombination rate using CEU, CHB+JPT, and YRI SNPs from the 1000 Genomes Project. Following previous work [72], recombination rates were estimated from non-overlapping 400 kb windows by dividing the genetic map distance of the two most distant SNPs by their physical distance. The genetic map, estimated by LDhat [73], was obtained from the 1000 Genomes Project. Windows that were less than 10 Mb away from the end of centromeres and telomeres, windows without a pair of SNPs greater than 360 kb apart, and windows with no aligned coding sequence were excluded. The remaining 3,666 windows were assigned into ten equal-sized bins by their recombination rates, and the number of synonymous, nonsynonymous deleterious and nonsynonymous neutral SNPs was counted per kb of aligned coding sequence in each bin. To account for the the proportion of codons that are conserved, which is correlated with both the rate of recombination and the number of G or C nucleotides within codon (Figure S2A), codons in each recombination bin were subdivided into four classes by the number of GC nucleotides within the human codon (j = 0, 1, …, 3). In cases of polymorphic codons, GC content of the ancestral codon were counted. A total of 6,248,078 codons were classified as significantly conserved or not by the LRT at a P-value cutoff of 0.001. The relationship between recombination and the ratio of deleterious to neutral SNPs was assessed using the logistic regression model:where DELi, j, and NEUi, j, are the number of deleterious and neutral nonsynonymous SNPs, respectively, ri is the average recombination rate of windows in bin i, and si, j adjusts for differences in the number of potentially deleterious sites. si, j was estimated by:where fconi, j is the fraction of conserved codons out of all aligned codons with j GC nucleotides in bin i, fconj is the mean of fconi, j over all i = 1, …, 10, and fdelj is the fraction of deleterious out of all tested nonsynonymous SNPs with the same j GC nucleotides.
To account for biased gene conversion, which has been previously proposed to explain a higher rate of GC-biased disease alleles in regions of higher recombination [74], we re-examined the relationship between the ratio of deleterious to neutral SNPs and recombination after excluding 13,995 AT-to-GC mutating SNPs potentially affected by biased gene conversion. SNPs within codons with zero GC nucleotides were also eliminated due to their relatively small number (N = 335). Using the logistic regression model that accounts for the variation in the number of potentially deleterious sites, the regression coefficient β1 of recombination rate remained similar (−0.097 to −0.101) and highly significant (P = 9.3×10−6).
Hitchhiking regions were defined by genomic intervals that were identified by two or more out of nine tests for hitchhiking, using intervals rounded to the nearest multiple of 10 kbp [33]. To compare different methods, we examined regions that were identified by one method and overlapped with any other method. Non-hitchhiking regions were defined as autosomal regions excluding hitchhiking regions as defined above. The density of deleterious and neutral nonsynonymous SNPs was measured relative to the accessible portion of aligned coding regions used for the likelihood ratio test. The accessible genome, which satisfies minimum read depth required for SNP calling, was obtained from the 1000 Genomes Project for CEU, CHB+JPT, and YRI [28], and their union was used for the combined analysis of all samples. The difference between the SNP density within hitchhiking and non-hitchhiking regions was tested by a two-proportion z-test.
To test whether a higher ratio of deleterious to neutral SNPs in hitchhiking relative to non-hitchhiking regions is caused by a higher recombination rate or a larger number of potentially deleterious sites in hitchhiking regions, the 400-kb genomic windows which were already binned by the rate of recombination and the number of GC nucleotides in a codon were further classified into hitchhiking and non-hitchhiking groups. After removing windows near centromeres and telomeres, there were 388 windows identified by three or more tests of hitchhiking that were assigned to the hitchhiking group (h = 1), and 2,917 windows without any hitchhiking regions that were assigned to the non-hitchhiking group (h = 0). The data were fit to the following logistic regression model:where ri is the rate of recombination, si,j,h adjusts for the density of conserved codons and h is an indicator variable for hitchhiking windows.
To study the decay of the ratio of deleterious to neutral SNPs as a function of distance from hitchhiking regions, we used regions identified in CEU by iHS [35] and Rsb [36]. For iHS, the top 5% of scanned genomic windows (a total of 127.6 Mb) were used as hitchhiking regions, as described below. For Rsb, we used regions identified in CEU in comparison to both YRI and CHB+JPT (a total of 119.7 Mb). Deleterious and neutral nonsynonymous SNPs outside iHS and Rsb regions were assigned into bins of non-overlapping 200-kb windows by their distance from the nearest hitchhiking region. The ratio of deleterious to neutral SNPs was modeled as a function of the distance (dk) of each window in bin k to the nearest hitchhiking region using logistic regression:
Population specific patterns of hitchhiking were examined using regions identified by multiple tests of selection and by iHS alone. Regions identified by multiple tests of selection were not differentiated by which population showed evidence of selection and so represent a composite view of hitchhiking [33]. iHS regions were identified in CEU, CHB+JPT, and YRI, using empirical cutoffs of 0.25%, 1%, and 5%. To identify iHS hitchhiking regions using HapMap Phase II data, iHS scores of individual SNPs (HapMap Phase II) were downloaded (http://hg-wen.uchicago.edu/selection/), and for each 100-kb non-overlapping genomic window the signal of selection was evaluated by the fraction of SNPs with iHS scores above +2 or below −2, as in Voight et al. [35]. Windows were grouped into bins by the number of SNPs within the window using increments of 25 SNPs. Empirical cutoffs were applied separately to each bin. Windows with less than 10 SNPs and bins with less than 100 windows (less than 400 for the 0.25% cutoff) were excluded.
Disease-associated alleles were obtained from OMIM (http://www.ncbi.nlm.nih.gov/omim), a catalog of published GWAS studies (http://www.genome.gov/26525384) and Google Scholar searches of the literature. For OMIM, dbSNP IDs (release #132) with OMIM links were downloaded (ftp://ftp.ncbi.nih.gov/snp/database/organism_data/human_9606/OmimVarLocusIdSNP.bcp.gz). Excluding InDels, unmapped variants, and variants on sex chromosomes, 10,775 OMIM variants were re-mapped to the reference genome using UCSC's LiftOver program. All OMIM variants included in HapMap Phase II (release #24) were considered common enough to be SNPs with the exception of those with minor allele frequency of zero. Average minor allele frequency across CEU, CHB+JPT, and YRI was compared between hitchhiking and non-hitchhiking regions. For allele frequency, HapMap Phase II+III (release #26) data were used [75]. For disease SNPs identified in the 1000 Human Genomes project [28], common and rare variants were distinguished by their mean allele frequencies across CEU, CHB+JPT, and YRI using a 5% allele frequency cutoff. SNPs without allele frequencies were set to an allele frequency of zero.
Disease-risk alleles were obtained from a catalog of published Genome-Wide Association Studies (GWAS) [50]. Excluding 115 regions without associated SNPs and 10 regions with multi-SNP haplotype associations, we obtained 3,383 non-redundant autosomal risk alleles with the strongest trait association at each locus from a total of 585 published studies. Allele frequencies in control population and odds ratios were available for 2,504 and 1,253 risk alleles, respectively. Reported risk allele frequencies were averaged over control populations if the risk allele was identified in more than two studies. However, reported odds ratios were not pooled over different studies and traits even if the risk allele was reported in multiple studies.
To examine common deleterious and neutral SNPs reported in the literature, we used Google Scholar (http://scholar.google.com) and the dbSNP rs number as the search term. The set of tested SNPs was based on 790 deleterious SNPs and 369 neutral nonsynonymous SNPs with an allele frequency of greater than 30% in the HapMap CEU panel. SNPs within known olfactory receptors were excluded. Neutral SNPs were matched to the frequency distribution of deleterious SNPs by acceptance-rejection sampling. As a result, derived allele frequencies are not significantly different between the two sets (Wilcoxon Rank Sum Test, P = 0.79). For each SNP, we searched for reported phenotype associations based on population association or cell-based functional assays. To minimize potential human biases, dbSNP identifiers of deleterious and neutral SNPs were mixed together and Google Scholar search results were manually examined without knowledge of SNP classification. Patents, eQTL associations, conference and poster abstracts, and journals without full-text access were excluded. SNP associations had to be significant after a multiple testing correction. SNP association studies with sample size less than 200 were also not included.
To identify genomic regions with exceptionally high rates of deleterious SNPs per coding sequence, 1-Mb sliding windows were scanned across all autosomes with a step size of 0.5 Mb. Assuming that the rate of deleterious SNPs per accessible coding sequence is constant across the genome, a Poisson distribution was used to evaluate the excess number of deleterious SNPs in each window. The expected number of deleterious SNPs per window was set to the product of the genome average (0.51 deleterious SNPs per 1 kb accessible CDS) and the length of accessible coding sequence in the window. Out of 3,549 windows with at least two deleterious SNPs, 70 (2%) with the highest P-value were selected (P<4.5×10−4). After excluding regions that were consecutive to or overlapped another region with a smaller P-value, we retained 43 regions.
|
10.1371/journal.pcbi.1003670 | A Digital Framework to Build, Visualize and Analyze a Gene Expression Atlas with Cellular Resolution in Zebrafish Early Embryogenesis | A gene expression atlas is an essential resource to quantify and understand the multiscale processes of embryogenesis in time and space. The automated reconstruction of a prototypic 4D atlas for vertebrate early embryos, using multicolor fluorescence in situ hybridization with nuclear counterstain, requires dedicated computational strategies. To this goal, we designed an original methodological framework implemented in a software tool called Match-IT. With only minimal human supervision, our system is able to gather gene expression patterns observed in different analyzed embryos with phenotypic variability and map them onto a series of common 3D templates over time, creating a 4D atlas. This framework was used to construct an atlas composed of 6 gene expression templates from a cohort of zebrafish early embryos spanning 6 developmental stages from 4 to 6.3 hpf (hours post fertilization). They included 53 specimens, 181,415 detected cell nuclei and the segmentation of 98 gene expression patterns observed in 3D for 9 different genes. In addition, an interactive visualization software, Atlas-IT, was developed to inspect, supervise and analyze the atlas. Match-IT and Atlas-IT, including user manuals, representative datasets and video tutorials, are publicly and freely available online. We also propose computational methods and tools for the quantitative assessment of the gene expression templates at the cellular scale, with the identification, visualization and analysis of coexpression patterns, synexpression groups and their dynamics through developmental stages.
| We propose a workflow to map the expression domains of multiple genes onto a series of 3D templates, or “atlas”, during early embryogenesis. It was applied to the zebrafish at different stages between 4 and 6.3 hpf, generating 6 templates. Our system overcomes the lack of significant morphological landmarks in early development by relying on the expression of a reference gene (goosecoid, gsc) and nuclear staining to guide the registration of the analyzed genes. The proposed method also successfully maps gene domains from partially imaged embryos, thus allowing greater microscope magnification and cellular resolution. By using the workflow to construct a spatiotemporal database of zebrafish, we opened the way to a systematic analysis of vertebrate embryogenesis. The atlas database, together with the mapping software (Match-IT), a custom-made visualization platform (Atlas-IT), and step-by-step user guides are available from the Supplementary Material. We expect that this will encourage other laboratories to generate, map, visualize and analyze new gene expression datasets.
| Deciphering and integrating the genetic and cellular dynamics underlying morphogenesis and homeostasis in living systems is a major challenge of the post-genomic era. Although full genome sequencing is available for a number of animal model organisms [1], quantitative data for the spatial and temporal expression of genes is still lacking [2].
Remarkable advances in photonic microscopy imaging [3],[4],[5] and labeling techniques [6] allowed gathering data at all levels of a multicellular system's organization with adequate spatial and temporal resolutions. Fluorescent in situ hybridization techniques [7], immunocytochemistry and transgenesis, combined with 3D optical sectioning, make it now possible to assess the dynamics of gene expression throughout animal development with precision at the single-cell level. However, moving forward from databases of gene expression that contain average values at low spatiotemporal resolutions—such as those obtained from DNA microarrays available for most model organisms—to a dynamic, cell-based 4D atlas is a major paradigm shift that requires the development of appropriate methods and tools.
In this context, the design and implementation of automated image analysis strategies to build a gene expression atlas with resolution at the cellular scale is an important methodological bottleneck towards greater biological insights [8],[9]. The task of assembling imaging data from cohorts of individuals, or analyzed embryos, onto a series of 3D prototypes, or templates (one per developmental stage), can be approached by finding a spatial correspondence between individuals based on registration methods, a technique used in medical imaging [10]. Yet, gathering and consolidating into a single prototype multimodal and multiscale features from different specimens that exhibit phenotypic variability remains a difficult challenge.
Recent studies on different model organisms have explored computational strategies for building atlases either by measuring cell positions to create prototypic specimens [11],[12] or by gathering gene expression patterns observed in cohorts of specimens [13],[14],[15],[16]. Yet, very few frameworks have combined both features. Long et al. [11] collected data from 15 C. elegans specimens at the earliest larval stage (L1 with 357 cells) to build a statistical 3D atlas of nuclear center positions. C. elegans presents a number of advantages facilitating the reconstruction process. The entire organism can be imaged with resolution at the single-cell level and its cell lineage tree is stereotyped enough to allow spatiotemporal matching of different individuals at this level. The same features allowed the reconstruction of a prototypic lineage for a cohort containing six specimens of Danio rerio (zebrafish) embryos throughout their first 10 cell division cycles [12]. Peng et al. [15] achieved the spatial matching of 2,945 adult Drosophila brains to collect the expression patterns of 470 different genes. Similarly, Lein et al. [13] constructed a comprehensive atlas of the adult mouse brain containing about 20,000 gene patterns. The first gene expression atlas with resolution at the cellular scale was produced by Fowlkes et al. [14]. They integrated 95 gene expression patterns observed at 6 different developmental stages in a total of 1,822 different Drosophila embryos within a common 3D stencil.
Applying this approach to vertebrate model organisms is more difficult because of higher cell lineage variability and heterogeneous levels of gene expression within highly dynamic patterns. In addition, the reconstruction of 3D gene expression templates at cellular scale for vertebrate species is likely to require the acquisition of partial volumes recorded at high resolution [15] from single specimens, and their precise mapping onto in toto reference specimens. The zebrafish, a vertebrate model organism increasingly used for its relevance to biomedical applications [17], cumulates good properties for investigating the reconstruction of the multiscale dynamics of early embryogenesis. The gene regulatory network (GRN) architecture of the zebrafish early embryonic development is under construction [18] and the embryo is easily accessible and amenable to transgenesis, multiple in situ staining and 3D+time imaging. The spatiotemporal data offered by a 4D atlas of gene expression with resolution at the cellular level is expected to provide the necessary measurements for further modeling of the GRN dynamics and possible integration of the genetic and cellular levels of organization [19]. Such data would make the zebrafish the first vertebrate model amenable to a systemic study. However, building 3D templates of gene expression for the zebrafish blastula and gastrula stages is especially problematic due to the lack of morphological landmarks required for the registration of patterns [20],[21].
We provide a methodology to construct, visualize and analyze a gene expression atlas composed of templates at various stages of vertebrate early development. We designed, implemented and now deliver two computational frameworks, Match-IT and Atlas-IT, to support the automatic mapping of 3D gene expression patterns from different individuals (the analyzed embryos) onto common reference specimens (the templates) with resolution at the cellular scale. This “virtual multiplexing” procedure [14] overcomes the limited number of gene products that can be jointly stained and measured in a single specimen.
Match-IT was used to produce the prototypic cartography of 9 gene expression patterns imaged from 3D double fluorescent in situ hybridization at 6 developmental stages (Table S1, Movie S1, Figs. S1, S2, S3, S4, S5, S6, S7). Atlas-IT was designed to interactively visualize gene coexpression patterns and their dynamics. We validated our 4D atlas construction methodology by an automated quantitative assessment of gene patterns' similarity and overlap through time. Analytical tools, such as clustering, were designed to identify morphogenetic domains and gene synexpression groups, i.e. groups of genes sharing the same spatiotemporal expression patterns. The proposed spatiotemporal atlas of zebrafish blastula and early gastrula preserves the information of the cell as the gene expressing unit, providing means for the integration of genetic and cellular data unavailable so far.
We designed a computational framework (Fig. 1), going from image acquisition to image data analysis, to perform the mapping of different stained gene expression patterns onto a common prototypic model at each developmental stage (Fig. S8), thus creating a series of 3D templates of gene expression with resolution at the cellular scale.
The processing workflow consisted of embryo staining, image data acquisition (Materials and Methods), nuclear center detection, gene pattern segmentation, mapping of the analyzed embryos onto a template at each stage, and selection of template cells positive for the expression of specific genes. This methodology was designed to document at a sufficient spatial and temporal resolution the gene expression dynamics underlying the formation of the Spemann organizer and the embryonic axis of zebrafish early embryos. To this end, we imaged the dorsal side of fluorescently stained embryos with cellular resolution from fixed specimens about every 30 min from 4 to 6.3 hpf. The resulting 6 templates comprised a stencil of in toto 3D images of the template specimens (Fig. 2a) at different stages, and mappings of the partial 3D views of the analyzed embryos (Fig. 2b).
In order to integrate 3D data into one template, our novel Match-IT tool (Software S1 and User Guide S1) performed the segmentation of gene expression domains, the mapping of analyzed embryos onto a common reference specimen and the identification of positive cells (Fig. 1 and Movie S2), eventually delivering a 3D database that summarized the genetic profile of single cells.
Analysis of the 3D templates produced by Match-IT required dedicated visualization tools to test hypotheses and derive biological insights. The available software kits did not fulfill our requirements, either because they were too specific for a given model organism (such as PointCloudXplore [26] for Drosophila) or because they were too generic as visualization and processing tools (such as Icy [27], Vaa3D [28], or CellProfiler [29]) and did not allow displaying selections of individual cellular positions or querying a template for coexpression domains with resolution at the cellular scale.
For these reasons, we designed, developed and deliver here the Atlas-IT interactive visualization interface (Fig. 4a and Software S2 and User Guide S2) to explore 4D atlas resources. With this tool, we can interact with the complete atlas data, in particular superimpose raw images (either as 3D volumes or orthoslices), segmented patterns, and the whole set of detected template nuclei or selected positive nuclei at any time point (Movie S3). Atlas-IT can be used to assess the dynamics of gene coexpression domains or the variability of gene expression patterns.
We used Match-IT and Atlas-IT together to reconstruct a 4D atlas of zebrafish early embryogenesis, which is now released. It comprises 6 developmental stages and 9 gene expression patterns chosen to study a specific embryological question, namely the genetic dynamics underlying the formation of the Spemann organizer at the dorsal midline [1] (a region in the zebrafish containing precursors of the segregation between the prechordal plate and the notochord [30]). The 9 genes are: gsc, sox32, tbx16, oep, snai1a, foxa2, ntla, flh, and egfp, where the latter was was detected in a custom-made transgenic line Tg(−4gsc:egfp)isc3. These genes appear as nodes in the axial mesendoderm GRN proposed by Chan et al. [18]. In addition, egfp allowed us to validate the transgenic line as a faithful reporter of early gsc gene expression (Fig. S15). The time series of 3D templates was chosen to explore gene expression dynamics from the onset of zygotic activation at 3 hpf until early gastrulation, and encompasses the following developmental stages: sphere (4 hpf), dome (4.3 hpf), 30% epiboly (4.7 hpf), 50% epiboly (5.3 hpf), shield (6 hpf) and late shield (6.3 hpf) according to the staging defined at . For each new gene expression to be mapped, a cohort of individuals was processed for double in situ hybridization and 3 of them were imaged. The atlas construction methodology was established by using one specimen of each cohort (Table S1).
The atlas was constructed to be able to compare gene expression patterns from different stained specimens. Establishing spatial relationships between gene patterns required assessing gene expression variability and calculating mean expression domains (Materials and Methods). The expectation was that the spatial relationships observed between two genes stained in the same embryo should be maintained between their mean expression domains in a template. The expression of gsc was revealed in 9 different specimens, which comprised 8 analyzed embryos and one template, at each developmental stage. It provided a paradigmatic case to calculate a mean expression domain and assess gene variability (Fig. S16). At any given stage, we quantified the mean distance from the complete outer surface of each individual gsc domain () to the closest boundary point of the mean domain (), following a leave-one-out protocol (Fig. 4b). The measured distance, which reflected both the accuracy of our mapping scheme and the inter-individual variability between the boundaries of the gsc expression domains, was on average less than , i.e. approximately one cell diameter (Fig. S17). This accuracy error remained within the same range independently from the thickness of the three main embryo planes (Fig. S18). Additionally, more than 80% of all the individual border points were less than one cell row away from the border, indicating that there were no large distance discrepancies along the contours (Fig. S19).
To demonstrate that this level of accuracy was maintained in regions far from the gsc expression domains, we replicated the same quality measure with another gene, tbx16, which spread across a much larger area than gsc. With Match-IT, we added two new tbx16 datasets, tbx16b and tbx16c, to the already existing tbx16a expression in the atlas at 6.3 hpf (Fig. S20a). The mean distance from the complete outer surface of each individual tbx16j domain to the closest boundary point of tbx16 remained under one cell diameter (Fig. S20b). Moreover, the histogram of distances between border points of tbx16j and tbx16 confirmed that most of the expression contours lay within two cell rows from each other (Fig. S20c). Note that this quality measure was an upper bound of the registration quality reflecting both the mapping variability and the intrinsic inter-embryo variability.
Additionally, we confirmed that the spatial relationships between every gene and the patterns in the analyzed embryos were the same in each template with respect to the domain. In particular, this was the case for the oep-gsc pair illustrated in Fig. 4c,d.
Various analysis tools for the quantitative analysis of a spatiotemporal atlas of gene expression were also developed (see Materials and Methods). We performed an automated identification of gene coexpression pattern dynamics in space and time, explored clustering strategies at the cellular level to automatically identify morphogenetic domains or spatiotemporal gene synexpression groups, and introduced an “entropy” analysis for gene expression.
We have designed, developed and delivered the Match-IT and Atlas-IT software tools dedicated to the reconstruction, analysis and visualization of a 4D atlas of gene expression in zebrafish early embryogenesis. The atlas comprises 6 different time points between 4 and 6.3 hpf, gathering data for 9 gene patterns into 6 different 3D templates.
So far, the only known method delivered for the reconstruction of gene expression atlases in the zebrafish was designed by Ronneberger et al. [21] for the brain and at late developmental stages, when a large number of morphological landmarks could already be recognized. Given the complexity of building a zebrafish brain atlas at late stages, the authors imposed strong constraints on the data in terms of staining protocols and imaging. Our own atlasing strategy was designed to map partial 3D volumes onto whole embryos chosen as templates. Specimens were only required to display, in addition to any pattern of interest, nuclear staining for single-cell counterstain and a common gene expression pattern, gsc in the present version of the atlas, used for the registration step. This gene was chosen as a relevant marker, with early, strong and well-regionalized expression, to serve as a reference for constructing the dorsal side's gastrulation atlas. Thus we have minimal prerequisites for data format and specimen preparation, which should facilitate the introduction of new data into the atlas. In addition, our scheme could be easily adapted to other vertebrate organisms, e.g. xenopus, dogfish or lamprey, at early stages of development, when too few morphological features are available to use landmark-based registration methods in the mapping process. The possibility of visual inspection and, if necessary, manual correction using our Match-IT graphical interface contributes to flexibility and accuracy when integrating new data into the atlas and validating the results.
Our choice to work with a hybrid automated/supervised method of nuclear center detection proved to be suitable for quantifying certain features of gene expression pattern dynamics at the cellular level. This opens the possibility to discuss, in terms of cell number, the overlap between gene expression patterns and their evolution in time. It also allows studying whether cell proliferation alone is enough to account for the expansion of gene expression patterns, by correlating internuclear distance and cell division, which, in zebrafish early development, happens at constant global cell volume (Fig. S26). On the other hand, the resolution of the atlas at the cellular scale is a requirement to exploit the correlation between gene expression dynamics and cell lineage. Cellular resolution enables further mapping of the atlas onto digital specimens reconstructed from live in toto imaging, starting with our transgenic line.
Working at the cellular resolution was also intended to tackle the problem of gene expression quantification. Current strategies for in situ hybridization could at best provide relative measurements suitable for quantifying graded patterns and fuzzy borders within each analyzed embryo. Such a relative quantification would be readily available from our atlas (Fig. S10). We expect future developments of the programmable in situ amplification technique [7] to help achieve quantification of gene expression comparable among different analyzed embryos at the cellular level.
The relevance of the atlas relies on its ability to represent and integrate the same information as would be obtained by inspecting different patterns in the same specimen. This depends on the accuracy of the registration strategies but most importantly on how the atlas construction scheme deals with individual variability. Every step of the mapping strategy has to cope with individual variations in terms of shape, cell number, cell density, and variability of the reference gene pattern. In this context, the choice of the template is crucial. The template should be closest to the mean of the population, based on geometric parameters and gene expression. Ideally, a multiscale model of individual variability should drive the choice of the atlas template as well as representative reference patterns or features to guide the mapping. In our case, the gsc pattern served as a guide for the registration step, based on the hypothesis that its expression is symmetric with respect to the bilateral plane. Although this is a reasonable assumption, it is an approximation that might be confronted to other features such as other reference gene patterns or additional morphological traits. The templates used in this paper were visually chosen to be the closest to the mean. Although this choice may not be fully representative of the average morphology, the concept of average is also not completely relevant for the released proof-of-principle atlas that comprises 9 specimens per developmental stage. The tools released here open the way for a broader population that could ideally produce a more representative template.
In this context, we calculated a mean gsc expression pattern after registering the domains from 9 different specimens. The resulting domain could be subsequently used as a new reference to refine the global mappings. Moreover, all the genes gathered in the atlas could be averaged, thus preventing potentially misleading conclusions based on single specimens that might be outliers. The increase in size of the cohorts will allow exploring the possible convergence of the averaging strategy toward a single or multiple prototypical specimens.
Atlas resources will only be fully exploited with the development and use of automated analysis methods and dedicated visualization tools. Toward this objective, we designed Atlas-IT to provide a number of functionalities not available in any of the visualization tools that we examined: augment/visualize/analyze raw data and segmented data, calculate mean gene expression domains, gene coexpression patterns, synexpression groups, and morphogenetic domains by cell clustering. Interactive visualization and data display are essential to reveal biologically relevant information. The exploration of analytical methods to highlight spatial and temporal correlations is also a major endeavor. Typically, clustering methods have been used to establish the gene expression profiles of cells and tissues from microarray data, and more recently to group anatomical regions according to their gene expression profile [13],[32],[33]. Although clustering of spatial gene expression patterns has been described elsewhere [34], it is the first time that this method is applied to gene expression profiles at the cellular level, , providing the means to reveal morphogenetic domains and synexpression groups. Additionally, whereas the Shannon entropy has been used to measure gene expression complexity [35], it is also the first time that this measure is applied to spatially mapped data. Introducing the concept of “genetic entropy” in the analysis of atlas data offers a new systematic way to assess cell diversification and its underlying genetic complexity. This analysis proved to be robust against the noise due to errors in the segmentation and/or spatial mapping. Although a relatively high proportion (100 out of 512) of all possible gene expression profiles were found in the atlas, only 30 of them (i.e. ) produced 75% of all the atlas genetic information (Fig. S24).
Making a gene expression atlas is a necessary step toward the integration of multiscale and multimodal data, which should be organized, displayed and annotated to provide and share as much relevant information as possible. Developmental biology remains far behind the biomedical field in the construction and sharing of this type of resources. Thus, before reaching a consensus and establishing standards in the field, a lot remains to be explored in terms of different schemes, their flexibility, their potential and limitations. The atlas construction process presented here allowed us to address some of the most difficult biological questions linked to individual variability, its components and characteristic scales. A gene expression atlas often comprises hundreds or even thousands of genes [36]. On the other hand, resources can grow and diffuse only if deployed together with appropriate algorithms and analytical tools. Our novel construction and manipulation methods, which led to the first release of the zebrafish blastula and early gastrula atlas, are meant as a contribution toward the complete reconstruction of the zebrafish embryonic physiome (or “embryome”) under different genetic and environmental conditions.
In vitro fertilization was used to synchronize the spawn from wild type (wt) or transgenic crosses from the custom made fish line Tg(−4gsc:egfp)isc3. Embryos, staged according to Kimmel et al. [37], were fixed 24 h at in PFA 4% then rinsed 3 times in PBS 0.1% Tween and stored at in ethanol. Double fluorescent in situ hybridization (FISH) was carried out as described in Brend et al. [38] using antisense RNA probes labeled with fluorescein or digoxygenin. Probes were detected with an anti-digoxigenin-POD Fab fragment and anti-fluorescein-POD Fab fragment (Roche) used at 1∶250 in a blocking reagent solution (Roche). Probe detection was done with Cy3 or Cy5 mono NHS ester (Amersham) or NHSFluoresceine (Pierce) tyramides as POD substrates. Nuclei were stained in DAPI (Invitrogen D3571).
As an input, our methodology used 3D images acquired by confocal laser scanning microscopy from fixed zebrafish embryos with fluorescent staining of gene expression patterns and DAPI counterstain to highlight cell nuclei. Image acquisition was performed with a Leica SP2 two-photon (for DAPI) and confocal laser scanning upright microscope with a Leica objective HCX APO 20X/0,5W U-V-I or HCX APO 10X/0,3. Embryos were mounted in a teflon mold at the bottom of a 3 cm Petri dish filled with 1×PBS, 01% twin 20, and maintained properly oriented with 1% agarose.
The nuclei and gsc expression domains were systematically revealed in all the analyzed embryos and templates, and used to compute the gene expression mappings. In addition to the reference gene, gsc, each analyzed embryo was stained for the expression of another gene of interest. The template data was obtained by imaging the whole embryo with a 10× objective while the analyzed specimens were imaged with a 20× objective providing a 3D view limited to the dorsal side of the embryo with a better spatial resolution (Fig. 2a,b and Movie S1).
The fluorescent in situ hybridization used a state-of-the-art protocol [38] and reproduced standard data (zfin.orgzfin.org). More details about data acquisition parameters and specimen features can be found in Table S1 and Fig. S1, S2, S3, S4, S5, S6, S7.
The Match-IT custom-made code was implemented in ITK and Matlab, including the MathWorks package “geom3D” redistributed under a BSD license. A public release of this software, together with sample datasets and a user guide, accompanies the publication of this article, http://bioemergences.iscpif.fr/documents/MatchIT.zip. The segmentation of the gene expression patterns in each analyzed embryo was carried out by a thresholding operation supervised by a biologist to best define the domain features. This operation was followed by “morphological closing” [39], a mathematical transformation based on a spherical structuring element the size of a typical cell diameter (i.e. internuclear distance). Finally, a converse “morphological opening” operation left only the largest connected pattern. The common referential extraction started by applying a spherical fit to the outer cell nuclei in all analyzed embryos and templates. The blastoderm margin was identified with a plane, , fitted to the 5% southernmost nuclei. The bilateral symmetry plane, , was found by connecting the spherical model center and the center of mass of the gsc segmented domain perpendicular to the blastoderm margin. The origin of the triplet was placed at the latitude of the blastoderm margin, and the longitude was defined by the center of mass of the gsc domain. The registration ([10],[40]) of the analyzed embryo images on the template employed the ITK registration toolkit to optimize the cross-correlation metric between the embryo shape of the template and that of the analyzed embryos according to a step gradient optimizer. The embryo shapes were weighted by the inverse distance function to the external blastoderm contour (i.e. half the average internuclear distance away from the outermost nuclear layer).
The Atlas-IT custom-made visualization platform was implemented in Processing. A public release of this software, together with sample datasets and a user guide, accompanies the publication of this article, http://bioemergences.iscpif.fr/documents/AtlasIT.zip
At each developmental time point, a total of 9 different analyzed embryos with staining were mapped onto the template where expression was also revealed. Consequently, every nucleus, , in the template was assigned a value, , ranging from 0 to 9, depending on the number of analyzed patterns that led to its selection as positive for the expression of . We used a Voronoi diagram to model the cell around each nucleus and assigned these cells their corresponding value . In order to measure the variability of the resulting mean expression, we studied the profile of across 3 cutting lines centered on the mean centroid and following the specimen anatomy along the lateral, radial and sagittal directions respectively (Fig. S16).
We demonstrated that the proposed clustering and entropy schemes are robust against changes in the thresholds employed to segment the gene expression patterns in the atlas. In particular, we chose two gene expressions in the atlas at 6.3 hpf: , which co-expresses with , and , which spreads through a much larger area than . For the expression of these two genes, we modified by the thresholds chosen by the biologist expert, computed the new segmented patterns and modified accordingly the number of positive cells found in the atlas. The entropy and clustering resulting from these modified atlases were compared to the original atlas and showed to be robust against these threshold changes (Fig. S25). To compare the modified vs. the original clustering (Fig. 5b–c) we used two metrics previously employed in literature: a) the correlation between the distance matrix that generate the modified and the original clustering hierarchical trees [42], b) the cophenetic correlation, a measure of how faithfully a hierarchical tree preserves the pairwise distances between the original data points [43], [44]. In this later case, the cophenetic coefficient was extracted by comparing the original hierarchical tree to the new pairwise distances generated by the modified atlases. To compare the modified vs. the original entropy we computed the difference in number of bits. The biggest difference between all the modified and the original atlas was 0.15 bits in entropy and a 0.03 decrease for both the cophenetic coefficient and the correlation between distance matrix (Fig. S25). To put these values in perspective, the minimal possible variation to the atlas (changing the value of one gene expression for one cell only) had an impact of 0.0003 bits of entropy and 0.005 in cophenetic correlation, whereas a substantial variation to the atlas (e.g. changing one third of the atlas values or substituting it by a random atlas) had an impact of 0.83 and 3.6 bits of entropy and 0.52 and 0.79 in cophenetic correlation respectively.
|
10.1371/journal.pcbi.1003636 | High-Resolution Modeling of Transmembrane Helical Protein Structures from Distant Homologues | Eukaryotic transmembrane helical (TMH) proteins perform a wide diversity of critical cellular functions, but remain structurally largely uncharacterized and their high-resolution structure prediction is currently hindered by the lack of close structural homologues. To address this problem, we present a novel and generic method for accurately modeling large TMH protein structures from distant homologues exhibiting distinct loop and TMH conformations. Models of the adenosine A2AR and chemokine CXCR4 receptors were first ranked in GPCR-DOCK blind prediction contests in the receptor structure accuracy category. In a benchmark of 50 TMH protein homolog pairs of diverse topology (from 5 to 12 TMHs), size (from 183 to 420 residues) and sequence identity (from 15% to 70%), the method improves most starting templates, and achieves near-atomic accuracy prediction of membrane-embedded regions. Unlike starting templates, the models are of suitable quality for computer-based protein engineering: redesigned models and redesigned X-ray structures exhibit very similar native interactions. The method should prove useful for the atom-level modeling and design of a large fraction of structurally uncharacterized TMH proteins from a wide range of structural homologues.
| Membrane proteins perform crucial cellular functions and can be involved in serious diseases but remain difficult to study experimentally. Hence, high-resolution membrane protein structures are scarce which hinders the design of selective therapeutics and of receptors with novel function for systems/synthetic biology applications. The computational modeling of membrane protein structures represents an important alternative approach but, to achieve high accuracy, usually requires structural information from closely related proteins currently unavailable for most membrane proteins. To address this limitation, we have developed a novel method to predict membrane protein structures from the structures of non-closely related proteins that differ both in loop and transmembrane regions. Using this approach, we show that a large diversity of membrane proteins can be reconstructed at a level of accuracy suitable for computer-based protein engineering applications. Because requiring information from a single distant homolog only, we expect that around 60% of human membrane proteins can reliably be modeled using our approach, thereby allowing precise structure/function studies on a large fraction of structurally uncharacterized membrane proteins.
| Membrane proteins perform a wide diversity of critical functions in living cells but are also involved in serious diseases and represent more than 60% of current drug targets [1], [2]. Despite recent tremendous progress in membrane protein expression, biochemistry and X-ray crystallography, eukaryotic membrane protein structures remain difficult to characterize experimentally [3]. The lack of high-resolution structures hinders the design of more effective therapeutics and of receptors with novel function for systems/synthetic biology applications which rely on atomic-resolution information [4]. The high-resolution prediction of membrane protein structures is therefore an important alternative approach but remains a major challenge in absence of close structural homologues [5]. Although numerous methods have been developed to model G protein-coupled receptor (GPCR) structures [6]–[9], much fewer techniques have been developed and applied to the entire class of alpha helical membrane proteins. Current state-of-the-art de novo structure prediction techniques of alpha helical membrane proteins can generate low-resolution models with native-like topologies [10]–[12] and, despite some insightful applications [13], most current comparative modeling methods do not significantly improve starting templates [14], [15]. The main structural differences between distant homolog transmembrane alpha-helical (TMH) proteins are found in loop regions and in helical conformations shaping TMH core structures and ligand/effector binding sites. While the problem of rebuilding protein loops has been extensively studied [16], [17], the accurate modeling of membrane protein structures from distant homologues diverging in both loop and TMH core regions is a remaining unsolved challenge [5]. The origins of TMH conformational diversity are multiple and range from the presence of localized sequence-specific distortions (e.g. Proline-induced kinks) to local bends and global tilts stabilized by specific tertiary contacts [18]–[23]. Many of these features cannot be accurately predicted from sequence information alone and requires the explicit modeling of atom-level physical interactions stabilizing these structures [18]–[20], [24]. The large size of TM proteins and associated number of degrees of freedom combined with the ruggedness of the all-atom energy landscape make their prediction at atomic resolution computationally intractable using an exhaustive conformational search in torsional angle space.
To address this problem, we have developed a general modeling strategy based on efficient sampling techniques of alternative TMH structures to reconstruct both TMH core and loop regions from distant structural homologues. The method was stringently validated in two blind predictions where the generated models were top-ranked [14], [15] and in a large benchmark dominated by pairs of membrane protein distant homologues where starting templates were almost all significantly improved. Computational design calculations suggest that the models should be of suitable accuracy for rational protein engineering applications.
As shown in Fig. 1, multiple sequence alignments using Hidden Markov Model (HMM)-based techniques [25] are first performed to identify structural homologues that best align with the target sequence. The quality of the alignment in the TMH regions leads to two different model rebuilding strategies: 1) If the alignment in the TMH regions does not exhibit significant gaps, if the positions of coils or residues promoting local distortions are identical and if TMHs are predicted to have similar length, then target and template TMH structures are likely very similar (Methods). In this situation, the template TMH structure is first kept fixed onto which loops diverging between target and template are reconstructed de novo using fragment insertion techniques [10], [16]. The reconstructed models are then refined at the all-atom level [24] (Fig. 1). 2) If one of the above-mentioned conditions is not satisfied however, target and template TMH structures may differ significantly and the target TMH region is also reconstructed as described below (Fig. 1). TMH structures mostly sit in the hydrophobic environment of the lipid membrane disfavoring any unsatisfied polar atom. Therefore, we reasoned that, except in local bends or kinked regions where hydrogen-bond networks may be partially disrupted, most TMH regions to be rebuilt adopt helical conformations. Previous work also suggests that most bent helices can be approximated by straight TMH fragments away from the local distortion [24] which can adopt diverse structures (from a 310 turn to a π helix) [18], [19], [23]. To efficiently identify alternative low-energy TMH conformations, each TMH fragment away from local predicted bends (that usually span from 4 to 6 residues) is first modeled as a rigid-body helix and its conformation optimized in a low-resolution search sampling helical rigid-body degrees of freedom (see Method). This low-resolution search averages out side-chain conformations, effectively flattening the conformational free energy landscape and allows the rapid identification of low-energy TMH conformations [10] with alternative interhelical and/or kink angles. Loop and local TMH regions around bends or kinks are then rebuilt using fragment insertion techniques and the fully-reconstructed low-resolution structures are refined at all-atom to identify the lowest-energy native-like structures. At this stage, global deformations of TMH stabilized by short-range atom-level tertiary interactions can be identified and selected by energy [24]. To avoid sampling regions of the conformational space unlikely to be occupied by the peptide chain, distance constraints are applied to the template structure at pairs of residues in proximity and conserved in both target and template sequences (Methods).
To assess the significance of our technique developed to model membrane proteins from distant homologs, we analyzed the space of structural homologs available to all human TMH proteins using HHpred [25]–[27], a toolkit for searching and aligning query sequences with sequences from existing structures. The resultant HHpred alignments were filtered by a range of percent sequence identity thresholds (i.e. of homolog hit versus target) and percent coverages (i.e. of total length of target sequence) of 90%, 75%, 60% or 50%. As shown in Fig. S1A, the percentage of human multi-pass TMH proteins sharing 15–25%, 25–35% and >35% sequence identity with their best structural homolog hit is 43%, 12% and 12%, respectively. Similar distributions were obtained for datasets including also human single-pass TMH proteins or consisting of human multi-pass TMH proteins truncated to their TM domains (Fig. S1C,D). These results indicate that only distant structural homologs are currently available for a large fraction of human TMH proteins. Moreover, as shown in Fig. S1B, only one single distant structural homolog is found for a large fraction of these TMH proteins. These results justify our approach and led us to test our technique on a benchmark where membrane protein structures were primarily modeled from single distant structural homologs.
The technique was tested in two challenging blind predictions of membrane receptor structures, i.e. GPCR-DOCK 2008 for the adenosine receptor (A2AR) [15] and GPCR-DOCK 2010 for the chemokine receptor (CXCR4) [14]. The closest homolog template to A2AR was the beta 1 adrenergic (B1AR) receptor structure [28] sharing 32% sequence identity and exhibiting excellent sequence alignment in the TM region with A2AR. Therefore, TMH remodeling of the template structure was not required. A total of 206 models were submitted by the participants but very few showed significant improvements compared to the initial template structure. Among the top 10 models for both receptor and ligand binding prediction accuracy, one of our submitted models ranked co-first and first for the receptor prediction accuracy over the full length (i.e. 283 residues) and TMH region (i.e. 214 residues), respectively (the reported model from Costanzi had a lower “full-length” RMSD but did not include the entire long ECL2 loop, see Table 1 in [15]). An additional model submitted without ligand (submission 3600_8, Supplementary information in [15]) was even closer to the target (Cα RMSD of 2.9 Å over 283 residues) and ranked first among all submitted models for both TMH and full-length structures with a Z-score of 1.51. For CXCR4, although HHpred identified the beta 2 adrenergic receptor (B2AR) as the best aligned structural homolog, B2AR is a distant homolog sharing only 22% sequence identity with CXCR4 [29] and its second TMH did not align well with the target near a proline-inducing kink. The C-terminal part of TMH2 starting from the kink and the loop structures were therefore remodeled, and 5 low-energy models with docked ligand were submitted. One model was ranked first for the accuracy of the receptor structure among all the 158 models submitted by the participants for the two CXCR4 structures (Z-score = 1.72) and 2 additional ones were ranked second and third for the prediction accuracy of the CXCR4/IT1t structure (Z-scores of 1.36 and 1.24) [14]. Both blind predictions demonstrate that our technique significantly improved starting templates and generated models exhibiting several structural features closer to the target X-ray structures than to the starting template. For example, the TMH shifts in the A2AR structure from B1AR (Fig. 2A), the local kink in TMH2 of CXCR4 (Fig. 3A–C) and the 27 residues long extracellular loop 3 of CXCR4 (residues G220-I246) (Fig. 4A) were predicted quite accurately. Although the conformation of the 16 residues long partially disordered extracellular loop 2 (residues A174-E179, R183- N192) was not predicted with near-atomic accuracy, its conformation was closer to the target CXCR4 than to the starting template (Fig. 4F) and was the most accurate prediction for that region among all submitted models [14]. We also attempted the modeling of D3DR but, since close homologs (sequence identity >30%) were available, the main interest for this target was not receptor modeling but ligand docking which is outside the scope of the present study. With a Z-score of 0.41, our best model of D3DR ranked within the top 35% of the population of models. However, the accuracy of the models may not reflect the ability of our method to model the receptor because the ICL3 loop was mistakenly not rebuilt (i.e. the polypeptide chain was not connected between TMH5 and TMH6 due to the presence of T4 lysozyme in the B2AR template), preventing an optimal all-atom refinement of the receptor structure.
To further test whether our method consistently improves homolog templates, we selected a representative dataset of 50 membrane protein structure pairs exhibiting a wide diversity of sequence identity (from 15% to 70%), length (from 183 to 420 residues and topology (from 5 to 12 TMHs) (Methods, Table S1). In this dataset, 28 pairs were GPCRs (class A or B), 22 pairs were non-GPCRs and 37 pairs were distant homologs sharing not more than 25% of their sequences. In each pair, one structure was assigned as the target to be modeled and the other one as the starting template. 36 pairs exhibited poor sequence alignment for at least one TMH and required both TMH and loop rebuilding prior to all-atom refinement. Specifically, 21 GPCR pairs required sampling alternative conformation of one distorted TMH (Table S2), 15 non-GPCR pairs required at least one TMH to be rebuilt and the Lac permease/EmrD pair sharing only 15% sequence identity required all TMHs to be simultaneously rebuilt (Table S1). The models were selected by all-atom energy and clustering (Methods). The quality of the predictions was analyzed for their accuracy over the full-length, TMH structures and individual distorted TMH conformations. They were compared to the starting template and to models generated with the same input information (e.g. alignment, template structure) using 1. MEDELLER [30], a comparative modeling technique developed for membrane proteins, 2. the widely used MODELLER comparative modeling method [31] and 3. I-TASSER, a widely-used protein structure prediction server [32], [33]. As shown in Fig. 5 and Table S1, our method significantly improves starting templates for all but 4 protein pairs over the full length structure and for all but 3 protein pairs over the TMH regions. The average improvements as measured by GDT-HA over the entire dataset (i.e. High Accuracy Geometric Distance Test measuring similarity between two protein structures [34]) are 0.07±0.04 and 0.10±0.05 for the full length structure and the TMH regions, respectively, and are statistically significant (p values <0.005 and <0.0001, respectively, as measured by student t-test). These improvements are particularly noticeable in the TMH regions where the percentage of residues lying within 1 Å of the native structure is increased by 17±10% thereby decreasing the Cα RMSD from 2.1±0.7 to 1.7±0.7 Å in these regions. In contrast, the models generated by MEDELLER, MODELLER and I-TASSER remain very close to the starting templates and do not exhibit significant improvements as measured by GDT-HA over TMH regions: 0.002±0.01, −0.006±0.05 and −0.01±0.05, respectively (p values >0.5; Fig. 5, Table S1).
The absence of improvements in the TMH regions was observed for 3 close homolog pairs: 3PBL from 3EML, 4EJ4 from 3RZE and 2IC8 from 2NR9. For 3PBL from 3EML, the template is already very close to the target structure (Cα RMSD = 1.1 Å). At this level of structural similarity, inaccuracies in the energy function and the lack of explicit modeling of buried water molecules, lipids and ligands in the current method may impede further significant improvements.
Table S2 summarizes the local improvements on the distorted TMH2, which was rebuilt in 21 GPCR pairs because of the poor sequence alignment between the target and the template in that region. The overall conformation of the kinked TMH2 was improved for all but two pairs as measured by GDT-HA which increased from 0.79±0.05 to 0.86±0.06 and by Cα RMSD which decreased from 1.33±0.42 to 0.84±0.33 Å. Importantly, as shown for 3ODU from 2RH1 (Fig. 3A–C) and for 3EML from 1U19 (Fig. 3D–F) and in Table S2, the precise conformation of the kinked regions that were rebuilt de novo was also improved as measured by the differences in dihedral angles between template or model and native structures. When averaged over the bend (i.e. 5 residues, Methods), these differences decreased from 26±13° and 34±19° to 15±7° and 14±10°, for phi and psi backbone dihedral angles, respectively.
Modeling the unusually distorted TMH2 of squid rhodopsin (2Z73) [35] was challenging. Proline 90 perturbs and partially breaks the hydrogen bond network between the backbone nitrogen and carbonyl groups of residues 85 to 90 which form a wide π turn splitting TMH2 in two helical fragments. In addition to the wide π turn, the relative position of these helical fragments is also unusual. Unlike many kinked helices [21], the interhelical (kink) angle is only 21 degrees and the C-terminal helix is displaced outside of the TMH core compared to the N-terminal helix, a conformation stabilized by the beta-strand forming an extracellular “lid” over the retinal binding site. In absence of this loop region during TMH rebuilding, the native conformation of the C-terminal helix is not stabilized by a large number of physical contacts with the rest of the TMH core making the selection of that conformation difficult by energy alone. Although our protocol improved starting templates overall, we expect that rebuilding TMH core and loop regions simultaneously may become a more effective strategy for helical conformation stabilized by loop regions and will be explored in future work.
The largest improvements in full length structure and TMH regions (defined as GDT-HA increases ≥0.12) were mainly observed for distant homologues and include both GPCRs and non-GPCRs: 1U19 from 3ODU, 2RH1 from 1U19, 2CFQ from 2GFP and 3P5N from 3RLB. GDT-HA increases ≥0.12 in the TMH region were also mainly observed for distant homologs, such as 1U19 from 2Z73 or 3EML, 2RH1 from 2Z73 or 3ODU, 2Z73 from 3ODU, 3PBL from 3ODU, 3V2Y from 3RZE, 3EML from 3UON, 1U7G from 3B9W, 3P5N from 4DVE, 3V5U from 4KPP and 3GD8 from 3KLY.
Within the GPCR targets, modeling the beta2 adrenergic receptor (2RH1) from bovine rhodopsin (1U19) led to the largest improvements in GDT-HA: 0.13 and 0.19 over the full-length and TM structures, respectively. Although these 2 GPCRs share only 20% sequence identity in the modeled regions, 73% of the model residues lie within 1 Å of the native TM structures compared to only 28% for the starting template and display very similar side-chain conformations compared to in the native structure (Table S1, Fig. 2B,C). Most of the residues not predicted at atomic resolution belong to the extracellular part of the first TMH which, unlike in 1U19, is poorly packed to the rest of the TM structure in the B2AR crystal structure and is difficult to predict accurately.
Within the non-GPCR targets, the largest improvements in GDT-HA were observed for the ECF-type riboflavin transporter (3P5N) from thiamine-specific S-component ThiT from an ECF-type ABC transporter (3RLB). Although these 2 transporters share only 15% sequence identity in the modeled regions, the overall fold is conserved. Three TMHs poorly aligned with the template were rebuilt leading to 0.14 and 0.16 improvements in GDT-HA over the full-length and TM structures, respectively. In contrast to the template, most of the TMH region in the selected Rosetta model is superimposable to that of the target allowing a large fraction of side-chains to adopt similar packing than in the native structure (Fig. 2F). Similar improvements of starting templates leading to close to atomic accuracy backbone and near-native side-chain conformation predictions in the TM region were observed for other distant homolog pairs such as 2RH1 from CXCR4 (Cα RMSD of 1.7 Å, Fig. 2D) and 1U7G from 3B9W (Cα RMSD of 1.1 Å, Fig. 2E).
Although most of the largest improvements were obtained for distant homologs, the method was also able to improve starting templates for most of the closer homolog pairs that are structurally more similar. For example, improvements in GDT-HA ≥0.1 were observed for the pairs 1U19 from 2Z73, 2Z73 from 1U19, 1J4N from 1FX8, and 1L7V from 2NQ2 sharing more than 25% sequence identity in the modeled regions (Table S1). In the latter, reconstruction of distorted TMHs with different kink patterns between target and template allowed accurate prediction of backbone and side-chain conformations in the TM region (Cα RMSD of 1.1 Å, Fig. 2G).
When loop sequences are well aligned between template and target, their structures from the template are, as for TM regions, accurately refined at all-atom. In absence of significant sequence alignment with the template, loops are rebuilt de novo from sequence and accurately predicting their structures remains a challenge in the field of protein modeling. Three scenarios are typically encountered: 1) When loops are short (typically <8 residues) (e.g. kinks in distorted TMHs) or 2) When loops are long (typically ≥8 residues) but not only composed of disordered segments (i.e. incorporating a significant fraction of secondary structure elements), our approach can rebuild these regions from sequence with near-atomic accuracy (Cα RMSD within 2.5 Å). Examples include the blind-predicted extracellular loop 3 of CXCR4 (residues G220-I246) and the blind-predicted intracellular loop 2 of DRD3 (residues V109-T118) as well as several GPCR and non-GPCR loops in our benchmark (Fig. 4A–E). 3) Loops such as the extracellular loop 2 of GPCRs can be long and mostly disordered and/or make numerous contacts with small molecules or with other subunits in the crystal structures. Because crystal contacts or ligands are not modeled by the current method, near-native conformations of loops stabilized by such contacts are very difficult to select by energy alone. Therefore, although our blind predicted model of the long disordered extracellular loop 2 of CXCR4 was significantly more accurate than any other submitted model in the blind prediction, future developments (e.g. integrated loop modeling and ligand docking) will be necessary to consistently reach high-accuracy prediction in these regions and allow accurate prediction of ligand-bound conformations. Nevertheless, our results suggest that our method should be useful in rebuilding and refining X-ray structures of membrane receptors where functionally important loop regions have missing densities or are often deleted to facilitate crystallization.
An important question in the field of protein modeling is the relationship between the accuracy of the models and their potential applications. Near-atomic resolution models should be accurate enough to guide the rational design of mutations and the interpretation of their effects [4]. As a stringent test of the accuracy of our predictions, we subjected the selected models from our benchmark to complete sequence redesign in the TMH regions and compared the results to similar calculations performed with the native X-ray and initial template structures (Methods). Single-state design calculations select combinations of amino acids that minimize the free energy of (i.e. predicted to stabilize) the protein. Previous sequence calculations performed on high-resolution transmembrane helical protein X-ray structures recapitulated a significant fraction of native sequences [24], suggesting that this fraction of residues is naturally selected for stability. Because the physical interactions underlying the selection of amino acids are very sensitive to the atomic details of the structure, the level of native amino acid recovery should be indicative of the accuracy of the protein structure. While redesigned template structures recovered only 23±6% of native amino acid sequences, redesigned X-ray and selected model structures recovered 35±10% and 42±8% of native amino acid sequences, respectively. Only 41±13% of the native residues recovered in redesigned templates were also recovered in redesigned X-ray structures. By contrast, 72±7% of the native residues recovered in redesigned selected models were also recovered in redesigned X-ray structures. These results indicate that the native interactions recovered in redesigned X-ray and selected model structures are similar, and suggest that the TMH regions of protein models generated using our method are in a range of accuracy suitable for rational design applications.
The prediction of membrane protein structures represents an important approach in light of their difficult experimental determination but remains a challenging problem. Current prediction techniques are limited to the generation of low-resolution models from sequence information alone [10]–[12] or of near-atomic resolution models from close structural homologues [5]. However, close structural homologs are currently not available for a large fraction of membrane proteins and often only one distant structural homolog hit can be found for these proteins (Fig. S1), making their structure prediction at high-resolution a real challenge. To address this problem, we developed a generic method that can efficiently reconstruct TMH and loop regions from single distant or closer homologues. The method was stringently validated in two blind predictions and in a large benchmark consisting of pairs of membrane protein homologues with wide diversity in length, topology and sequence identity. Submitted models were first-ranked in the blind predictions [14], [15] for the accuracy of the full-length receptor structure and the method was able to improve most starting templates in the benchmark to reach near atomic accuracy prediction in the TMH regions (Fig. 2, Fig. 5, Table S1). In local regions of the TMH structures where distortions differed between template and target, the method was able to significantly improve the starting template and to predict distorted helical structures with an average Cα RMSD of only 0.8 Å to the native structures (Fig. 3, Table S2). As a stringent proof of the model's accuracy, complete redesign of their TMH regions recapitulated similar native interactions than the redesign of the same regions in the X-ray structures. In contrast, the methods MEDELLER [30], a comparative modeling technique developed for membrane proteins, the widely used homology modeling software MODELLER [31], and I-TASSER, a web-server for protein structure prediction [32], [33], did not significantly improve homologous templates (Fig. 5).
The improvements observed for most distant or closer homologues with diverse length and topology indicate that the method provides a general and efficient approach for reconstructing the structure of a large diversity of transmembrane helical folds. Starting templates with sequence identity to the target as low as 15% were significantly improved, suggesting that the technique should be effective at generating atomic-level models more accurate than available templates for many structurally uncharacterized TMH proteins (Fig. S1).
Because the conformational heterogeneity and poor stability of eukaryotic membrane proteins in detergents is a major bottleneck to their crystallization, their stabilization has been a very intensive area of research but has only been achieved with limited success using labor-intensive cycles of random or scanning mutagenesis [36]–[38]. According to our design calculations, our technique can predict stabilizing physical interactions in structurally uncharacterized receptors and should therefore be particularly useful for predicting mutational effects on receptor's conformational stability, for engineering receptors with altered conformational energy landscape and for precisely guiding structure/function studies.
Future developments will involve 1) the explicit modeling of water molecules to improve the prediction of TMH core regions, and 2) the simultaneous modeling of loop and bound ligand conformations to improve the prediction of loop structures and allow accurate prediction of receptor-ligand bound conformations and interactions for ligand docking and virtual screening applications.
In conclusion, the method may prove useful for the atom-level modeling and design of structurally uncharacterized classes of alpha-helical membrane receptors which are particularly challenging to study experimentally and for which close homologues are currently often not available.
To analyze the coverage potential of homology modeling of membrane proteins, HHpred [25], a toolkit for searching and aligning query sequences with sequences from existing structures, was run on three datasets of human transmembrane helical proteins. Two datasets were taken from the Survey of the Human Transmembrane Proteome [39] and consisted in: 1) full-length sequences of human transmembrane proteins with at least two predicted transmembrane helices (3838 sequences), and 2) full-length sequences of human transmembrane proteins with at least two predicted transmembrane helices truncated to the transmembrane domain (i.e. from the first to last predicted transmembrane helix residues) (3838 sequences). Additionally, a full-length human transmembrane proteome dataset (6521 sequences) was created by supplementing the aforementioned 3838 full-length multi-pass sequences with 2683 human single-pass transmembrane helical proteins from Uniprot database [40]. Each of these datasets were clustered at 98% sequence identity using USEARCH [41], yielding non-redundant dataset sets of 3405, 3079, and 5818 for the full-length human multi-pass transmembrane helical proteins, transmembrane domain truncated human multi-pass transmembrane helical proteins, and full-length combined single- and multi-pass human transmembrane helical proteins, respectively. These were used as inputs to HHpred search for structurally characterized homologs. HHpred was run by using the HHsuite programs HHblits [26] (to generate HMM alignment from searching Uniprot database) and HHsearch [27] (to match the HMM-HMM alignment to PDB database). DSSP [42] and Psipred [43] were used for secondary structure prediction annotation as part of the HHpred protocol. The resultant HHpred alignments were filtered by a range of percent sequence identity thresholds (i.e. of homolog hit versus query) and percent coverages (i.e. of total length of query sequence) of 90%, 75%, 60% or 50%.
A representative dataset of 50 membrane protein structure pairs was selected that samples a wide range of sequence identity (from 15% to 70%), length (from 183 to 420 residues) and topology (from 5 to 12 TMHs). As outlined below, the dataset was selected to be representative of the entire classes of membrane proteins that can be modeled using the method described in this study.
Membrane protein targets were selected by filtering the OPM database with the following criteria that reflect the current scope of the method. Firstly, selecting for “transmembrane” and “alpha-helical polytopic”, 936 proteins in 75 superfamilies were identified. Next, families were removed that 1) have less than two unique protein structures (need at least one homolog) –or– 2) consist of multi-protein complexes –or– 3) consist of very large proteins (>15 secondary structure elements or >600 residues) –or– 4) contain large cofactors (e.g. heme groups) –or– 5) formed from many symmetrical subunits. This reduced the number of superfamilies to 18. Additionally, four of the remaining superfamilies did not contain proteins with structurally characterized homologs with sequence identity >15% and were also removed. The remaining 14 superfamilies are the following (as categorized by OPM database): 1) Rhodopsin-like receptors and pumps, 2) ABC transporters, 3) General secretory pathway, 4) Major Intrinsic Protein, 5) Ammonia and urea transporters, 6) Major Facilitator Superfamily, 7) APC (Amino acid-Polyamine-organoCation) superfamily, 8) Monovalent cation-proton antiporter, 9) Chloride transporter, 10) Multidrug/Oligosaccharidyl-lipid superfamily, 11) Energy-coupling factor transporters, 12) Rhomboid protease, 13) Sodium/calcium exchanger, and 14) Peptidase family M48. Our dataset of modeling targets covers 12 of 14 superfamilies. The available target/template homologs for the Monovalent cation-proton antiporters and the Peptidase family M48 are too distant (structural alignment between template and target is extremely poor: Calpha rmsd = 25 Å) and too homologous (38% identity), respectively, to be considered relevant for this study. In total we selected 50 representative modeling cases combining different target/template pairs, and 31 unique targets. Of our modeling targets, 12 are GPCRs (11 Class A and 1 Class B) and 19 are non-GPCRs membrane proteins.
The following X-ray structures and corresponding pdb codes were selected from the protein database:
Several methods including the consensus method 3D-Jury [45] and HHpred [25] based on HMM-HMM comparisons were tested to generate optimal sequence-sequence alignments. HHpred gave the best alignments in our benchmark and was subsequently used for all predictions. The following parameters were used: ten PSI-BLAST iterations with an E-value threshold of 1E-3, local alignment with global final realignment. For the blind predictions, the best alignment was systematically considered to select homologues and construct templates. For a few of the most distant pairs (3L1L from 3GIA, 2CFQ from 2GPF, 3P5N from 4DVE, 3GIA from 4DJK, 2GPF from 2CFQ, 3KLY from 3GD8, 3GD8 from 3KLY, 3HD6 from 1U7G, 3VVO from 3MKT and 4HZU from 3RLB), the sequence alignment generated by HHpred was adjusted manually, guided by topology prediction of TMHs given by Octopus [46] and secondary structure prediction given by Psipred [43], to improve the alignment of the TMH region and minimize the number gaps or insertions in this region.
The template structures and alignments between template and target sequences for each protein pair in the benchmark were used as inputs to the Homology Modeling software MEDELLER [30]. MEDELLER was run using the online MEDELLER server (http://opig.stats.ox.ac.uk/webapps/medeller/home.pl?app=MEDELLER) with default settings to generate “complete” models. The MEDELLER server does not provide a benchmarking option that excludes the target structure from its loop modeling process, which uses FREAD [39], a database search loop modeling algorithm. Therefore, all models generated by MEDELLER were checked for loops that incorporated fragments from the target structure. For all but two protein pairs in the benchmark, the complete models generated by the MEDELLER server did not include target loops and were directly used for analysis. The two MEDELLER models (3EML from 3UON and 1U7G from 3B9W) that included target loops were run again on the online FREAD server (http://opig.stats.ox.ac.uk/webapps/fread/php/index.php) and the best loop fragment hits excluding those from the target were used for analysis.
Homology modeling with MODELLER [31] was run using an online MODELLER server (http://toolkit.tuebingen.mpg.de/modeller) with default settings. The template structures and alignments between template and target sequences for each protein pair in the benchmark were used as inputs.
The I-TASSER server (http://zhanglab.ccmb.med.umich.edu/I-TASSER/) was provided with the same target sequence, target/template alignment and template structure than Rosetta, MODELLER and MEDELLER (option I: Specify template with alignment). To ensure that I-TASSER would not use any additional homolog templates closer to the target than the one assigned in each protein pair of the benchmark, other templates with sequence identity higher than 25% to the target or closely related to the homolog template assigned in each protein pair were excluded (option II: Exclude homologous templates/Exclude specific template proteins). I-TASSER usually generated 5 models and the most accurate one is reported in our study.
The method consists of three parts: 1. Rebuilding of TMH structures, 2. Rebuilding of non-TMH (e.g. loops, helical bends) structures, 3. All-atom refinement of reconstructed structures.
1. Rebuilding of TMH structures is performed if 1) gaps in the sequence alignment occur in these regions, 2) bends have different predicted positions (e.g. unaligned Prolines or coil motifs, non-conservation of Prolines between template and target sequences) or 3) TMHs have different predicted lengths (i.e. significantly different secondary structure prediction) indicating potential different tilt angles with regards to the membrane plane. Concerning the prediction of residues promoting helical bends, we limited ourselves to the presence of prolines in the target or in 10% of the homolog sequences which, depending on the membrane protein structure databases analyzed, account for between 60% [21] and 90% [18] of TMH kinks. Sequence motifs other than prolines have been reported to induce helical bends but current sequence-based predictions do not exhibit a combined sensivity/specificity high enough to be used as an automated input in the rebuilding of TMHs. Even if they cannot be identified by sequence or secondary structure information alone, helical bends and distortions promoted by local strain in the backbone structure or by specific tertiary interactions can still be identified and modeled during the all-atom structure refinement stage.
Rigid-body helical degrees of freedom of TMHs to be rebuilt are sampled based on a kinematic description of the polypeptide chain where the protein system is represented in internal coordinates by a tree of atoms which can have any structure provided there is no closed loop [10], [47]. The atom-tree representation was further developed so that the edges in the tree can be any bond connections or rigid body transformations, making the protein a single continuous bonded chain or multiple domains connected by virtual long-range “jumps” between residues. This new atom-tree representation allows torsional and rotameric sampling within each individual TMH segment as well as perturbations in the rigid body degrees of freedom around the “jump” connecting these segments. Loops and local distorted regions connecting full-length or fragments of TMHs to be rebuilt are stripped out from the template and alternative TMH conformations are generated by randomly sampling rigid body degrees of freedom along and off the helical axis. At this stage, the protein template is represented at the coarse-grained level where side-chain conformations are averaged out, thereby drastically decreasing the number of degrees of freedom to be sampled. Moves are accepted using a Metropolis Monte-Carlo criterion (1000 to 5000 steps for each TMH fragment constrained by a Gaussian function to 1–1.5 Å of the starting structure) and followed by loop rebuilding and full structure gradient-based minimization (see below).
More specifically, as shown in Fig. 1, kinked TMHs are represented by two TMH fragments and a distorted helical turn around the kink. Each helix fragment is treated as a rigid-body and defined by a helical axis (m for “moveable” and f for “fixed” defining the reference state). Following the distribution of kink angles and distances between TMH fragments of kinked TMHs in membrane protein structures, the two TMH fragments adopt relative orientations that are constrained in space. The m helix is moved with regard to the f helix according to the following degrees of freedom:
In addition to these constrained moves, each TMH fragment is allowed to spin around its helical axis. Finally, the f and m helices are also allowed to move as a single unit and to sample the conformational degrees of freedom of a standard alpha helix rigid body.
2. Non-TMH (e.g. loops, helical bends) structures with low sequence identity to the template or exhibiting gaps/deletions in the sequence alignment with the template are rebuilt de novo [10]. This step follows the previously developed de novo folding protocol for membrane protein structures and involves random peptide fragment insertions subjected to acceptation by the Metropolis criteria based on the total energy of the system. At this stage, the system is still represented at the coarse-grained level and the low-resolution energy function of Rosetta is used to compute the energy of the system. Cyclic coordinate descent (CCD) is used to close the chain break in the rebuilt region and to maintain the connectivity of the protein chain, and is achieved by iteratively inserting fragments and increasing the chain break penalty. If after twelve rebuilding steps, any chain break remains larger than 0.2 Å, the region to be rebuilt is expanded by one residue on both sides until a continuous peptide chain is recovered. The libraries of fragments to be inserted are generated for fragments of size 9 and 3. Fragments of larger size were tested but didn't provide any significant improvements in the accuracy of the rebuilt regions.
Helical bends in kinked TMHs are typically modeled as four residues loop insertion connecting two helical fragments and can sometime result in distorted loop conformations which are not usually observed in native kinked helices. Such local structures involve either a combination of non-helical turns and 310 helix or helical distortions extending 2 or 3 residues Cterminal to the residue responsible for the bend. In such situation, starting from the selected all-atom refined model, a larger window of residue (e.g. 5 to 8) is rebuilt and locally refined using the loop modeling protocol.
3. The fragment insertion protocol described above involves fragment insertion moves that sample a large conformational space to identify a broad range of physically-realistic conformations. The coarse-grained models are then subjected to all-atom refinement which searches the all-atom conformational energy landscape for local minima in the vicinity of these structures. This step combines an all-atom energy function developed for transmembrane protein structures with an efficient search for low-energy conformations. As described previously [24], the energy function mainly consists of short-range interactions, e.g. Lennard-Jones, hydrogen-bond. Knowledge-based potentials describe torsional states of both backbone and side-chain atoms and the solvation energy of each atom as a function of both its depth in the membrane and its burial in the protein. A Monte-Carlo minimization procedure with discrete side-chain optimization is used to efficiently sample low-energy conformations in the rugged all-atom energy landscape. A single move involves the following steps: 1) random backbone perturbations, 2) discrete side-chain optimization for the new backbone conformation, 3) minimization of the energy of the system with respect to all conformational degrees of freedom. Several cycles of small backbone perturbations are first applied to the entire receptor starting with a smooth Lennard-Jones potential followed by an iterative ramping up of the repulsive part of the potential. This procedure allows a smooth transition from a coarse-grained to a full-atom representation without loosing the compactness of the initial structures.
To avoid the sampling of conformational space unlikely occupied by the target sequence, the all-atom energy function is also supplemented by a constraint potential maintaining conserved regions that are in vicinity in the template structures. These constraints are defined between pairs of strictly conserved or similar residues in both target and template sequences and that are in vicinity in the template structures. In our calculations, a constraint is defined by a distance between the Cα atoms of the interacting residue pairs and a constraint width (i.e. the deviation from the assigned distance at which the constraint score begins to ramp up). Any deviation from these distances during refinement is penalized by a harmonic potential. Small constraint widths were assigned for short-range contacts (e.g. 0.2 Å for contacts ≤5 Å) while larger constraint widths were assigned for longer-range constraints (e.g. 0.5 Å for contacts of ∼8 Å). To avoid over constraining the models to the starting template, the average number of selected constraints was around 5% of the total number of residues for the most distant homolog pairs (i.e. sequence identity of 15–20%) and between 5% and 10% for intermediate homolog pairs (i.e. sequence identity of 20–25%). For closer homolog pairs (i.e. sequence identity >25%), models were highly constrained to the starting templates at most positions not rebuild de novo.
Between 10000 and 40000 all-atom refined models are generated per target. 1000 or up to 10% lowest energy structures are selected and their transmembrane region clustered into structurally-related families using Rosetta's clustering protocol. The most accurate model among the five lowest-energy structures which cluster in one of the five largest families of models is selected and discussed in this study. Accuracy of the models to the target structure is computed using TM-SCORE [48] over full length and TM regions.
Sequence design calculations were performed as described previously [24], [49]. Briefly, the backbone coordinates from the X-ray structure, the selected model and the initial template were selected to perform the design calculations. All 20 amino acids were allowed at the TMH positions and the native residues were kept in the loop regions. The combination of amino acids and side-chain conformations minimizing the free energy of the system was selected by Monte Carlo sampling of discrete side-chain conformations (i.e. rotamers) followed by energy minimization over all conformational degrees of freedom. The Dunbrack rotamer library [50] expanded by rotamers at +−1 standard deviation around the mean values for the dihedral angles chi1 and chi2 was used to repack the structures. The energy of each structure was computed using the all-atom RosettaMembrane energy function [24]. 100 independent design calculations were performed starting from each individual backbone structure. The percentage of native sequence recovery was calculated from the lowest energy designed structures.
|
10.1371/journal.pgen.1007337 | Threshold-dependent repression of SPL gene expression by miR156/miR157 controls vegetative phase change in Arabidopsis thaliana | Vegetative phase change is regulated by a decrease in the abundance of the miRNAs, miR156 and miR157, and the resulting increase in the expression of their targets, SQUAMOSA PROMOTER BINDING PROTEIN-LIKE (SPL) transcription factors. To determine how miR156/miR157 specify the quantitative and qualitative changes in leaf morphology that occur during vegetative phase change, we measured their abundance in successive leaves and characterized the phenotype of mutations in different MIR156 and MIR157 genes. miR156/miR157 decline rapidly between leaf 1&2 and leaf 3 and decrease more slowly after this point. The amount of miR156/miR157 in leaves 1&2 greatly exceeds the threshold required to specify their identity. Subsequent leaves have relatively low levels of miR156/miR157 and are sensitive to small changes in their abundance. In these later-formed leaves, the amount of miR156/miR157 is close to the threshold required to specify juvenile vs. adult identity; a relatively small decrease in the abundance of miR156/157 in these leaves produces a disproportionately large increase in SPL proteins and a significant change in leaf morphology. miR157 is more abundant than miR156 but has a smaller effect on shoot morphology and SPL gene expression than miR156. This may be attributable to the inefficiency with which miR157 is loaded onto AGO1, as well as to the presence of an extra nucleotide at the 5' end of miR157 that is mis-paired in the miR157:SPL13 duplex. miR156 represses different targets by different mechanisms: it regulates SPL9 by a combination of transcript cleavage and translational repression and regulates SPL13 primarily by translational repression. Our results offer a molecular explanation for the changes in leaf morphology that occur during shoot development in Arabidopsis and provide new insights into the mechanism by which miR156 and miR157 regulate gene expression.
| Leaves produced at different stages in the development of an Arabidopsis shoot vary predictably in shape and size. Previous studies have shown that this phenomenon is regulated by variation in the abundance of the miRNAs, miR156 and miR157, but how miR156/miR157 produce the changes in leaf morphology that occur during shoot development is not understood. To answer this question, we measured the abundance of miR156/miR157 and their SPL targets in successive leaf primordia, and characterized the effect of variation in the abundance of miR156/miR157 on leaf morphology and the abundance of SPL transcripts and SPL proteins. miR156/miR157 are present at very high levels in the first two rosette leaves, where they act as buffers to stabilize leaf identity. They are present at lower and steadily declining levels in subsequent leaves, where they act to modulate leaf morphogenesis. In these later-formed leaves, a small decrease in the abundance of miR156/miR157 produces a disproportionately large increase in SPL activity, primarily as a result of the increased translation of SPL transcripts. Our results provide a new view of vegetative phase change in Arabidopsis and the mechanism by which miR156 and miR157 regulate this process.
| Leaves produced at different times in shoot development are often morphologically distinct. In Arabidopsis, for example, successive rosette leaves differ in size, length:width ratio, the angle of the leaf base, hydathode number, the complexity of the vascular system, cell size, sensitivity to gibberellic acid, and the absence vs. the presence trichomes on the abaxial surface of the leaf blade [1–7]. Some of these so-called "heteroblastic" traits change gradually throughout shoot development, others change early in shoot development and are then expressed more-or-less uniformly, while still others are present at one stage of development and absent at a different stage. These latter two patterns allow the shoot to be divided into several discrete phases, the transition between which is termed "vegetative phase change" [8].
miR156 is the master regulator of vegetative phase change in Arabidopsis [3,9] and other flowering plants [10–15]. It is initially expressed at a very high level, and declines as the shoot develops [3,9,16–18]. This decrease is associated with an increase in the expression of its targets, SQUAMOSA PROMOTOR BINDING PROTEIN-LIKE (SPL) transcription factors, and is responsible for the transition to the adult phase. Most species also possess another miRNA, miR157, that differs from miR156 at 3 nucleotides [19]. miR157 has the same targets as miR156 and produces an over-expression phenotype similar to that of miR156 [20]. However, the normal function of miR157 is still unknown.
Although it is clear that miR156, and possibly miR157, regulate many of the changes that occur during shoot development, the function of these miRNAs at specific times in development and in specific leaves is poorly understood. In particular, it remains to be determined if miR156 is responsible for the graded changes in leaf morphology that occur during the juvenile phase and, if so, how it produces this variation. It is also important to determine if miR156 plays a role in shoot morphogenesis during the adult phase. Although miR156 is present at lower levels in the adult phase than in the juvenile phase, a comparison of the expression patterns of miR156-sensitive and miR156-insensitive SPL reporters suggests that miR156 represses SPL gene expression during both phases, albeit to different extents [21]. Finally, it is important to determine the mechanism by which miR156 represses gene expression. Previous studies have shown that miR156—as well as several other plant miRNAs (reviewed in [22])—mediates both transcript cleavage and translational repression [23–27], but the relative importance of these processes for the activity of miR156 remains to be determined. This question is of particular interest in light of the observation most SPL transcripts change very little during shoot development, despite the significant decrease in miR156 that occurs during this process [21].
To address these questions, we characterized the morphological and molecular phenotype of loss-of-function mutations in MIR156 and MIR157 genes, and measured the absolute amount of miR156/miR157 in successive leaf primordia. We also quantified the effect of varying miR156/miR157 levels on the expression of their SPL targets. Our results demonstrate that miR156 and miR157 have different expression patterns, different activity, and mediate transcript cleavage and translational repression to different extents at different SPL genes. We also show that variation in the level of miR156/miR157 only has a significant effect on SPL gene expression when these miRNAs are present at relatively low levels. These results provide a foundation for detailed studies of the molecular mechanism of miR156/miR157 activity and their role in shoot morphogenesis.
In Arabidopsis, miR156 is encoded by 8 genes and miR157 is encoded by 4 genes. We characterized the contributions of these genes to the overall pool of miR156/miR157 by sequencing small RNAs from the FRI FLC and FRI flc-3 genotypes [28,29]. We chose these genotypes because they represent common genotypes in naturally-occurring accessions of Arabidopsis [30], and because the vegetative and flowering phenotype of FRI flc-3 is nearly identical to that of wild-type Col [2, 31]. Sequencing of small RNAs from 11-day-old shoot apices (2 replicates of each genotype) revealed an abundant 20 nt transcript that maps to MIR156A, B, C, D, E, and F, an abundant 21 nt transcript that maps to MIR157A, B, and C, a moderately abundant 21 nt transcript that maps to MIR156D, and 3 rare transcripts that map uniquely to MIR156G, MIR157D and MIR156H (Table 1). Unexpectedly, miR157-related transcripts were more abundant than miR56-related transcripts.
To determine which genes produce these transcripts, we identified T-DNA insertions in MIR156A, MIR156C, MIR156D, MIR157A, and MIR157C, and used site-directed mutagenesis to produce mutations in MIR156B. RT-qPCR analysis of these alleles demonstrated that they eliminate or greatly reduce the primary transcripts of the affected genes (S1 Fig). We then examined the amount of miR156 and miR157 in these stocks by hybridizing RNA blots with probes for miR156, miR157, and a combination of both probes. We used this approach instead of RNA sequencing because libraries constructed with two different RNA adaptors revealed that different miRNAs ligate with different efficiencies to each adaptor [32]. Although the miR156 and miR157 probes cross-hybridize to some extent, the source of the hybridization signal could by determined by comparing the effect of mir156 and mir157 mutations on these signals.
The effect of mir156 and miR157 mutations on the levels of miR156 and miR157 in 11-day-old seedlings and in 1mm primordia of leaves 1 & 2 is shown in Fig 1. In Col, the miR156 probe hybridized to 20 nt and 21 nt transcripts, with the 20 nt transcripts being more abundant than the 21 nt transcripts (Fig 1A). The abundance of the 20 nt transcripts was reduced to 62 ± 10% (± SD, n = 4) of wild-type in mir156a-2 (hereafter, mir156a), to 51 ± 8% (± SD, n = 4) of wild-type in mir156c-1 (hereafter, mir156c), and to 11 ± 1% (± SD, n = 3) of wild-type in the mir156a/c double mutant (Fig 1A). These genes are therefore the major source of the 20 nt miR156 transcripts. mir156d-1 (hereafter, mir156d) had very little effect on the overall abundance of miR156 in 11 day-old seedlings and leaf primordia (Fig 1). However, the intensity of the 21 nt band was slightly reduced in mir156d-1 and in genotypes containing this mutation; for example, the 21 nt miR156-hybridizing band was slightly less intense in the mir156a/c/d mir157a/c pentuple mutant than in the mir156a/c mir157a/c quadruple mutant (Fig 1A). MIR156B also makes a minor contribution to the miR156 pool because the intensity of the miR156-hybridizing bands was essentially identical in the mir156a/b/c/d and mir156a/c mutants (Fig 1A), and there was no detectable difference between the intensity of the 20 nt and 21 nt bands in leaf primordia (LP) of the mir156a and mir156a/b mutants (Fig 1B).
In wild-type Col, the miR157 probe hybridized strongly to 21 nt transcripts and more weakly to 20 nt transcripts (Fig 1A). The 20 nt band was absent in mir156a/c, and thus represents cross hybridization of the miR157 probe with miR156. The intensity of the 21 nt miR157-hybridizing band was reduced to 81 ± 13% (± SD, n = 4) of wild-type in mir157a-1, to 26 ± 5% (± SD, n = 7) of wild-type in mir157c-1 (mir157c), and to 13 ± 2% (± SD, n = 3) of wild-type in the mir157a/c double mutant. The remaining 21 nt signal in the mir157a/c double mutant partly reflects cross-hybridization of the miR157 probe with 21 nt miR156 transcripts because the intensity of this band was slightly reduced in the mir156a/c mir157a/c and mir156a/c/d mir157a/c mutants compared to mir157a/c. These results demonstrate that the major miR157 transcript is 21 nt, and that MIR157C is the major source of this transcript.
Hybridization with a 1:1 mixture of miR156/miR157 probes revealed that 21 nt transcripts are significantly more abundant than 20 nt transcripts in 11 day-old seedlings and in the primordia of leaves 1&2 (Fig 1A and 1B). The 21 nt band was reduced significantly in mir157a/c, and therefore corresponds primarily to miR157, whereas the 20 nt band was nearly absent in mir156a/c, and therefore corresponds to miR156. These results are consistent with the results of RNA sequencing (Table 1), and demonstrate that miR157 is more abundant than miR156 in young seedlings. Northern analysis using a mixed miR156/miR157 probe revealed that the amount of miR156 and miR157 in the mir156a/c/d mir157a/c pentuple mutant is about 10% of the wild-type level (Fig 1A and 1B). Assuming that the mutations present in this pentuple mutant are null alleles, the amount of miR156/miR157 in this line represents the combined output of MIR156E,F,G,H and MIR157B,D. These six genes therefore contribute relatively little to the production of miR156 and miR157 in seedlings.
The morphology of rosette leaves changes qualitatively and quantitatively during shoot development. In plants grown in SD to delay flowering, the first two rosette leaves are small and round, and lack serrations and abaxial trichomes [1,2] (Fig 2A). Leaves 3 and 4 are larger than leaves 1 and 2, but also have round leaf blades with no serrations and no abaxial trichomes. Leaves 5 through 9 are larger, more elongated, and more serrated than the first four leaves. Depending on light quantity and quality, abaxial trichome production begins between leaf 7 and 9, and is accompanied by a decrease in the angle of the leaf base and by the production of more prominent serrations (Fig 2B). Previous studies have shown that the juvenile forms of these traits require the activity of miR156/miR157 [9], but the relationship between the abundance of these miRNAs and the changes in leaf morphology that occur during shoot development is still unknown. To begin to answer this question, we measured the abundance of miR156 and miR157 in successive rosette leaves of wild-type plants, and characterized the effect of mir156 and mir157 mutations on leaf morphology.
RT-qPCR (S2A Fig) and Northern analysis (S2B Fig) demonstrate that miR156 and miR157 increase as leaves expand. However, the expression pattern of these miRNAs in successive fully expanded leaves (S2C and S2D Fig) and 1 mm LP (Fig 2) is quite similar, indicating that the factors responsible for variation in the expression of miR156/miR157 during shoot development operate at all stages of leaf development. Consistent with our previous analyses of shoot apices [21], miR156 and miR157 decrease significantly from LP1&2 to LP3&4, and then decline more gradually before reaching a relatively constant level around leaf 13 (Fig 2). LP3&4 had approximately 25%, LP9 had 12%, and LP13 had 8% of the amount of miR156 present in LP1&2. miR157 declined to a lesser extent: LP 3&4 had approximately 50%, LP9 had 25%, and LP13 had 17% of the amount of miR157 present in LP1&2 (Fig 2). The expression pattern of miR156 in fully-expanded (FE) leaves matched its expression pattern in LP, but miR157 did not decline as dramatically between FE1&2 and FE3&4 as it did between LP1&2 and LP3&4 (S2C and S2D Fig).
We then determined the absolute amount of these miRNAs in LP by comparing the RT-qPCR results obtained with leaf samples to the results obtained using known quantities of miR156 and miR157. Synthetic miR156 and miR157 transcripts were serially diluted in 600ng/μl E.coli RNA, and a standard curve was produced by plotting the concentrations of these miR156 and miR157 standards against 2-ct of the corresponding RT-qPCR reaction. RT reactions were performed in parallel using 600ng of total RNA from LP1&2. The 2-ct value of the LP1&2 sample was then fitted to the standard curve, and the concentration of miR156 or miR157 was calculated using linear regression. This information, and the results of the experiment shown in Fig 2A, were then used to calculate the absolute amount of miR156 and miR157 in other LP (Fig 2B). miR156 was present in LP1&2 at a concentration of 1.96 ± 0.1 x 105 copies per ng total RNA, whereas miR157 was present at a concentration of 2.45 ± 0.2 x 105 copies per ng total RNA (Fig 2B). miR156 subsequently declined to approximately 2.6 x 104 copies per ng total RNA in LP9, whereas miR157 declined to 6.1 x 104 copies per ng total RNA. Thus, the transition between leaves 1&2 and leaves 3&4 is accompanied by a major decline in the level of miR156 and miR157 whereas subsequent changes in leaf morphology are associated with much smaller changes in the abundance these miRNAs. The juvenile-to-adult transition occurred during the period when miR156 and miR157 were declining very gradually, and was accompanied by a relatively small change in the abundance of these transcripts.
The relative importance of different MIR156 and MIR157 genes in shoot development was determined by characterizing the morphological phenotype of plants singly or multiply mutant for mir156a, mir156b, mir156c, mir156d, mir157a and mir157c (Figs 3 and 4). Plants were grown in SD to eliminate the effect of floral induction on leaf morphology [2]. We measured two traits that change with leaf position—the production of trichomes on the abaxial surface of the leaf blade and the angle of the leaf base. In wild type plants, the angle of leaf base became more acute starting with leaf 5, and abaxial trichome production started at leaf 9 (Figs 3 and 4).
The effect of mir156 and mir157 mutations on abaxial trichome production and leaf shape was correlated with abundance of miR156/miR157 in different leaves. Leaves produced early in shoot development, which have a relatively high level of miR156/miR157, were less sensitive to these mutations than leaves produced later in shoot development, which have a relatively low level of miR156/miR157 (Figs 3 and 4). For example, mir156a, mir157a, mir157c, and mir157a/c caused abaxial trichomes to be produced on leaves 7 and/or 8, but did not affect abaxial trichome production or the shape of leaf 1, 3, and 5. mir156c produced abaxial trichomes on leaves 7 and 8 and significantly reduced the angle of the leaf base in leaves 3 and 5, but had no effect on leaf 1. miR156a/c and mir156a/b/c/d reduced the angle of the leaf blade in leaves 1, 3 and 5, but had a more significant effect on leaves 3 and 5 than on leaf 1; these genotypes only produced abaxial trichomes on leaves 6 and above. miR156a/c mir157a/c and mir156a/c/d mir157a/c had a significant effect on the shape of leaves 1, 3, and 5, but rarely produced abaxial trichomes on leaf 2, and never produced abaxial trichomes on leaf 1 (Figs 3 and 4). The absence of abaxial trichomes on leaves 1 and 2 is attributable to the small amount of miR156/miR157 remaining in these mutants because 35S::MIM156 consistently produced abaxial trichomes on both of these leaves (Fig 4A). These results demonstrate that abaxial trichome production is more sensitive to miR156/miR157 than leaf morphology, and is strongly repressed by even low levels of these miRNAs. They also reveal that the amount of miR156/miR157 in leaves 1&2 far exceeds the amount required to specify their identity. Only genotypes with very low levels of miR156/miR157 (e.g., miR156a/c mir157a/c, mir156a/c/d mir157a/c, 35S::MIM156) cause these leaves to resemble adult leaves (Figs 3B and 4B).
In general, the morphological phenotype of mir156/mir157 mutations was correlated with their effect on the abundance of miR156 or miR157. mir156a and mir156c have a relatively large effect on the level of mir156 (Fig 1) and also have a relatively large effect on shoot morphology. mir156c has a more significant effect on the morphology of leaves 3 and 5 than mir156a (Fig 4B), which is consistent with its slightly larger effect on the abundance of miR156 (Fig 1A). mir156b and mir156d have very minor effects on the abundance of mir156 (Fig 1) and also have minor effects on shoot morphology; mir156b did not significantly enhance the phenotype of mir156a or mir156a/c, and mir156d only produced a significant effect on leaf morphology in combination with mir156a/c and mir157a/c. The only unexpected result was the phenotype of mir157a/c. miR157 is more abundant than miR156 and was therefore expected to play a larger role in vegetative phase change than miR156. However, mir157a/c had a significantly weaker effect on abaxial trichome production and leaf shape than mir156a/c (Figs 3 and 4), even though these double mutants have approximately the same amount of miR157 and miR156, respectively (Fig 1). This observation demonstrates that miR157 is less important for vegetative phase change than miR156, and suggests that it may be less active than miR156.
miRNAs with a 5’ terminal uridine, such as miR156 and miR157, repress the expression of their targets via their association with AGO1 [33]. To determine if miR156 and miR157 are loaded onto AGO1 with different efficiencies, we measured the amount of miR156 and miR157 associated with AGO1 in planta. For this purpose, we took advantage of an ago1-36 line transformed with AGO1-FLAG [34]. Extracts from 2-week-old seedlings of this transgenic line and wild-type Col (as negative control) were treated with an antibody to the FLAG epitope, and small RNAs were extracted from immunoprecipitad (IP) and non-IP samples and assayed using Northern blots. Hybridization with a mixed miR156/miR157 probe revealed that miR157 (21 nt band) was more abundant than miR156 (20 nt band) in the input fraction, but that miR156 was as abundant as miR157 in the IP fraction (Fig 5). This result indicates that miR156 is more efficiently loaded onto AGO1 than miR157.
This cannot be the only reason for the difference in the phenotypes of mir156a/c and mir157a/c because the amount of miR156 and miR157 associated with AGO1 is quite similar. Another possibility is that the AGO1-miR157 complex is inherently less active than the AGO1-miR156 complex. miR156 and miR157 bind to the SPL2, SPL9, SPL10, SPL11, and SPL15 transcripts with only one mismatched nucleotide, although the position of this nucleotide is different for the two miRNAs (Fig 6A). In addition to two internal nucleotides, miR157 differs from miR156 in possessing an additional U at its 5' end. This 5’ U is unpaired in the miR157-SPL13 duplex (Table 1). To determine if this extra nucleotide might influence the activity of miR157 we compared the relative strengths of miR156a and miR156d. The miR156d transcript is identical to the miR156a transcript, except for the presence of an additional 5’U (Fig 6A). The phenotypes of 5 transgenic lines constitutively expressing a genomic fragment containing MIR156A under the regulation of the CaMV 35S promoter, and an equal number of lines containing a similar construct encoding MIR156D [3], were compared under LD conditions. The lines used for this analysis were selected because they possessed a single T-DNA insertion site. The 35S::MIR156A lines produced approximately twice as many leaves without abaxial trichomes and approximately twice as many cauline leaves as the lines transformed with 35S::MIR156D (Fig 6B). This result demonstrates that MIR156D is less effective than MIR156A, and suggests that the additional 5' U in miR157 is partly responsible for its lower biological activity.
To determine the molecular basis for the effect of mir156 and mir157 mutations on leaf morphology, we compared SPL transcript levels in LP1&2 and LP3&4 in wild-type and mir156/mir157 mutant plants (Fig 7). Consistent with their modest effect on leaf morphology, single mir156c and mir157c mutations had a very small effect on SPL transcripts. However, plants with multiple mir156 and/or mir157 mutations displayed a significant increase in the level of some SPL transcripts. SPL3 transcripts were particularly responsive to a decrease in the level of miR156, increasing about 4-fold in mir156c and 5-to-6-fold in mir156a/c. In contrast, SPL3 transcripts were relatively insensitive to a decrease in miR157, except in genotypes that were also deficient for miR156. For example, SPL3 was elevated nearly 20 fold in LP3&4 of the mir156a/c/d mir157a/c pentuple mutant. SPL9 and SPL15 transcripts increased very slightly in mir156a/c and mir157a/c but increased up to 6-fold in mir156a/c mir157a/c and mir156a/c/d mir157a/c. SPL2, SPL10 and SPL11 increased 2-fold or less in mir156a/c and mir157a/c, and only about 3-fold in mir156a/c mir157a/c and mir156a/c/d mir157a/c. SPL13 transcripts were unaffected in mir157a/c, were elevated about 2-fold in both mir156a/c and mi156a/c mir157a/c, and were only slightly more abundant than this in mir156a/c/d mir157a/c.
The abundance of SPL3 is regulated directly by miR156/miR157 via miRNA-induced transcript cleavage, and indirectly by the effect of miR156/miR157-regulated SPL proteins on the expression of miR172, which in turn represses a group of AP2-like genes that repress the transcription of SPL3, SPL4, and SPL5 [3,9,23,35]. This combination of direct post-transcriptional regulation by miR156/miR157 and indirect transcriptional regulation via the miR172-AP2 pathway may be responsible for the hypersensitivity of SPL3 transcripts to variation in the abundance of miR156/miR157. In contrast, the only way in which miR156/miR157 have been found to regulate the expression of other SPL genes is through a direct interaction with their transcripts. These findings therefore suggest that the SPL2, SPL9, SPL10, SPL11, SPL13 and SPL15 transcripts are differentially sensitive to destabilization by miR156 and miR157.
Both the transcripts and the protein products of miR156/miR157-regulated SPL genes increase during shoot development [3,18,21,36,37]. The expression patterns of miR156/miR157-resistant reporter genes suggest that this increase is largely mediated by miR156/miR157 [21], but whether miR156/miR157 are entirely responsible for the temporal expression pattern of SPL genes is still unknown. To answer this question, we measured the abundance of the SPL2, SPL3, SPL9, SPL10, SPL11, SPL13 and SPL15 transcripts in successive leaf primordia of the mir156a/c mir157a/c mutant (Fig 8). Most of these transcripts were present at either the same level or at slightly lower levels in adult LP (LP5,6,9,10) compared to juvenile LP (LP1,2,3,4). The only exception was SPL3, which increased 3–4 fold from LP1&2 to LP9&10. This result suggests that miR156/miR157 are entirely responsible for the temporal increase in the SPL2, SPL9. SPL10, SPL11, SPL13 and SPL15 transcripts, whereas the temporal increase in SPL3 transcripts may be partly regulated by factors that operate independently of miR156/miR157. Alternatively, the temporal increase in SPL3 may be attributable to the small amount of miR156/miR157 remaining in this quadruple mutant.
The degree to which the abundance of different SPL transcripts changes in response to changes in the level of miR156/miR157 does not necessarily reflect the developmental importance of these SPL genes because miR156/miR157 can mediate both transcript cleavage [3,38–40] and translational repression [23–26]. For example, SPL3 is highly expressed in vegetative shoots and is more sensitive to miR156/miR157 than any other SPL gene, but the phenotype of spl3 mutations demonstrate that it plays little or no role in vegetative development [21]. We were particularly interested in determining whether SPL9 and SPL13 contribute to the precocious phenotype of mir156/mir157 mutants because SPL9 transcripts increase as miR156/miR157 levels decline, whereas SPL13 transcripts are relatively insensitive to changes in these miRNAs (Fig 7). To address this question, we introduced spl9 into a mir156a/c mutant background and introduced spl13 into a mir156a/c mir157a/c mutant background. spl9 completely suppressed the precocious abaxial trichome phenotype and partially suppressed the leaf shape phenotype of mir156a/c, whereas spl13 partially suppressed the effect of mir156a/c mir157a/c on both of these traits (Fig 9). Thus, SPL9 and SPL13 both play important roles in miR156-mediated developmental transitions.
We studied the mechanism by which miR156/miR157 regulate the expression of SPL9 and SPL13 by comparing the abundance of the SPL9 and SPL13 mRNAs with the abundance of their protein products. Antibodies against SPL9 and SPL13 are not available, so we used previously described [21] and newly generated SPL9-GUS and SPL13-GUS translational reporters to visualize these proteins in transgenic plants. Leaf primordia were harvested sequentially as the shoot developed, and the abundance of the SPL9-GUS, SPL13-GUS and miR156 transcripts was measured by RT-qPCR, while the abundance of the SPL9-GUS and SPL13-GUS proteins was measured using the MUG assay. Consistent with previous results [21], a nearly10-fold decrease in the level of miR156 between LP1&2 and LP9&10 was accompanied by very modest (2-fold or less) increase in the level of the SPL9-GUS and SPL13-GUS transcripts (Fig 10A and 10B). In contrast, the activity of the SPL9-GUS protein increased 10-fold between LP1&2 and LP9&10 (Fig 10A), whereas the activity of SPL13-GUS protein increased 15-fold between LP1&2 and LP7&8 (Fig 10B). The relationship between the change in miR156 levels and the change in SPL9-GUS and SPL13-GUS expression varied from leaf to leaf. The 4-fold decrease in miR156 between LP1&2 and LP3&4 was associated with a 3-fold increase in SPL9-GUS activity and a 9-fold increase in SPL13-GUS activity, but subsequent smaller changes in miR156 were associated with disproportionately large increases in the expression of these reporters. For example, in the SPL9-GUS line, miR156 declined by about 2-fold between P3&4 and LP9&10, while the amount of SPL9-GUS protein increased 9-fold. In the SPL13-GUS line, miR156 declined by only 10% between LP3&4 and LP7&8, while the amount of SPL13-GUS protein doubled.
To examine the quantitative relationship between miR156 and SPL13 expression in more detail, we took advantage of a transgenic line containing an estrogen-inducible miR156 target-site mimic (Ind-MIM156), which enabled us to decrease the activity of miR156 by exogenous application of β-estradiol. One-week-old plants homozygous for the SPL13-GUS and In-MIM156 transgenes were given mock and β-estradiol treatments, and LP1&2 were harvested 24 hours later and analyzed by RT-qPCR and the MUG assay. This treatment reduced the abundance of miR156 by about 3-fold and produced a 2-fold increase in SPL13-GUS mRNA, but increased the abundance of the SPL13-GUS protein by greater than 15-fold (Fig 10C). Because the amount of active miR156 in In-MIM156 may not be measured accurately by RT-qPCR, we also examined the abundance of SPL3 transcripts in mock- and estradiol-treated plants. The abundance of SPL3 mRNA is hypersensitive to variation in miR156 and thus serves as a proxy for the abundance of miR156 (Fig 7). SPL3 transcripts were 4-fold more abundant in induced plants relative to mock-treated plants (Fig 10C), which is similar to difference in the amount of SPL3 transcripts in Col vs. mir156c (Fig 7). mir156c reduces miR156 by about 50% (Fig 1B). Consequently, this result implies that estradiol-treated plants had approximately 50% less active miR156 than mock-treated plants, which is consistent with amount of miR156 detected by RT-qPCR. In summary, these results provide further evidence that miR156/miR157 regulate the expression of SPL13 primarily by promoting its translational repression, and also demonstrate that SPL13 activity responds non-linearly to changes in the abundance of these miRNAs.
SPL9 transcripts are more sensitive to changes in miR156/miR157 than SPL13 transcripts (Fig 7), suggesting that transcript cleavage may play a larger role in the regulation of SPL9 than SPL13. To address this possibility, we introduced the miR156-sensitive SPL9-GUS reporter into mir156a/c, mir157a/c, and mir156/c mir157a/c mutant backgrounds, and measured the abundance of the SPL9-GUS mRNA and protein in LP1&2. mir157a/c did not have a significant effect on SPL9-GUS mRNA or protein levels, but mir156a/c produced a 2-fold increase in the SPL9-GUS transcript and a 5-fold increase in the SPL9-GUS protein (Fig 10D). mir156a/c mir157a/c had an even more dramatic effect on the expression of SPL9-GUS, producing a 4-fold increase in the SPL9-GUS transcript and an ~36 fold increase in the SPL9-GUS protein (Fig 10D). These results demonstrate that miR156/miR157 repress SPL9 both by destabilizing the SPL9 transcript and by repressing its translation. The increase in SPL9 activity that occurs during shoot development [21] is probably attributable primarily to a reduction in miR156/miR157-mediated translational repression because the SPL9-GUS protein increases more significantly in response to a decrease in miR156/miR157 than the SPL9-GUS transcript.
To compare the sensitivity of the SPL9 and SPL13 transcripts to miR156/miR157-mediated cleavage, we used a modified form of 5’ RNA Ligase Mediated Rapid Amplification of cDNA Ends (5’ RLM-RACE) [3,41] to quantify the ratio of un-cleaved/cleaved SPL9 and SPL13 transcripts in wild-type Col and mutants deficient for miR156 and miR157. Equal amounts of total RNA from LP1&2 were ligated to a 5’-end RNA adaptor, and the purified RNA ligation products were then used in RT reactions using a poly-T primer. The levels of un-cleaved and cleaved SPL transcripts were then measured by qPCR, using primers specific for each type of transcript. These results were normalized to elf4A1, and the un-cleaved/cleaved transcript ratio in each genotype was then calculated by dividing the relative expression values. This ratio does not necessarily reflect the actual difference between these transcripts because primers for un-cleaved and cleaved transcripts may have different amplification efficiencies. Consequently, instead of using this ratio to compare the relative abundance of cleaved SPL9 and SPL13 transcripts, we asked whether the cleavage of these transcripts is differentially sensitive to variation in the level of miR156/miR157. This was done by normalizing the un-cleaved/cleaved transcript ratio from different mutants to the value in Col. The ratio of un-cleaved:cleaved SPL13 transcripts was about 2-fold greater in mir156a/c and mir156a/c mir157a/c than in Col, whereas the ratio of un-cleaved:cleaved SPL9 transcripts was 5-fold greater in mir156a/c and 15-fold greater in mir156a/c mir157a/c than in Col (Fig 10E). Thus, SPL9 is more sensitive than SPL13 to miR156/miR157-directed transcript cleavage.
The stoichiometry of a miRNA and its target can influence the mechanism of gene silencing [42]. To determine if the mode of action of miR156 is related to the relative abundance of miR156 and its targets, we measured the absolute quantity of several SPL transcripts and miR156 in LP3&4—the leaves in which the translational reporters for SPL3, SPL9, and SPL13 are first expressed [21]. This was done using known concentrations of SPL transcripts and miR156 as standards, and performing RT-qPCR on these standards in parallel with RNA from LP3&4. There was a 5-fold range in the abundance of different SPL transcripts, with SPL5 and SPL15 being the least abundant, and SPL3 and SPL13 being the most abundant (Fig 11). miR156 was 100 times more abundant than SPL3 and SPL13, about 200 times more abundant than SPL6 and SPL9, and about 500 times more abundant than SPL5 and SPL15 (Fig 11). This result therefore suggests that greater than a 100-to-200-fold excess of miR156 is required to completely repress SPL genes. Assuming that the transcription rate of these SPL genes is the same in LP1&2 and LP3&4, we predict that the amount of miR156 in LP1&2 (where all miR156-regulated genes are completely repressed [21]) is approximately 300–600 times greater than the amount of SPL3 and SPL13 transcripts, and approximately 1,500 times greater than the amount of SPL5 and SPL15 transcripts. Although the relative abundance of miR156 vs. SPL9 and SPL13 might suggest that translational repression is favored by a relatively low miR156:SPL transcript ratio (SPL13) whereas transcriptional cleavage is favored by a high miR156:SPL transcript ratio (SPL9), this seems unlikely because a 90% reduction in the level of miR156 in mir156a/c produced only a slight increase in the level most SPL transcripts, including SPL9 (Fig 6). Indeed, we only observed a major increase in SPL transcripts in the mir156a/c mir157a/c quadruple mutant, implying that transcript cleavage does not require high levels of these miRNAs. Thus, the miR156/SPL transcript ratio cannot explain the difference in the sensitivity of the SPL9 and SPL13 transcripts to miR156-directed translational repression.
The phenotype of plants over-expressing miR156 or miR157 reveals that these miRNAs promote the juvenile vegetative phase, but how they specify heteroblastic patterns of leaf morphology, as well as the individual functions MIR156 and MIR157 genes, are unknown. We addressed these questions by accurately measuring the abundance of miR156 and miR157 in leaves with different developmental identities, and by characterizing the phenotype of plants deficient for different MIR156 and MIR157 genes. We found that miR156 and miR157 have similar—but not identical—expression patterns, and that miR157 is more abundant, but less effective, than miR156. We also found that heteroblastic variation in leaf morphology is correlated with the relative abundance of miR156/miR157, and that different features of leaf morphology are differentially sensitive to the level of these miRNAs. Our observation that variation in the abundance of miR156/miR157 produces non-linear changes in the protein level of their targets suggests a molecular mechanism for the qualitative and quantitative changes in leaf morphology that occur during shoot development.
The vegetative period of shoot development is typically divided into two phases—a juvenile phase and an adult phase. However, many species display considerable morphological variation during these phases. In some species the first few leaves are referred to as "seedling leaves" because they are anatomically or morphologically distinct from other juvenile leaves [43–47]. In Arabidopsis, leaves 1&2 differ from other juvenile leaves in that they are smaller and rounder, have a less complex vascular system, and are less sensitive to exogenously applied gibberellin than other juvenile leaves [1,2,6]. Leaves 1&2 also have much lower levels of SPL proteins than other juvenile leaves [21]. We found that leaves 1&2 have a significantly more miR156/miR157 than other juvenile leaves, and that the largest absolute as well as relative decrease in these miRNAs occurs between leaves 1&2 and leaves 3&4. Additionally, the amount of miR156/miR157 in leaves 1&2 far exceeds the amount that is actually required to determine their identity; mutations that nearly completely eliminate miR156 have a similar effect on leaves 3 and 5, but a much weaker effect on leaf 1. The only genotypes that caused leaf 1 to resemble leaves 3 and 5 were those that reduce miR156/miR157 by 90%. These results suggest that leaves 1&2 represent a distinct developmental phase.
Starting with leaf 3, leaf size, the number of leaf serrations, and the angle of the leaf base change gradually from leaf-to-leaf. These gradual changes in leaf morphology are accompanied by an increased ability to produce abaxial trichomes [1], which first appear between leaf 6 to 9. In contrast to the morphological stability of leaves 1&2, the morphology of these “late” juvenile leaves is influenced by light intensity, photoperiod, and the reproductive state of the shoot [2, 5]. This combination of quantitative and qualitative changes, as well as the morphological plasticity of late juvenile leaves, can be explained by the relatively low and gradually decreasing level of miR156/miR157 in successive leaves, and by the non-linear response of some SPL genes to changes in the abundance these miRNAs. Features of leaf morphology that change continuously from leaf-to-leaf are likely to be controlled by pathways or processes whose activity is directly correlated with the level of SPL gene expression, whereas all-or-none traits, such trichome initiation, may only appear when the expression of these genes exceeds a threshold.
Our results also explain why the expression of phase-specific traits becomes dissociated under certain conditions. For example, abaxial trichome production is more responsive to conditions that promote floral induction than either hydathode number or leaf shape [2]. Juvenile and adult vegetative traits can also be dissociated in English ivy, resulting in plants that display different combinations of these traits [48,49]. We suspect that this phenomenon is attributable to functional differentiation between SPL genes, coupled with variation in their sensitivity to miR156 and miR157. Some SPL genes are expressed at relatively high levels and respond nearly linearly to changes in the level of miR156/miR157, whereas others are expressed at relatively low levels and only respond significantly to a change in the level of miR156/miR157 when these miRNAs are present at very low levels. Given that SPL genes are not functionally identical [9,21,50], conditions that produce small changes in miR156/miR157, or which elevate the transcription of particular SPL genes above the threshold established by miR156/miR157, could lead to unusual combinations of phase-specific traits.
miR156 is one of the oldest and most highly conserved miRNAs in plants [51,52]. Although miR157 is nearly as old and as highly conserved as miR156, this is not widely appreciated because miR157 is frequently annotated as miR156 in small RNA sequencing studies and in miRBase (http://www.mirbase.org). The failure to distinguish these miRNAs is likely based on the assumption that they have the same function. Our comparison of the expression patterns of miR156 and miR157, as well the phenotype of plants lacking one or both of these miRNAs, demonstrates that these miRNAs work together to regulate vegetative phase change, but are not functionally identical. miR157 is more abundant than miR156 and is expressed in a similar temporal pattern, but miR156 plays a more important role in vegetative phase change and is a more potent repressor of SPL gene expression. The difference in the activity of these miRNAs may be due, in part, to the lower efficiency with which miR157 is loaded onto AGO1. However, this is not the only reason for the difference in their activity because the amount of miR157 associated with AGO1 is not dramatically lower than the amount of miR156 associated with AGO1. Another factor that may contribute to the difference in their activity is the structure of the miRNA:target-site duplex [41,53–55]. miR156 and miR157 bind to most of their targets with a single mismatch, but this mismatch is located one nucleotide from the cleavage site in the case of miR157 and 3 nucleotides from the cleavage site in the case of miR156. miR157 also has an additional 5' nucleotide (relative to miR156), which is unpaired in the miR157:SPL13 duplex. SPL13 plays a major role in vegetative phase change [21] and if this mismatch reduces the ability of miR157 to repress the activity of SPL13, this would be expected to have a significant phenotypic effect. We suspect that this extra 5' uracil is primarily responsible for the relatively low activity of miR157 because miR156d also has an extra 5' uracil and is significantly less active than miR156, despite being otherwise identical to miR156. However, we cannot rule out the possibility that the difference in the activity of miR156 and miR157 is a consequence of the difference in their length, rather than the specific features of the miRNA:SPL duplex. Most miRNAs in plants are 21 or 22 nt in length, but several evolutionarily conserved miRNAs are 20 nt, or exist as both 20 nt and 21 nt variants [52,56]. miRNAs that are 22 nt are uniquely capable of generating tasiRNAs and other types of phased siRNAs [57,58], but it is still unknown if 20 and 21 nt miRNAs are functionally distinct. The 20 nt miR156 transcript is present in the moss, Physcomitrella patens [59,60], and in virtually all other plants that have been examined to date [51,52]. miR157 is absent in Physcomitrella, but is present in Selaginella and most, but not all, higher plants [61]. The fact that miR157 has been conserved along with miR156 during plant evolution suggests that it is not completely redundant with miR156, and raises the possibility that the relative activity of these miRNAs may differ in different species.
The effect of mutations in different MIR156 and MIR157 genes on the abundance of miR156 and miR157 demonstrates that MIR156A and MIR156C produce most of the miR156 in the shoot whereas MIR157C produces most of the miR157. MIR156D and MIR157A are expressed at a much lower level than these loci, but the ability of mir156d and mir157a to enhance the phenotype of plants mutant for mir156a, mir156c, and mir157c demonstrates that they are functionally significant. We do not know if the miR156 and miR157 transcripts that remain in the mir156a/c/d mir156a/c mutant are derived from one or more these loci or from other MIR156/MIR157 genes because we cannot be certain that the mutations present in this mutant stock are completely null. Whatever the case, these remaining transcripts are functionally active because the phenotype of this pentuple mutant is not as strong as the phenotype of plants over-expressing a miR156 target site mimic [62]. A complete picture of the function of this gene family will require identifying loss-of-function mutations in MIR156E, F, G and H, and MIR157B and D.
The phenotype of slicer-defective AGO1 mutants suggests than plant miRNAs destabilize transcripts exclusively by transcript cleavage, and that this is the primary mode by which they regulate gene expression [63,64]. However, the results of this and previous studies [23–26] indicate that miR156 and several other plant miRNAs act primarily by promoting translational repression [24–27,65,66]. This observation begs the question of why translational repression is so important for the function of these miRNAs, and how the choice between transcript cleavage and translational repression is regulated. Our results suggest that the amount of miR156 and miR157 present in both juvenile and adult leaves is sufficient to almost completely saturate the cleavage machinery at most of their targets. However, the response of different SPL transcripts to miR156/miR157 varies between transcripts, suggesting that the susceptibility of these transcripts to miR156/miR157-induced cleavage depends on sequences outside the miR156/miR157 target site. In plants, the importance of sequences flanking a miRNA target site has been demonstrated for miR159 [42,67] and for several miRNA-cleaved transcripts that generate phasiRNAs [68]. However, it is unclear if the sequence environment of a miRNA target site influences the mechanism by which a miRNA represses gene expression. A comparison of the molecular mechanism by which miR156/miR157 regulate the expression of SPL9 and SPL13 will be informative because these genes respond very differently to changes in the level of these miRNAs. miR156/miR157 are among the oldest miRNAs in plants, and it is therefore reasonable to conclude that miRNA-induced translational repression is an ancient regulatory mechanism in plants. Identifying the biochemical factors that induce AGO1 to direct transcript cleavage vs. translational repression, and defining the functional consequences of these modes of regulation, are important problems for future research.
All of the lines used in this study were in a Col genetic background. The mir156a-2 and mir156c-1 mutations have been described previously [69]. mir156d-1 (SALK_40772), mir157a-1 (Flag_375C03), mir157c-1 (SALK_039809) were obtained from the Arabidopsis Biological Resource Center (Ohio State University, Columbus, OH) and were crossed to Col at least 3 times before further analysis. mir156b-1 was generated by TALEN-directed mutagenesis [70] in a mir156c-1 background, and is a 42 nt-deletion within the MIR156B hairpin sequence (AACAGAGAAAACTGACAGAA—-42 bp deletion—GCGTGTGCGTGCTCACCTCTC) that removes most of the miR156 sequence. Multiple mutant lines were generated by inter-crossing mutations and then screening F2 populations for the desired genotypes using the allele-specific primers listed in S1 Table. Seeds were sown on Farfard #2 Mix and placed at 4° C for 3 days before moving to a Conviron growth chamber, where they were grown under either long day (16 hrs light/8 hrs dark; 80 μmol m-2 s-1) or short day (10 hrs light/ 14 hrs dark; 130 μmol m-2 s-1) conditions, with illumination provided by a 6:2 ratio of broad spectrum (Interlectric Tru-lite) and red light-enriched (Interelectric Gro-lite) fluorescent lights.
The miR156-sensitive and miR156-resistant SPL13-GUS reporter lines used in this study were described previously [21]. The previously described SPL9-GUS reporter lines [21] silenced when they were crossed into a miR156a-2 background, so it was necessary to produce new lines for these reporters. For this purpose, miR156-sensitive and miR156-resistant SPL9:SPL9-GUS genomic sequences [21] were inserted into the pCAM-NAP:eGPF vector [71] using the restriction enzymes XmaI and SbfI. These constructs were then introduced into the miR156a-2/miR156c-1 miR157a-1/miR157c-1 lines by Agrobacterium-mediated transformation. Homozygous single insertion lines were selected as described previously [71], and crossed to Col and further genotyped to obtain SPL9-GUS reporters in different genetic backgrounds.
The estradiol-inducible MIM156 line (Ind-MIM156) was constructed using a Gateway compatible version of the XVE system, as described by Brand and colleagues [72]. The MIM156 sequence described by Franco-Zorilla and colleagues [53] was cloned into pMDC160 by standard Gateway cloning using the primers in S1 Table (referred to as pMDC160-MIM156). Plants containing pMDC150-35S [72] were crossed to transgenic pMDC160-MIM156 plants and made homozygous. Induction of gene expression was performed by spraying 10μM 17-ß-estradiol (0.01% Silwet 77) on seedlings at the desired time point. Tissues were harvested at 24hr after induction.
Tissue samples were harvested into 2ml tubes submerged in liquid nitrogen, and then homogenized using a bead-beater. 300μl of extraction buffer (10 mM EDTA pH 8.0, 0.1% SDS, 50 mM sodium phosphate pH 7.0, 0.1% Triton X-100; 10 mM ß-mercaptoethanol and 25 μg/ml PMSF added fresh before experiment) was then added to each tube. Samples were mixed well and incubated on ice for 10 mins, after which they were centrifuged at 4°C (13000 rpm) for 15 mins. to remove cell debris. 96ul of supernatant was removed and incubated with 4ul of 25mM 4-MUG at 37°C. Incubation time varied among reporters to ensure the end fluorescence readings fell within a linear range. The reaction was terminated by adding 100ul of 1M sodium carbonate to each tube, and fluorescence was measured using a Modulus fluorometer (E6072 filter kit). The amount of MU in each sample was then calculated by comparing this reading to a standard curve constructed by plotting the fluorescence readings from serial dilutions (100nM, 250nM, 500nM, 1000n) of 4-MU. The 4-MU equivalent was divided by the incubation time and this value was then normalized to the amount of protein in the sample, which was determined by performing a Bradford assay on the supernatant remaining in the original tube. For each sample, GUS activity was expressed as 4-MU equivalent/min/mg protein. Values were then normalized to the control sample of each experiment.
RNA was extracted from leaf primordia no larger than 1mm in length using Trizol (Invitrogen), and samples were then treated with DNase (Ambion) following the manufacturer’s instructions. To measure the abundance of miRNAs, 600ng of RNA was used in a reverse transcription reaction with a SnoR101 reverse primer and a miRNA-specific RT primer. To measure the abundance of SPL transcripts, 600ng RNA was used in a reverse transcription reaction primed with Oligo(dT). qPCR was performed on the resulting products, using the primers listed in S1 Table. Reactions were performed in triplicate for each biological replicate.
Tissue was homogenized in liquid nitrogen, and total RNA was then extracted using Trizol (Invitrogen). Extracts were incubated in 500mM NaCl and 5% PEG8000 on ice for 2 hours, and centrifuged at 13,000 rpm for 10min. The supernatant was incubated with a 10% volume of 3M NaOAc and 2 volumes of 100% ethanol at -20°C for 2 hours. Small RNAs were precipitated by centrifugation at 13,000 rpm for 10min, and washed in cold 75% ethanol twice. RNA blotting was performed as described previously [3]. A 1:1 ratio of miR156 and miR157 probes was used for mixed probe hybridizations.
Sequencing libraries were generated from small RNAs isolated from shoot apices of FRI FLC and FRI flc-3 seedlings grown in the conditions described by Willmann and colleagues [2]. The shoot apex samples consisted of the shoot apical meristem and leaf primordia 1 mm or less in length. Libraries were generated using a lab-assembled version of Illumina's 2007 small RNA library sample preparation protocol, followed by high-throughput sequencing with Illumina's Genome Analyzer II platform.
The miR156 and miR157 transcripts used as references were synthesized by IDT, and the SPL transcripts used as references were synthesized by in vitro transcription. The template for each in vitro transcription reaction was generated by PCR, using the primers listed in S1 Table and cDNA from Col. Each purified SPL transcript was assayed by denaturing gel electrophoresis to confirm that the in vitro transcription product was a single species of the expected size. To quantify SPL transcripts, the reference mRNA generated by in vitro transcription was diluted to 1.00E-8 M and this sample was then used to create a 10x dilution series in 600ng/μl total RNA from E. coli. This dilution series was analyzed by RT-qPCR in parallel with RNA isolated from LP3&4. A series of 2x dilutions of the reference mRNA sample whose concentration was similar to that of the experimental sample was then constructed, and run along with the experimental sample in a second RT-qPCR reaction. The 2-CT values of the reference samples were plotted against their known concentrations, and the CT value of the unknown sample was then placed on this graph to determine the RNA concentration.
Ligation reactions were performed with 5 μg of total plant RNA and 1 μg of the GeneRacer (Invitrogen) RNA adapter following the manufacturer’s instructions, but without carrying out the de-capping reaction. After 2 hrs of incubation at 37°C, the reaction mixture was diluted with nuclease- free water and RNA was extracted in phenol: chloroform. The purified ligation product was dissolved in 10μl nuclease-free water, and 5μl of this solution was used in a reverse transcription reaction with an oligo(dT) primer. qPCR was performed using primers listed in S1 Table to quantify cleaved and un-cleaved SPL transcripts.
Two-week-old seedlings were harvested in liquid nitrogen and homogenized in a cold motar and pestle. For each sample, approximately 1mL ground powder was dissolved in 2mL lysis buffer (50mM Tris HCl, pH 7.4, with 150mM NaCl, 1mM EDTA, 1% Triton X-100, 1mM PMSF, 1% Protease Inhibitor) followed by 15 min incubation on ice. 20% of the homogenized sample was saved for RNA extraction, and the rest was centrifuged at 13,000 rpm at 4°C for 20min to remove cell debris. The resulting supernatant was then filtered through a 45μm filter. Immunoprecipitation was performed using Anti-FLAG M2 Magnetic Beads (Sigma) following the manufacturer’s instructions. RNA was extracted from the beads using Trizol (Invitrogen) and analyzed by Northern blotting.
Small RNA sequence data are available in the NCBI Gene Expression Omnibus database under series accession number GSE72303.
|
10.1371/journal.pgen.1001138 | Critical Functions of Rpa3/Ssb3 in S-Phase DNA Damage Responses in Fission Yeast | Replication Protein A (RPA) is a heterotrimeric, single-stranded DNA (ssDNA)–binding complex required for DNA replication and repair, homologous recombination, DNA damage checkpoint signaling, and telomere maintenance. Whilst the larger RPA subunits, Rpa1 and Rpa2, have essential interactions with ssDNA, the molecular functions of the smallest subunit Rpa3 are unknown. Here, we investigate the Rpa3 ortholog Ssb3 in Schizosaccharomyces pombe and find that it is dispensable for cell viability, checkpoint signaling, RPA foci formation, and meiosis. However, increased spontaneous Rad11Rpa1 and Rad22Rad52 nuclear foci in ssb3Δ cells indicate genome maintenance defects. Moreover, Ssb3 is required for resistance to genotoxins that disrupt DNA replication. Genetic interaction studies indicate that Ssb3 has a close functional relationship with the Mms1-Mms22 protein complex, which is required for survival after DNA damage in S-phase, and with the mitotic functions of Mus81-Eme1 Holliday junction resolvase that is required for recovery from replication fork collapse. From these studies we propose that Ssb3 plays a critical role in mediating RPA functions that are required for repair or tolerance of DNA lesions in S-phase. Rpa3 orthologs in humans and other species may have a similar function.
| Proteins that bind single-stranded DNA (ssDNA) are essential for DNA replication, most types of DNA repair including homologous recombination, DNA damage signaling, and maintenance of telomeres. In eukaryotes, the most ubiquitous and abundant ssDNA binding protein is Replication Protein A (RPA), a 3-subunit protein complex consisting of large (Rpa1), medium (Rpa2), and small (Rpa3) subunits. Rpa1 and Rpa2 directly bind ssDNA, whilst the function of Rpa3 is largely unknown. Here, we discover that in fission yeast a 2-subunit complex of Rpa1 and Rpa2 is sufficient for the essential DNA replication function of RPA and its role in homologous recombination repair of double-strand breaks. Rpa3 is not required for these functions, but it is needed for survival of many types of DNA damage that stall or collapse replication forks. Genetic studies indicate close functional links between the Rpa3-dependent activities of RPA, the repair of collapsed replication forks by Mus81-Eme1 Holliday junction resolvase, and the newly discovered Mms1-Mms22 protein complex that is essential for resistance to genotoxins that disrupt DNA replication.
| Preserving genome integrity in eukaryotic organisms depends on integrated mechanisms of DNA replication, DNA repair and telomere maintenance, which are all overseen by checkpoint control systems. Most genome maintenance proteins target specific types of DNA lesions, but a few have much more generalized functions. Of the latter class, perhaps the best-known example is replication protein A (RPA). Also known as single-stranded DNA-binding protein (SSB) or replication factor A (RFA), RPA consists of Rpa1 (∼70 kDa), Rpa2 (∼36 kDa) and Rpa3 (∼14 kDa), that together comprise the major single-stranded DNA (ssDNA) binding activity in eukaryotic cells [1]–[4]. RPA was originally described as a factor that is essential for replication initiation and elongation of SV40 virus DNA in cell extracts [5]–[7], and has since been shown to be required for nucleotide excision repair (NER) and mismatch repair (MMR) in vitro [8], [9]. It also stimulates the activity of homologous recombination (HR) repair proteins in vitro. Indeed, RPA is thought to be a critical factor in every DNA replication or repair process that involves ssDNA [1], [3].
All the known cellular functions of RPA depend on its ability to bind ssDNA [1], [3], [10], [11]. Although a complete 3-dimensional structure of RPA is lacking, structural and biochemical analyses have provided a detailed picture of its domain organization. In essence, RPA is made of 6 oligosaccharide/oligonucleotide binding (OB)-folds: 4 in Rpa1, and 1 each in Rpa2 and Rpa3. Direct binding to ssDNA is mediated by the OB-folds in Rpa1 and Rpa2. The OB-fold in Rpa3 is thought to mediate protein interactions that are required to stabilize the RPA heterotrimer [12], [13].
Teasing apart the in vivo functions of RPA subunits has been a challenging task because RPA is essential for cell viability. Almost all of this work has been carried out with the budding yeast Saccharomyces cerevisiae, where gene disruption studies established that all 3 subunits are required for cell viability [14]. Most genetic studies have been carried out with Rfa1, the ∼70kDA subunit, in which analyses of temperature sensitive or hypomorphic mutants have uncovered defects in DNA replication, recombination, repair, telomere maintenance, and DNA damage checkpoint signaling [3], [4]. Participation in checkpoint signaling has been traced to an interaction with Mec1-Ddc2 checkpoint kinase, which is orthologous to ATR/ATRIP in mammals and Rad3/Rad26 in the fission yeast Schizosaccharomyces pombe [15], [16]. Studies of Rfa2 mutants in budding yeast demonstrate its importance for DNA replication, recombination, repair, and telomere maintenance, although a checkpoint-signaling defect has yet to be established [1], [3]. In contrast to Rfa1 and Rfa2, very little is known about the function of Rfa3 in vivo, with the analyses limited to an N-terminal truncation mutant and a temperature sensitive allele [17].
We have been using fission yeast to investigate the cellular responses to DNA damage in S-phase. Many of these studies have focused on the effects of camptothecin (CPT), which is the prototype of a class of anticancer drugs that stabilize covalent DNA-topoisomerase I complexes by preventing the religation step of topoisomerase I [18]. When a replication fork encounters the CPT-Topoisomerase I complex, it can break, either through direct collision with the CPT-Topoisomerase I complex, or through formation of positive supercoils that stall the fork and can lead to its collapse [19]. In either case, the resulting DNA damage is a one-ended DSB that is subsequently repaired by a homologous recombination repair pathway that creates a Holliday junction in the process of reestablishing the replication fork [20]. Notably, the Mus81-Eme1 Holliday junction resolvase is essential for CPT resistance but plays no role in survival of DSBs created by ionizing radiation (IR), which are repaired primarily by a synthesis-dependent strand annealing (SDSA) mechanism that does not require resolution of Holliday junctions [21]–[23]. We also recently described the Mms1-Mms22/Mus7 protein complex in fission yeast, which like its counterpart in budding yeast, appears to play a very important but as yet poorly understood role in the survival of genotoxins that interfere with DNA replication [24]–[28].
In an effort to more fully characterize the response of fission yeast to replication-associated DNA damage, we carried out a genome-wide screen to identify genes required for CPT resistance. Using a haploid deletion library, we identified a group of CPT sensitive mutants, amongst the most sensitive were mutants for mms22 or ssb3, the latter of which encodes Rpa3. In this report we characterize Ssb3 and explore its role in recovery from DNA damage in S-phase.
We screened an S. pombe haploid deletion library to identify genes required for CPT resistance. In agreement with another recent study [29], we found that mms22Δ and ssb3Δ mutants were amongst the most CPT-sensitive strains in the library (see below). Identification of mms22Δ was anticipated from other studies [24]–[27]. However, as the three subunits of RPA are essential for cell viability in S. cerevisiae, and at least the large subunit Rad11Ssb1/Rpa1 is essential in S. pombe [30], it was unexpected that an ssb3Δ mutant should be viable in S. pombe.
As some alleles in the Bioneer S. pombe deletion library are incomplete deletions, and errors can arise when screening arrayed mutant libraries, it was important to characterize the structure of the presumptive ssb3::KanMX4 mutant in the library. This analysis revealed that the ssb3::KanMX4 mutant was correctly arrayed in the library but the deletion was incomplete. The Bioneer ssb3::KanMX4 allele can potentially encode a protein having the first 21 amino acids of Ssb3 (S. Cavero and P. Russell, unpublished data). Mindful that the C-terminal 52 amino acids of S. cerevisiae Rfa3 are sufficient for function in vivo [17], we designed a new ssb3::KanMX6 construct that completely eliminates the ssb3+ open reading frame (see Materials and Methods). Haploid cells harboring this allele were viable and sensitive to CPT, confirming that Ssb3 is not required for cell viability in fission yeast but is required for CPT resistance (Figure 1A).
We next used serial dilution assays to assess the sensitivity of ssb3Δ cells to a range of genotoxins (Figure 1A). These studies confirmed that ssb3Δ cells are very sensitive to CPT. These cells were also sensitive to a low dose (0.0025%) of methyl methanesulfonate (MMS), which interferes with replication fork progression by alkylating DNA, and to hydroxyrurea (HU), which inhibits DNA replication by poisoning ribonucleotide reductase (Figure 1A). The ssb3Δ cells were also sensitive to UV, which creates cyclobutane dimers and other lesions that impede replication forks. Interestingly, ssb3Δ cells were only modestly sensitive to ionizing radiation (IR), the primary toxic effects of which are DSBs that are repaired in G2 phase (Figure 1A,B). These data show that Ssb3 is required for resistance to a range of genotoxins, particularly those that interfere with DNA replication.
The weak IR sensitivity of ssb3Δ cells suggested that Ssb3 is largely dispensable for homologous recombination-mediated repair of DSBs in mitotic cells. To confirm and extend these findings we analyzed meiosis, in which homologous recombination repair and crossover resolution of programmed meiotic DSBs is required for proper chromosome segregation and spore viability [31]. Tetrad analysis of an ssb3Δ×ssb3Δ cross yielded 91% spore viability, which was only a slight decrease from the 94% spore viability of a ssb3+×ssb3+ cross (Figure 2). Consistent with this high spore viability, microscopic observation showed that the asci from the ssb3Δ×ssb3Δ cross were indistinguishable from wild type. From these results we conclude that Ssb3 is not required for meiotic DSB repair.
Whilst ssb3Δ cells are viable and have only a slightly reduced growth rate, they are elongated relative to wild type (Figure 3A). As this phenotype typically results from activation of a cell cycle checkpoint, we crossed ssb3Δ into strains lacking the Cds1Rad53/Chk2 DNA replication checkpoint kinase or the Chk1 DNA damage checkpoint kinase [32]. Microscopic analysis revealed that the ssb3Δ elongated cell morphology phenotype required Chk1 but not Cds1 (Figure 3A). Various types of DNA damage activate Chk1, whereas Cds1 is primarily activated in response to stalled replication forks. Therefore, these findings indicate that the elongated morphology of ssb3Δ cells is caused by spontaneous DNA damage activating a DNA damage checkpoint that delays the onset of mitosis.
The suppression of the ssb3Δ elongated cell morphology phenotype by chk1Δ suggested that the DNA damage checkpoint is intact in the absence of Ssb3. To further test this hypothesis, we observed the response of ssb3Δ cells to CPT or HU treatment. Both wild type and ssb3Δ cells underwent cell cycle arrest in response to these treatments, whereas cells lacking Rad3Mec1/ATR failed to arrest division (Figure 3B). Consistent with these findings, we observed that Chk1 became hyper-phosphorylated in response to IR or CPT treatment (Figure 3C), which is indicative of an intact DNA damage checkpoint response [33].
These data showed that Ssb3 is not required for the Rad3-Chk1 branch of the DNA damage checkpoint pathway. To confirm this proposition, we carried out genetic epistasis studies with rad3Δ, chk1Δ and cds1Δ mutations. By crossing rad3Δ and ssb3Δ mutants, we discovered a strong genetic interaction between the two mutations, with the double mutant growing much slower than either single mutant (untreated panel in Figure 3D). Although the chk1Δ ssb3Δ double mutant did not have a strong growth defect in the absence of genotoxins, there were obvious synergistic interactions in the presence of IR, UV, HU, MMS and CPT (Figure 3D). The interactions between cds1Δ and ssb3Δ were more complicated: we observed no interactions in response to IR or UV, an additive effect in HU, and suppression in the presence of MMS or CPT (Figure 3D).
The Chk1-dependent cell elongation in ssb3Δ cells suggested that they suffer spontaneous DNA damage. To explore this idea further, we monitored Rad22-YFP foci in ssb3Δ cells. Rad22 is the fission yeast ortholog of Rad52, which is essential for homologous recombination (HR) repair, and many mutants that have genome maintenance defects have increased numbers of Rad22 foci [34]–[36]. We observed a large increase in cells with Rad22-YFP foci in the ssb3Δ strain (∼34%) compared to wild type (∼6%) (Figure 4). We consistently observed that a large fraction of the Rad22-YFP foci in ssb3Δ cells were brighter than in wild type, indicating more extensive recruitment of Rad22. Approximately 33% of the cells with Rad22-YFP foci were septated or attached, which correlates with S-phase in fission yeast, whilst most of the other cells with Rad22-YFP foci appeared to be in G2 phase (Figure 4). These findings suggest that ssb3Δ cells suffer increased rates of DNA damage in S-phase.
To address whether Ssb3 relocalizes to sites of DNA damage, we created a strain in which genomic ssb3+ was modified to encode Ssb3-GFP. This strain was not noticeably sensitive to CPT, indicating that Ssb3 function is undisturbed by the C-terminal GFP fusion (S. Cavero and P. Russell, unpublished data). By live cell analysis we observed that Ssb3-GFP was exclusively nuclear, with ∼7.5% of the cells in an asynchronous population having a bright nuclear focus. This pattern is typical of HR repair proteins such as Rad22 [36]. Upon treatment with CPT, there was a large increase in the number of cells with one or more Ssb3-GFP foci (Figure 5A). Again, this is typical of HR proteins, indicating that Ssb3 is a subunit of an RPA complex involved in homologous repair of DSBs.
To confirm whether Ssb3-GFP localizes at sites of ongoing DSB repair, we created a strain that co-expresses Ssb3-GFP and Rad22-RFP from their genomic loci. In the absence of genotoxic stress, ∼50% of the Ssb3-GFP foci co-localize with Rad22-RFP foci (Figure 5B–D). However, following CPT treatment, there was nearly complete overlap of the Ssb3-GFP and Rad22-RFP foci. This result indicates that Ssb3-GFP foci represent sites of actual DSBs and support a direct role for this protein in the repair of broken replication forks. It should be noted that Ssb3-GFP has a higher nuclear fluorescence, with brighter foci than Rad22-RFP, especially in the absence of DNA damaging agents. This fluorescence difference may cause an underestimation of Rad22-RFP foci, especially in the absence of genotoxins, and hence an underestimation of colocalization of Ssb3-GFP and Rad22-RFP foci.
To further investigate the role of Ssb3 in RPA function, we performed a genetic cross to express GFP-tagged Rad11RPA1 in ssb3Δ cells. Although the rad11-GFP and ssb3Δ parents grew well at the standard growth temperature of 30°C, tetrad analysis uncovered a strong synthetic lethal interaction between the two alleles at 30°C (Figure 6A). Microscopic analyses performed revealed that the rad11-GFP ssb3Δ spores formed microcolonies of very elongated cells at 30°C (Figure 6A). A similar effect was observed with HA-tagged rad11 (S. Cavero and P. Russell, unpublished data). These data indicated that the tags slightly impair Rad11 function or destabilize the RPA complex, making it highly dependent on Ssb3. In support of this idea, we found that spore germination at 25°C yielded healthier rad11-GFP ssb3Δ cells that were suitable for localization studies. As expected, the growth of these cells was temperature sensitive (S. Cavero and P. Russell, unpublished data).
We carried out live-cell analysis of Rad11-GFP localization, in the absence or presence of CPT, in wild type (wt) or ssb3Δ cells grown at 25°C. About 5% of wild type cells had Rad11-GFP nuclear foci in the absence of DNA damage (Figure 6B). This number increased ∼3.5-fold when cells were incubated with 30µM CPT for 4 hours, indicating that Rad11-containing RPA complex localizes to DSBs, as expected. Interestingly, the percentage of nuclei with Rad11-GFP foci in the absence of DNA damage was greater in ssb3Δ cells, probably resulting from increased spontaneous DNA damage. Similarly, after CPT treatment, the increase in Rad11-GFP foci formation was also greater in ssb3Δ cells compared to wild type (Fig. 6B). Therefore, the relocalization of RPA complex to DSBs does not depend on Ssb3, and instead it appears to be enhanced.
It should be noted that the frequency of Rad11-GFP foci observed in wild type cells in this experiment was less than seen for Ssb3-GFP in Figure 5. This difference most likely arises from the effect of temperature; indeed, the frequency of Ssb3-GFP foci is reduced at 25°C versus 30°C (S. Cavero and P. Russell, unpublished data). Moreover, co-expression of Rad11-GFP and Ssb3-RFP revealed almost 100% overlap of nuclear foci (Figure 6C).
Having found that Ssb3 is required for survival of S-phase DNA damage and that it colocalizes with Rad22 and RPA at CPT-induced DNA damage sites, we lastly performed a genetic epistasis analysis of ssb3Δ with genes involved in different pathways of DNA replication or repair. Double mutant strains were created by mating and assessed for growth in dilution series on rich growth medium in the absence or presence of various genotoxins (Figures 7 and 8). The data are summarized in Table 1. In the absence of genotoxins, we detected negative genetic interactions with the Swi1Tof1-Swi3Csm3 replication fork protection complex (FPC), which is required for stable fork pausing and efficient activation of Cds1 [34], [37]; Ctf18, which is a subunit of an alternative Replication Factor C (RFC) complex that is involved in fork stabilization [38]; Rfc3, a subunit of RFC [39], [40]; Rad2, the 5′ flap endonuclease required for Okazaki fragment processing during DNA replication; Rhp55Rad55, a Rhp51Rad51 paralog that forms a heterodimer with Rhp57Rad57 that mediates the formation and/or stabilization of the Rad51-DNA filament required for HR repair [40]; and Brc1Rtt107, a 6-BRCT domain protein that is involved in survival after DNA damage in S-phase and binds phospho-histone H2A (γH2A) at sites of DNA damage [41], [42]. Under the same conditions we detected only weak or no negative genetic interactions with Rad13 and Swi10, which are required for nucleotide excision repair (NER); Rhp51, which is the Rad51 ortholog required for most types of HR repair; or with Mms22, which is required for recovery after DNA damage in S-phase [24], [25]. In the presence of genotoxins, additional negative genetic interactions with ssb3Δ were revealed for Rad13, Swi10 and Rhp51. Mms22 did not show a genetic interaction with Ssb3 in the presence of genotoxins, with the exception of a weak interaction in cells treated with UV (Figures 7 and 8).
Mms22 forms a protein complex with Mms1, and we have found that mms1Δ mms22Δ double mutants behave identically to either single mutant [28]. From these genetic relationships we predicted that ssb3Δ and mms1Δ should have an epistatic genetic interaction in all conditions except for UV, as was seen for ssb3Δ and mms22Δ (Figure 8). Genetic studies confirmed this prediction (Figure 9A).
Mutations that inactivate Mms1 and Mms22 have negative genetic interactions with deletions of many genome maintenance genes. One exception is Mus81, which together with Eme1 forms a Holliday junction resolvase that is required for recovery from collapsed replication forks and resolution of crossovers in meiosis [21]–[23], [43]. Having found that ssb3Δ has an epistatic relationship with mms1Δ and mms22Δ, we explored the genetic interactions involving ssb3Δ and mus81Δ. In the absence of genotoxins, mus81Δ cells grew slowly compared to wild type, but this defect was not enhanced by eliminating Ssb3. Similarly, in the presence of genotoxins, deleting Ssb3 appeared to only slightly enhance the growth defect of mus81Δ cells, indicating a close functional relationship between Ssb3 and Mus81 (Figure 9B).
Our studies identified a negative genetic interaction between ssb3Δ and rhp55Δ, with the double mutant being similar to the rhp51Δ strain (Figure 8A). Rhp55-Rhp57 protein complex is one of two Rhp51 mediators, with the other consisting of Sfr1 and Swi5, which form a protein complex that serves as an alternative mediator for Rhp51 [43], [44]. Inactivating both mediators severely impairs Rhp51 function. In view of these relationships, we tested the genetic interactions between ssb3Δ and sfr1Δ or swi5Δ. In contrast to the interaction between ssb3Δ and rhp55Δ in the absence of genotoxins (Figure 8A), our data indicated there was little or no interaction between ssb3Δ and sfr1Δ or swi5Δ in the absence of genotoxins (Figure 9C). However, there were obvious negative interactions between ssb3Δ and sfr1Δ or swi5Δ in the presence of MMS or CPT, and weaker interaction in double mutant cells treated with UV or HU (Figure 9C). The possible interpretations of these data are discussed below.
As a central component of all DNA transactions involving ssDNA, RPA has been the subject of many genetic, biochemical and structural studies. The large majority of the in vivo functional studies have been carried out with S. cerevisiae, in which it is clearly established that all 3 subunits of RPA are essential for cell viability. It was reasonable to assume the same was true in all organisms. We were therefore surprised to find that the small subunit of RPA was apparently not essential in fission yeast. To eliminate doubts, we reengineered an ssb3Δ mutation to completely eliminate the ssb3+ open reading frame, and found that this mutant was also viable. From these studies we conclude that a heterodimeric complex consisting of Rad11Rpa1and Ssb2Rpa2 can carry out the essential DNA replication functions of RPA in fission yeast. As functional studies unfold in other model organisms, it will be interesting to determine whether the essentiality of Rpa3 is the rule or the exception.
Whilst our studies demonstrate that Ssb3 is not required for the essential functions of RPA, they nevertheless show that Ssb3 has important effects on RPA function. One observation supporting this conclusion is the temperature sensitive genetic interaction between Rad11-GFP and ssb3Δ. It is likely that both alleles modestly destabilize or impair the function of RPA complex, to the degree that combining the alleles causes an acute temperature sensitive phenotype. This hypothesis is consistent with data indicating that Rpa3 mediates protein interactions that help to stabilize the RPA heterotrimer [12], [13].
In the absence of exogenous DNA damaging agents, ssb3Δ mutants have a modest growth defect, moderate cell elongation dependent on Chk1, and an elevated number of Rad11Rpa1 and Rad22Rad52 foci. These phenotypes most likely result from defects in DNA replication, leading to DNA structures recognized as DNA lesions by the DNA damage checkpoint and HR repair machinery. However, ssb3Δ and chk1Δ mutations do not have an obvious synergistic growth defect. These data suggest that the DNA lesions leading to a Chk1-dependent mitotic delay in ssb3Δ cells are unlikely to be DSBs. In support of this conclusion, we did not detect a obvious genetic interaction between ssb3Δ and rhp51Δ when mutants were tested in the absence of genotoxins. Thus, it is likely that ssb3Δ cells accumulate gapped ssDNA structures that activate the DNA damage checkpoint and are substrates for Rad22.
We found that cds1Δ partially suppresses the MMS and CPT sensitivity of ssb3Δ cells (Figure 3D). This genetic interaction was unexpected. The toxicity of MMS and CPT is thought to derive primarily from replication fork collapse; therefore, these data suggest that activation of the replication checkpoint is detrimental to restoration of collapsed replication forks in ssb3Δ cells. It will be interesting to discover which substrates of Cds1 kinase mediate this effect.
Whilst ssb3Δ cells are sensitive to CPT, MMS and UV, they display only weak sensitivity to IR. The toxic effects of CPT, MMS and UV are thought to arise primarily from creating DNA lesions that interfere with DNA replication, whereas IR directly causes DSBs in all phases of the cell cycle. These data suggest that Ssb3 is not critical for typical HR-mediated DSB repair in mitotic cells, which occurs through a synthesis-dependent strand-annealing (SDSA) pathway requiring Rad22Rad52, Rhp51Rad51, and other core HR proteins. Furthermore, ssb3Δ cells exhibit only a very minor defect in spore viability. Since formation of viable spores depends on carrying out HR repair of programmed meiotic DSBs through a double-strand break repair (DSBR) pathway requiring resolution of Holliday junctions by Mus81-Eme1 resolvase [21], our findings show that both pathways of HR repair are largely functional in the absence of Ssb3. However, these data do not exclude the possibility that deletion of ssb3+ might have modest effects in meiotic recombination.
The HR-mediated repair of CPT-induced broken replication forks also depends on Rad22, Rhp51 and Mus81, and our data indicate that Ssb3 plays a relatively important role in this pathway. It is unclear why Ssb3 should appear to be more important for repair of CPT-induced damage than IR-induced DSBs. Ssb3 might have an important role in stabilizing stalled replication forks, such as those that are proposed to form as the result of positive supercoils forming ahead of the fork in cells treated with CPT [19]. It is noteworthy that previous genetic studies have identified other genome maintenance factors that are not required for repair of IR-induced DSBs but are critical for survival of CPT treatment. The most relevant factors may be Mus81-Eme1 and Mms1-Mms22 protein complexes, which are critical for survival of CPT treatment but are not required for repair of IR-induced DSBs [21]-[23], [25], [28]. In this respect, it is likely to be particularly significant that ssb3Δ displays little or no synergistic genetic interactions with mus81Δ, mms1Δ and mms22Δ mutations in the absence or presence of genotoxins, indicating that there are likely to be close functional connections between the Ssb3-dependent functions of RPA and the activities of Mus81-Eme1 Holliday junction resolvase and Mms1-Mms22 protein complex. This possibility is entirely consistent with studies showing that mms1Δ and mms22Δ have negative genetic interactions with mutations of many genome maintenance genes, but not with mus81Δ [24], [25], [28]. However, it is important to note that unlike Mus81, Ssb3 is not required for resolution of Holliday junctions in meiosis, and thus it is unlikely that Ssb3 has an integral role in the Mus81-dependent resolution of Holliday junctions that form upon restoration of collapsed replication forks [21], [44].
Regardless of the genotoxins assayed in our studies, the phenotypes of ssb3Δ cells do not match that of an rhp51Δ mutant that is severely defective in HR-mediated DSB repair. However, the phenotypes of ssb3Δ and rhp55Δ mutants are similar, and there is a strong genetic interaction when ssb3Δ and rhp55Δ are combined. In fact, the ssb3Δ rhp55Δ double mutant is approximately equivalent to rhp51Δ. These data suggest that there is a nearly complete breakdown of HR in the ssb3Δ rhp55Δ double mutant. Rhp55 is a Rad51 paralog that forms a heterodimer with Rhp57. Studies of the analogous Rad55-Rad57 complex in budding yeast have shown that it can function as a mediator in the strand-exchange reaction necessary for Rad51 to nucleate on ssDNA in the presence of RPA [45]. One possible interpretation of these findings is that both ssb3Δ and rhp55Δ mutations cause defects in Rad51 nucleation on ssDNA, resulting in a synergistic interaction of the mutations. RPA foci are actually enhanced in ssb3Δ cells, suggesting that a key role of Ssb3 may be promoting the disassembly of RPA from ssDNA.
Interestingly, in the absence of exogenous genotoxins, ssb3Δ does not have obvious genetic interactions with mutations deleting genes encoding Sfr1 or Swi5, which form an alternative mediator complex for Rhp51 [46], [47]. These data are consistent with the possibility that Ssb3 might act with Sfr1-Swi5 complex. However, when tested in the presence of genotoxins, there are clear negative genetic interactions involving ssb3Δ and sfr1Δ or swi5Δ. The same is true for ssb3Δ and rhp51Δ mutations. Thus, Ssb3 might act with the Sfr1-Swi5 complex, but it also has functions that do not involve Rhp51 and its mediators.
The strains used in this study are listed in Table S1. Standard procedures and media for S. pombe genetic and biochemical analyses were used as previously described [48]. Cells were cultured in YES (yeast extract, glucose and supplements) or EMM (defined minimal medium with supplements) as described [48]. The complete deletion of the ssb3 open reading frame was made using pFA6a-KanMX6 as template and the primers “Ssb3 KanMX6 forward” (5′-TCG TGT CAA CAA GTA GTT AAC TAC CTG GTC TGA TAC ATA CTT CAC TTC CAC CAC TTT ATA AAC AAC GCG TAT AAA ATA ATC GGA TCC CCG GGT TAA TTA A-3′) and “Ssb3 KanMX6 reverse” (5′-CGT TTA TTC TTC CAT GTT TAT TTG TAC TGT GCA TGA GAA ATG AAA GAG AAA TCT GTG TTG TAT GAT CCA TAA AAT GTT TCG AAT TCG AGC TCG TTT AAA C-3′). The PCR product was transformed into PR110 (h+ leu1-32 ura4-D18) cells by electroporation, G418-resistant colonies were obtained, and PCR and DNA sequencing verified ssb3 disruption.
Cells were photographed using a Nikon Eclipse E800 microscope equipped with a Photometrics Quantix CCD camera and IPlab Spectrum software. All fusion proteins were expressed at their own genomic locus. Rad22-YFP and Ssb3-GFP/Rad22-RFP-expressing strains were cultured in EMM, while Ssb3-GFP and Rad11-GFP-expressing strains were grown in YES until mid-log phase for foci quantification assays. Experiments with Rad11-GFP expressing strains were carried out at 25°C because of the thermosensitivity of the rad11-GFP ssb3Δ strain. In the case of CPT treatment, 30µM CPT in DMSO (or DMSO alone as a control, 0.3% final concentration) was added to mid-log phase cultures for 4 hrs at 30°C, washed out and cells were resuspended in YES for foci quantification. In all cases, at least 800 nuclei were scored in three independent experiments. All microscopy was conducted with live cells. Cell length, nuclei number and position, and the presence of a division plate were used to assess cell cycle position.
In chronic exposures to drugs, mid-log phase cultures were resuspended to 0.5 OD600 and serially diluted tenfold. Dilutions were spotted onto YES agar plates containing the indicated amounts of CPT, MMS or HU. Note that the effective concentration of these drugs, particularly MMS, can vary depending on the age and source of the stock solution, which accounts for the different concentrations used in some of the survival assays. For exposure to IR, cells were irradiated using a 137Cs source with the indicated dose and then serially diluted onto triplicate YES plates. In the case of UV treatment, cells were serially diluted onto YES plates and irradiated using a Stratagene Stratalinker UV source. Cell survival was determined after 3–4 days at 30°C unless otherwise indicated.
Whole cell extracts were prepared from exponentially growing cells disrupted in lysis buffer (50mM Tris-HCl pH 8, 150mM NaCl, 2.5mM EDTA, 10% Glycerol, 0.2% NP-40, 50mM NaF, 5mM PMSF, and Complete Protease Inhibitor tablets (Roche)) with a glass bead beater and resolved in 8% SDS-PAGE. Proteins were transferred to nitrocellulose membranes, blocked with 5% dry milk, 0.05% Tween-20 in TBS (150mM NaCl, 50mM Tris-HCl pH 7.6) and probed with HA (Roche) antibody.
|
10.1371/journal.pgen.1003382 | Genome-Wide Control of RNA Polymerase II Activity by Cohesin | Cohesin is a well-known mediator of sister chromatid cohesion, but it also influences gene expression and development. These non-canonical roles of cohesin are not well understood, but are vital: gene expression and development are altered by modest changes in cohesin function that do not disrupt chromatid cohesion. To clarify cohesin's roles in transcription, we measured how cohesin controls RNA polymerase II (Pol II) activity by genome-wide chromatin immunoprecipitation and precision global run-on sequencing. On average, cohesin-binding genes have more transcriptionally active Pol II and promoter-proximal Pol II pausing than non-binding genes, and are more efficient, producing higher steady state levels of mRNA per transcribing Pol II complex. Cohesin depletion frequently decreases gene body transcription but increases pausing at cohesin-binding genes, indicating that cohesin often facilitates transition of paused Pol II to elongation. In many cases, this likely reflects a role for cohesin in transcriptional enhancer function. Strikingly, more than 95% of predicted extragenic enhancers bind cohesin, and cohesin depletion can reduce their association with Pol II, indicating that cohesin facilitates enhancer-promoter contact. Cohesin depletion decreases the levels of transcriptionally engaged Pol II at the promoters of most genes that don't bind cohesin, suggesting that cohesin controls expression of one or more broadly acting general transcription factors. The multiple transcriptional roles of cohesin revealed by these studies likely underlie the growth and developmental deficits caused by minor changes in cohesin activity.
| The cohesin protein complex binds to chromosomes and helps ensure that chromosomes are divided equally into the daughter cells when a cell divides. Cohesin also affects how genes are expressed. Small changes in cohesin alter gene expression and development, causing Cornelia de Lange syndrome, a genetic disease. Cohesin influences the amount of RNA produced by many genes, but the reasons are poorly understood. We investigated this question by measuring how changes in cohesin levels affect the level of RNA polymerase, the enzyme that transcribes genes to make RNA, at all genes in Drosophila cells. We find that genes that bind cohesin have higher average levels of RNA polymerase and produce more final processed RNA per RNA polymerase than genes that don't bind cohesin. We also find that cohesin binds nearly all DNA sequences located outside of genes that are predicted to regulate gene expression. Reducing cohesin affects RNA polymerase levels at many genes and the predicted regulatory sequences, indicating that cohesin facilitates communication between regulatory sequences and genes. Our data also show that cohesin affects transcription of most genes that don't bind cohesin, likely by controlling transcription of broadly acting transcription factors that regulate many genes.
| Cohesin is a large protein ring that topologically encircles DNA and participates in several chromosome functions, including sister chromatid cohesion, chromosome segregation, DNA repair, and gene expression (reviewed in [1]–[3]). It is loaded onto chromosomes by the kollerin complex, and removed by the releasin complex.
Modest changes in cohesin, kollerin or releasin activity alter gene expression, growth, and animal development without measurable defects in chromatid cohesion or chromosome segregation. For instance, minor alterations of kollerin or cohesin activity in humans cause Cornelia de Lange syndrome (CdLS, OMIM #122470, #300590, #610759, #614701) which is associated with significant physical and intellectual deficits (reviewed in [4]). Cohesin also influences gene expression in non-dividing cells [5], [6]. Thus, cohesin's role in gene expression appears largely independent of its role in cell division, and considerably more sensitive than its other cellular functions to changes in cohesin dosage.
Current evidence argues that cohesin directly influences gene transcription. In animal cells, cohesin and kollerin preferentially bind genes important for growth and development near the transcription start site and in the transcribed region [7]–[11]. In Drosophila, cohesin is largely absent from silent genes, and selectively binds active genes in which transcriptionally-engaged RNA polymerase II (Pol II) pauses just downstream of the start site [9], [12]. Upon depletion of cohesin or kollerin, mRNAs from cohesin-binding genes are more likely to be affected than those from non-binding genes, and can change within a few hours [5], [13]. Current evidence argues that cohesin regulates transcription by multiple mechanisms, including facilitating enhancer-promoter and insulator looping, and by controlling the transition of promoter-proximal paused Pol II to efficient elongation [1], [2].
The prior studies of how cohesin regulates gene expression measured steady state mRNA levels, and thus do not clearly differentiate the roles of cohesin in transcription from other processes such as RNA splicing, transport, and stability. To gain more direct insights into the mechanisms by which cohesin influences transcription, we measured the effects of cohesin depletion on the genome-wide distribution of Pol II, Pol II phosphorylated at the serine 2 residue in the heptad repeats in the C terminal domain of the Rpb1 subunit (Ser2P Pol II), P-TEFb, and Cdk12 in Drosophila cells derived from central nervous system. Ser2P Pol II is actively elongating and formed by the action of the P-TEFb and Cdk12 kinases. We also measured the effects of cohesin and kollerin depletion on transcriptionally-engaged Pol II by precision global run-on sequencing (PRO-seq). We deduce that cohesin directly promotes the transition of promoter-proximal paused Pol II to elongation at many genes that it binds from comparing the changes in Pol II occupancy and activity in control and cohesin-depleted cells. The evidence indicates, that in many cases, cohesin likely facilitates this transition by supporting long-range enhancer-promoter interactions, but also has other roles directly at the promoter. Surprisingly, we also find that cohesin influences Pol II activity at most genes that don't bind cohesin, possibly through control of broadly-acting transcription factors.
To directly assess the influence of cohesin on gene transcription, we compared the genome-wide occupancy of Pol II and Pol II kinases relative to cohesin binding, and measured the effects of cohesin depletion on Pol II and kinase occupancy. We used genome-wide chromatin immunoprecipitation with tiling microarrays (ChIP-chip) to measure the genome-wide binding of Pol II, the Cyclin T (CycT) subunit of the P-TEFb complex, and the Cdk12 Pol II kinase in ML-DmBG3 (BG3) Drosophila cells derived from larval central nervous system. We used antibodies against the Rpb3 subunit of Pol II [14] to measure the total Pol II occupancy, and antibodies specific for Ser2P Pol II to measure elongating Pol II. All ChIP-chip experiments were performed with two independent biological replicates and averaged.
Genome-wide, Rpb3 correlates well with Ser2P Pol II (r = 0.87), especially on gene bodies (Table 1; Figure 1). Pol II positively correlates with the CycT subunit of the P-TEFb Pol II kinase (0.64–0.67), and somewhat less, although significantly, with the Cdk12 kinase (0.39–0.45) (Table 1). The Rad21 cohesin subunit strongly overlaps Pol II (r = 0.67–0.68), consistent with prior findings [9], and has a similar correlation with CycT (0.73), but less with Cdk12 (0.49) (Table 1). We often detect CycT and Cdk12 at promoters, and enrichment in the gene bodies is frequently similar in strength, as in diminutive, the Drosophila myc gene (dm, FlyBase FBgn0262656; Figure 1).
ChIP does not determine if Pol II is transcriptionally engaged, or the direction it is transcribing. We thus used precision global run-on sequencing (PRO-seq; [15]), a variation of GRO-seq [16] that gives improved resolution to measure the levels and orientation of transcription-competent Pol II genome-wide. PRO-seq varies from GRO-seq in that biotin-labeled ribonucleotides are used to allow run-on for a nucleotide or two, instead of the longer run-on with BrUTP used in GRO-seq. PRO-seq, like GRO-seq [17], is highly sensitive, and unlike ChIP, does not depend on crosslinking efficiency or antibody specificity, and detects elongation-competent Pol II regardless of the phosphorylation status. Nuclei were isolated under conditions of ribonucleotide depletion to halt transcription, but leave Pol II transcriptionally engaged. The nascent RNA transcripts produced upon restart of transcription were used to generate a cDNA library for high-throughput sequencing. Inclusion of sarkosyl in the run-on transcription reaction prevents new transcription initiation, so that only Pol II that is already transcriptionally engaged is detected, and gene body and promoter paused Pol II are detected with equal efficiency [17]. Two independent biological replicates were used for each PRO-seq measurement (control, Rad21 RNAi, Nipped-B RNAi).
The number of PRO-seq reads was quantified for nearly 17,000 annotated transcription units, and after normalization for the total number of reads, the genome-wide correlations between the two biological replicates were 0.98 for all three groups (Table S1). We selected approximately 7,000 “PRO-seq active” transcription units for detailed analysis by using only those transcription units that had at least 1 read per million in the 200 bp region surrounding the annotated transcription start site, and in the gene body in the control cells (Table S2). Because genes only bind cohesin when they are active [9], restricting the analysis to active genes is essential for valid comparisons of cohesin-binding to non-binding genes. Many genes have more than one active transcription start site, and thus the 7,000 active transcription units represent approximately 6,000 genes.
Cohesin-binding genes have more Pol II on average than non-binding active genes as measured by both PRO-seq and ChIP-chip. When active genes are subdivided into four groups (Figure S1A) from low to high cohesin binding levels based on the mean Rad21 ChIP signal in the 400 bp region surrounding the transcription start site, the average PRO-seq read density and Rpb3, Ser2P Pol II, Cdk12, and CycT ChIP signals at the promoter all increase with cohesin (Figure 1A–1E). Similar results are obtained for both promoters and gene bodies when PRO-seq active genes are split into cohesin-binding and non-binding genes, and Pol II occupancy is measured by ChIP-chip (Figure S1F).
A prior report indicated that cohesin preferentially binds genes with promoter-proximal paused polymerase, based in part on genome-wide overlap of cohesin with the Negative Elongation Factor (NELF) pausing factor, and the higher levels of short promoter-proximal transcripts produced by cohesin-binding genes [12]. The PRO-seq data, which directly measures pausing, confirms these findings. The pause index is defined as the ratio of the PRO-seq signal density (normalized reads per base pair) in the 200 bp promoter region to the density in the rest of the gene body. The average pause index increases with cohesin occupancy, and the genes with the highest cohesin levels have substantially higher pausing (Figure 1F). Conversely, when active genes are divided into four groups ranging from low to high pausing (Figure S1B), the average cohesin occupancy at the promoter increases with the pause index (Figure S1C). Pausing can also be measured by the ratio of the Rpb3 ChIP signal at the promoter to the signal in the gene body [18], and this analysis also confirms that cohesin-binding genes have higher levels of pausing (Figure S1D). Although Rpb3 ChIP is not as sensitive as PRO-seq, and is not specific for transcriptionally-engaged Pol II, the concordance between the PRO-seq and Rpb3 measures of pausing agrees with the finding that most Pol II at the promoter is transcriptionally-engaged [17].
The Drosophila Nipped-B kollerin subunit was discovered in a genetic screen for factors that control long-range activation of the cut (FlyBase FBgn0004198) and Ultrabithorax (FlyBase FBgn0003944) genes by remote tissue-specific enhancers [19], and cohesin binds and facilitates the activity of transcriptional enhancers for pluripotency, β-globin, and T cell receptor genes in mammalian cells [6], [7], [20]. We thus examined the cohesin and Pol II occupancy of predicted extragenic cis-regulatory modules (CRMs) in BG3 cells. Active CRM/enhancer features include DNAseI hypersensitive sites (DHS), and the H3K4me1 and H3K27ac histone modifications (reviewed in [21]). The modENCODE project generated these data for BG3 cells [22], and by these criteria, there are 2,353 potential CRMs, 557 of which are not within annotated transcription units and are at least 500 bp from a transcription start site (Table S3). Forty-two of the predicted CRMs overlap 21 tissue-specific CRMs curated by the REDfly database that are functional in transgenic reporter constructs [23]. Strikingly, we find that virtually all predicted extragenic CRMs (96%) bind cohesin and Nipped-B (Figure 2A). A similar fraction (94%) of all 2,353 CRMs, which includes those located within transcribed regions, bind cohesin. Cohesin levels at the extragenic CRMs correlate positively with both the H3K27ac (r = 0.65) and H3K4me1 histone modification levels (Figure S2). Somewhat less than half of the extragenic CRMs associate with Pol II, and a similar fraction bind Pol II kinases (Figure 2A). Association of Pol II and Pol II kinases with a large fraction of these extragenic sequences supports the idea that they are functional CRMs, and the finding that virtually all bind cohesin is consistent with the idea that cohesin facilitates their function.
The average cohesin occupancy of the extragenic CRMs is higher than that for all active promoters, while the Pol II occupancy of active promoters is higher than that of the CRMs (Figure 2B). PRO-seq density indicates that much of the Pol II detected by ChIP at the CRMs is not transcriptionally engaged (Figure 2B). While the median Pol II occupancy of the predicted CRMs by ChIP is only some 3-fold lower than for promoters, the median PRO-seq density at the CRMs is indistinguishable from zero, given that less than 50% of CRMs have PRO-seq signals (Figure 2B). As seen in S2 cells [17], the mean signals at CRMs are substantially lower than those at promoters, such that the ratio of the mean PRO-seq to mean Pol II ChIP ratio is approximately 50-fold lower at CRMs than at promoters. We theorize, therefore, that most of the Pol II detected by ChIP at CRMs is promoter-bound Pol II that associates with the CRMs through DNA looping, although we cannot rule out the possibility that Pol II is directly recruited by CRM-bound proteins, but cannot initiate transcription.
Figure 2C shows clustered CRMs some 68 kb upstream of cut, in a region without genes, and which produces no mRNA. The surrounding region contains several enhancers that regulate the cut gene throughout development. The wing margin enhancer whose function is sensitive to Nipped-B dosage in vivo is 12 kb upstream of these putative CRMs, and several other tissue-specific enhancers are downstream [19], [24], [25]. The region with the CRMs contains enhancers critical for differentiation of multiple sensory cells. Gypsy transposon insulator insertions just upstream of the predicted CRMs cause primarily cut wing phenotypes, while insertions just downstream also cause head capsule defects, including deformed antenna [26].
Cohesin (Rad21) depletion substantially reduces the level of elongating Pol II on the cut gene as measured by Ser2P Pol II ChIP (Figure 2C), and the PRO-seq signals decrease some 40% in the gene body (Table S2). Cohesin depletion also modestly reduces the Ser2P Pol II ChIP signal in the region containing the predicted cut CRMs, lending support to the idea that this is a functional remote enhancer. By ChIP-chip, cohesin also influences association of Pol II with many of the other predicted extragenic CRMs around the genome. Figure 2D shows that Rpb3 and Ser2P Pol II occupancy decrease significantly on 15 to 25% of the predicted CRMs upon Rad21 depletion, consistent with the idea cohesin facilitates interactions of many CRMs with promoters.
Stable topological binding of cohesin to chromosomes requires loading by kollerin. Thus, depletion of cohesin and kollerin would be expected to have comparable genome-wide effects on Pol II if topologically-bound cohesin is the form that influences transcription. We compared PRO-seq measurements in mock-treated control BG3 cells to cells in which the Rad21 cohesin subunit or the Nipped-B kollerin subunit were depleted by approximately 80% using RNAi. Under these depletion conditions, there are no measurable defects in sister chromatid cohesion or chromosome segregation, and a modest decrease in the rate of cell division, which may reflect decreased expression of the Drosophila myc (dm) gene that promotes cell growth [13], [27]. These RNAi conditions reduce cohesin chromosome binding by at least 3 to 4-fold at all genes examined by ChIP, including genes that start with very high cohesin and show some of the largest changes in mRNA levels [12].
The effects of Rad21 and Nipped-B depletion on the PRO-seq signals in the promoter regions and gene bodies of the PRO-seq active genes are remarkably similar. The maximal changes included increases and decreases approaching 15-fold at promoters (Figure 3A), and some greater than 16-fold in gene bodies for both cohesin-binding and non-binding genes (Figure 3B). These results indicate that topologically-bound cohesin is the form that influences transcription.
Cohesin and kollerin depletion also had very similar effects on the pause index, which measures the efficiency with which paused Pol II enters into elongation. Upon Rad21 or Nipped-B depletion, genes with high cohesin levels showed increased and decreased pausing at similar frequencies (Figure 3C, 3D). Thus, depletion of cohesin or the loading factor have remarkably similar effects on regulatory steps of transcription.
Overall, cohesin depletion did not substantially change the median pausing index at cohesin-binding genes, with similar numbers of genes showing increases and decreases (Figure 3D, Figure S1E). This is consistent with the prior findings that cohesin increases expression of some genes and decreases expression of others [13]. One possibility is that in addition to facilitating enhancer-promoter interactions, cohesin might also facilitate interactions of silencers that inhibit transition of Pol II to elongation. Prior studies also show that cohesin blocks transition of paused Pol II to elongation at some genes [12]. Some of these, such as invected and engrailed, are simultaneously targeted by Polycomb silencing proteins, and increase dramatically in expression upon cohesin depletion. PRO-seq confirms that such genes are among those that show the largest pausing decreases (Table S2). The presence of repressor proteins may be one factor, therefore, that determines when cohesin inhibits transition to elongation.
Unexpectedly, cohesin depletion indirectly reduced pausing at most genes that lack cohesin, with a median decrease of 25% (genes in lowest cohesin group in Figure 3D). In control cells, the median pause index at the genes with the highest cohesin levels is 3.7-fold higher than at the genes without cohesin (Figure 1F). However, cohesin depletion increases this ratio to 8.7, primarily because of the broad decrease in pausing at genes that lack cohesin. The overall reduction in pausing might suggest that pausing factors are diminished, but the mRNA levels for all NELF and DSIF subunits are virtually unaffected by cohesin or Nipped-B depletion [13]. Both cohesin-binding and non-binding genes show frequent decreases in promoter PRO-seq density, but these decreases are substantially more frequent at genes that lack cohesin, which likely explains why they also show more frequent decreases in the pause index (Figure 4A). If this indirect general pausing decrease caused by cohesin depletion also occurs at cohesin-binding genes, then it will counteract and obscure many direct increases in pausing caused by cohesin depletion. If so, it can be inferred that cohesin directly facilitates transition to elongation even more frequently than the raw data indicates.
By facilitating enhancer-promoter contact, cohesin could increase the rates of distinct steps of transcription: Pol II recruitment, transcription initiation, or the transition of paused Pol II to elongation. In addition, cohesin bound at the promoters of cohesin-binding genes could directly influence all three steps. The finding that cohesin depletion reduces promoter PRO-seq density less frequently at cohesin-binding genes than at genes that lack cohesin (Figure 4A) argues that recruitment or initiation are less often directly influenced by cohesin. Strikingly, although PRO-seq density is more frequently decreased at the promoters of genes that lack cohesin, there is little difference in the overall effect of cohesin depletion on total Pol II occupancy at cohesin-binding and non-binding promoters as measured by Rpb3 ChIP, further supporting the idea that Pol II recruitment is not usually directly affected by cohesin (Figure 4A). This predicts that genes that lack cohesin would not show as dramatic pausing decrease upon cohesin depletion if pausing was calculated using Rpb3 ChIP instead of PRO-seq data, which was confirmed (Figure S1E). Because the average decrease in PRO-seq at the promoters that lack cohesin is greater than that at cohesin-binding promoters, but the average change in total Pol II occupancy is similar, we deduce that transcription initiation is frequently reduced at genes that lack cohesin.
The more frequent increase in transcriptional pausing at cohesin-binding genes relative to genes that lack cohesin in response to cohesin depletion predicts that cohesin more often directly facilitates transition of paused polymerase to elongation at many genes. To confirm this idea, we compared the frequency of absolute changes in Pol II occupancy of promoters and gene bodies caused by cohesin depletion using the genomic ChIP data for Rpb3 and Ser2P Pol II. We set a statistical threshold for increases or decreases (see Materials and Methods) to determine how many promoters and gene bodies show significant changes upon cohesin depletion. This revealed that total or Ser2P Pol II occupancy rarely increased at the promoters or in the bodies of either cohesin-binding or non-binding genes upon cohesin depletion (Figure 4B). Decreases in Pol II at the promoters were also rare, but more frequent than increases. Cohesin depletion caused significant absolute decreases in Rpb3 and Ser2P Pol II in the bodies of more than half of the cohesin-binding genes, almost twice as often as in genes that lack cohesin (Figure 4B). We conclude, therefore, that cohesin often directly increases transition of paused Pol II to elongation, and less frequently directly influences Pol II recruitment or transcriptional initiation.
Although infrequent, absolute reductions in total Pol II promoter occupancy after cohesin depletion that met the statistical threshold were detected twice as often at cohesin-binding genes than at genes that lack cohesin (Figure 4B). This is still consistent with the finding that the average fold-changes in total Pol II promoter occupancy at cohesin-binding and non-binding genes are similar (Figure 4A), because cohesin-binding genes have higher levels of Pol II at the promoter (Figure 1B). The same absolute change in Pol II occupancy would therefore be a smaller fold-change at most cohesin-binding genes than at most genes that lack cohesin.
We suspect that the reduced pausing that reflects reduced transcription initiation at most genes that lack cohesin is caused by altered expression of factors that act broadly at many or all genes, such as basal transcription factors. Cohesin depletion, however, does not significantly reduce expression of known basal factors such as TFIIB [13]. Prior work has shown that cohesin directly promotes dm/myc expression, and the global pattern of decreases in mRNA upon depletion of cohesin in BG3 cells strongly overlaps those seen in dm/myc mutants [13], [27], [28]. Thus another possibility, consistent with the recent reports that Myc directly amplifies transcription of most if not all active genes in a variety of mammalian cell types [29], [30], is that reduced dm/myc expression could contribute to the broad indirect effect of cohesin depletion on most genes that lack cohesin.
The P-TEFb Pol II kinase, which can be recruited by transcriptional activator proteins bound to enhancers or promoters, stimulates transition of paused Pol II to elongation by phosphorylating NELF, DSIF, and the C-terminal domain of the large subunit of Pol II (reviewed in [31]). Cdk12 is also responsible for a large fraction of Ser2P Pol II phosphorylation [32]. We tested the idea that cohesin promotes transition of paused Pol II to elongation by facilitating recruitment of P-TEFb or Cdk12 by comparing the CycT and Cdk12 ChIP signals in control cells and cells in which cohesin was depleted. We restricted the analysis to those genes in which CycT or Cdk12 was detected in the control cells, to make it possible to detect both decreases and increases. Surprisingly, after cohesin depletion, decreases in CycT or Cdk12 in any transcription units are very rare (Figure 4C). Indeed, CycT and Cdk12 both increase more frequently at promoters and gene bodies than they decrease upon cohesin depletion, and more than twice as often in the bodies of cohesin-binding than in non-binding genes (Figure 4C). Similar frequencies of CycT and Cdk12 increases are seen when all active genes are scored, indicating that increases also occur when the kinases are not detected prior to cohesin depletion. These increases are generally modest, but usually occur in genes with Ser2P Pol II decreases, and are strong enough to give up to a 1.5-fold increase in ratios of the kinases to total Pol II in the bodies of cohesin-binding genes (Figure S3). Because there are several heptapeptide repeats in Pol II, a decrease in the fraction of heptapeptide repeats that are phosphorylated within each Rpb1 molecule could increase the net number of unmodified sites available for kinase binding, even with a decrease in the level of Pol II in the gene body. Based on these findings we conclude that the frequent reduction in phosphorylated Pol II in gene bodies upon cohesin depletion is not caused by reduced presence of the Pol II kinases, and theorize instead that cohesin may facilitate efficient modification of Pol II.
The higher Pol II occupancy of cohesin-binding genes predicts that they should produce more mRNA on average, assuming that RNA processing, transport and stability do not differ substantially between cohesin-binding and non-binding genes. To test this idea, we used existing mRNA measurements [13] to calculate the ratio of steady-state mRNA to PRO-seq density in the gene body, which we define as “efficiency”. This surprisingly revealed that the average efficiency increases significantly with the cohesin level, and that the genes with the highest cohesin levels produce some 2-fold more steady-state mRNA per transcribing Pol II complex than genes that lack cohesin (Figure 5A).
Cohesin is not responsible for the higher efficiency. Upon Nipped-B or Rad21 depletion, the average efficiency of the genes with the highest cohesin levels actually increases modestly (Figure 5B, 5C). We currently do not know why cohesin-binding genes are more efficient, but note that they are highly transcribed, lack histone H3 lysine 36 trimethylation (H3K36me3), and are highly enriched for UG repeats in the nascent transcripts [12]. H3K36me3 and UG repeats regulate RNA processing, and binding of the TDP-43 protein to UG repeats stabilizes long nascent transcripts and reduces missplicing in mammalian neural tissue [33]–[36].
These studies provide compelling evidence that cohesin directly influences transcription. Comparing the effects of cohesin depletion on Pol II occupancy and activity shows that on average, cohesin-binding genes respond differently to cohesin depletion than non-binding genes, allowing us to infer that cohesin directly influences Pol II occupancy and activity at genes that it binds. This direct influence is likely mediated by facilitating looping interactions with enhancers, and also direct effects on the transition of paused Pol II to elongation at the promoter (Figure 6).
Beyond the generally higher levels of Pol II, the most remarkable differences between cohesin-binding and non-binding genes are in promoter-proximal transcriptional pausing. Cohesin-binding genes have a substantially higher average pausing index, and are much more likely than non-binding genes to show increased pausing upon cohesin depletion. Coupled with the decreases in Ser2P Pol II in the bodies of most cohesin-binding genes, the increased pausing upon cohesin depletion argues that cohesin facilitates the transition of paused polymerase to elongation at many genes that it binds.
Cohesin can increase the rate of Pol II transition to elongation by facilitating enhancer-promoter looping, which would bring transcriptional activators and the P-TEFb they recruit into contact with the paused Pol II to stimulate transition to elongation (Figure 6). Indeed, genetic evidence from Drosophila and chromosome conformation capture (3C) data from mammalian cells supports the idea that cohesin facilitates communication and looping between enhancers and promoters [6], [7], [19], [20]. In mammals, cohesin is present at the extragenic enhancers for several mammalian pluripotency genes, the β-globin gene and the T cell receptor locus, and at many CRMs defined by the binding of multiple tissue-specific transcription factors [6], [7], [20], [37]. In Drosophila BG3 cells, cohesin occupies essentially all CRMs, and the reduced Pol II occupancy at many upon cohesin depletion further expands the idea that enhancer-promoter communication is one of cohesin's key roles at several genes. Several studies indicate that cohesin also facilitates looping between sites binding the CTCF protein in mammalian cells to regulate gene expression, but this function is not conserved in Drosophila (reviewed in [1], [2]).
Many studies support a role for enhancers in the assembly of pre-initiation complexes at promoters, but also indicate that they can control other steps, including the transition of Pol II at the promoter to elongation [21]. The steps in activation controlled by a particular enhancer likely depend on the constellation of enhancer-bound transcription factors. If an enhancer's main function is pre-initiation complex formation, then we would expect to see frequent Pol II decreases at promoters upon cohesin depletion. Such decreases, however, are actually infrequent compared to gene body decreases in our experiments. Our data suggest, therefore, that once a gene is active, the primary function of most enhancers is to stimulate paused Pol II to enter elongation.
The analysis presented here cannot definitively address to what extent reduced enhancer-promoter communication explains Pol II decreases in the bodies of cohesin-binding genes caused by cohesin depletion. A critical limitation is that we do not yet know all the contacts between enhancers and promoters, and whether such contacts are cohesin-dependent. We note, however, that the high levels of cohesin at promoters, including at many genes that likely lack enhancers, raises the possibility that cohesin directly interacts with the paused Pol II complex and influences the transition to elongation. These interactions may involve increasing the efficiency with which P-TEFb and Cdk12 modify Pol II or the NELF and DSIF pausing complexes (Figure 6). We suggest that cohesin is more critical for kinase efficiency than for kinase recruitment because at most genes where cohesin depletion reduces Pol II phosphorylation, the kinase level in the gene body actually increases. Also consistent with the idea that promoter-bound cohesin directly influences transition to elongation is the finding that cohesin interacts with the Mediator complex [7]. In addition to facilitating assembly of the pre-initiation complex, Mediator is implicated in recruitment of elongation factors and efficient transcriptional elongation post-initiation [38]–[40].
The idea that promoter-bound cohesin directly influences transition of Pol II to elongation is also supported by prior work showing that cohesin inhibits transition to elongation at several cohesin-repressed genes [12]. In those studies, cohesin and pausing factor depletion experiments revealed that cohesin inhibits transition of Pol II to elongation at a step distinct and likely downstream from those controlled by the NELF and DSIF pausing factors. This inhibition is unlikely to be physical obstruction of Pol II movement because cohesin depletion did not increase the rate of elongation along the induced EcR gene. Moreover, many of the cohesin-repressed genes are among the rare cohesin-binding genes targeted by the PRC2 Polycomb group silencing complex. Thus the presence of repressor proteins may be one factor that determines whether promoter-bound cohesin facilitates or inhibits transition to elongation. Many cohesin-repressed genes are those that show the largest increases in mRNA upon cohesin depletion [13], and more Pol II in the gene bodies in this study. In general, these cohesin-repressed genes show little or no change in Pol II occupancy at the promoter upon cohesin depletion, further supporting the idea that repression largely reflects inhibition of entry into elongation and not Pol II recruitment [12, this study].
We unexpectedly observed that cohesin depletion reduces promoter-proximal Pol II pausing at most genes that don't bind cohesin. Cohesin depletion does not alter expression of genes encoding subunits of the NELF and DSIF pausing factors or the Pol II kinases, and very modestly increases expression of some Mediator subunit genes [13]. The reduction in transcriptionally-engaged Pol II at the promoter measured by PRO-seq is also more significant than the effect on total Pol II occupancy at genes that lack cohesin. We theorize, therefore, that cohesin controls expression of factors that operate broadly to facilitate transcription initiation.
The key suspects for general factors controlled by cohesin are general basal transcription factors, or possibly Diminutive (Dm), the Drosophila Myc protein (Figure 6). Cohesin depletion does not significantly decrease the mRNAs that encode the known basal transcription factors, but does substantially reduce dm/myc transcription. Recent studies in mammalian cells show that Myc directly amplifies transcription of most active genes [29], [30] and therefore reduction of dm/myc expression upon cohesin depletion is expected to alter transcription of many genes, including those that do not bind cohesin. The mammalian studies also indicate, however, that chemical ablation of Myc function increases pausing at Myc target genes [29], [30], [41], while our PRO-seq measurements argue that pausing generally decreases upon cohesin depletion. The mammalian experiments measured pausing by Pol II ChIP, which does not distinguish between promoter-bound Pol II that is transcriptionally-engaged from Pol II that has not initiated transcription, or is somehow otherwise blocked from elongation. In our experiments, Pol II ChIP did not show the same pausing decrease as PRO-seq upon cohesin depletion. Thus, although Myc appears to function as an anti-pausing factor, we cannot rule out the possibility that reduced dm/myc expression is responsible for many of the indirect effects of cohesin depletion on transcription initiation. Direct positive regulation of myc by cohesin occurs in Drosophila, zebrafish, mice and humans [8], [13], [27], [42]. As a key regulator of growth and protein synthesis, it is likely that reduced myc expression contributes to the poor growth of individuals with Cornelia de Lange syndrome and Nipbl(+/−) mice [42], [43].
Based on their higher Pol II occupancy, we expected that cohesin-binding genes would produce more mRNA on average, in proportion to the Pol Il levels. We observed, however, that they produced disproportionately more steady-state mRNA per transcriptionally-engaged Pol II complex, with the genes that have high cohesin levels being twice as efficient as the genes that lack cohesin. Cohesin depletion did not reduce the efficiency, indicating that these genes have other features that make them more efficient. Prior studies show that cohesin-binding genes lack the H3K36me3 histone modification, which is found on other active genes, and is mediated by the Set2 protein that travels with the phosphorylated C terminal domain of the Rpb1 Pol II subunit [44]. H3K36me3 influences RNA processing and vice versa [34], [35]. We currently favor the idea, therefore, that co-transcriptional RNA processing, which also affects RNA transport and stability, is more efficient at cohesin-binding genes. Alternatively, elongation rates, which can be influenced by the higher Pol II density at these genes, may be higher. Cohesin-binding genes are also highly enriched for TG repeats in transcribed plus-strand non-coding sequences 50 to 800 bp downstream of the promoter, and thus the nascent RNAs contain UG repeats [12]. One factor that binds UG repeats is TDP-43 (TBPH in Drosophila), which influences RNA processing, and increases the stability of many long nascent RNAs and splicing fidelity in mouse brain [33], [36]. It is possible that these repeats also participate in cohesin recruitment, which could explain the correlation between cohesin-binding and high efficiency.
Culture of ML-DmBG3-c2 (BG3) cells and RNAi depletion of Nipped-B and Rad21 were conducted as previously described [13].
Genomic chromatin immunoprecipitation of RNAi-treated and mock-treated BG3 cells was performed using Affymetrix Drosophila 2.0R genome tiling arrays as previously described [9] except chromatin sonication was performed under standardized conditions with a Diagenode Bioruptor, and precipitated DNA was amplified using commercial Whole Genome Amplification reagents (Sigma-Aldrich). Reverse-crosslinked chromatin was used to prepare probes for input control arrays. All ChIP-chip data generated for this study is the average of two independent biological replicates. Karen Adelman (NIEHS) provided Rpb3 antibodies, Akira Nakamura (Riken, Japan) provided CycT antibodies, and Bart Bartkowiak and Arno Greenleaf (Duke) provided Cdk12 antibodies. Ser2P Pol II antibodies were purchased from Abcam (ab5095). The Drosophila Rpb3 antibody has been previously been validated for ChIP-chip [18]. The Ser2P Pol II antibody was previously validated for specificity in Drosophila by in vivo inactivation of P-TEFb by the Pgc protein followed by immunostaining and western blots [45]. We retested the Ser2P Pol II antibody by treating BG3 cells with flavopiridol to inhibit P-TEFb followed by western blotting and observed that the major band decreases in intensity over time, although there is an unaffected minor band that co-migrates with the unmodified Rpb1 detected by the 8WG16 antibody (Figure S4). The Cdk12 antibody has previously been validated for ChIP [32]. The Drosophila CycT antibody was previously validated [45] and in our tests, it recognizes a single major protein of the expected size in western blots of whole cell extracts that is reduced by CycT RNAi treatment (Figure S5).
MAT software [46] was used to calculate ChIP enrichment across the Drosophila genome. MAT performs within-array normalization using individual probe DNA sequences, and MAT scores measure enrichment relative to an input control. MAT scores scale linearly with log2 IP/control enrichment values as determined by processing the same data with TiMAT (http://bdtnp.lbl.gov/TiMAT/). MAT is the optimal algorithm for analysis of Affymetrix array ChIP-chip, and provides peak detection sensitivity equivalent to ChIP-seq performed at a density of one read per genome base pair [47], [48]. ChIP-chip data has been deposited in the GEO database (accession no. GSE42399).
Precision global run-on sequencing for control cells, and cells depleted for Rad21 and Nipped-B, was conducted as described elsewhere [15], except that a simplified cell permeabilization nuclear isolation protocol was used. All steps were conducted at 4° unless indicated otherwise. 2.5 to 7.5×108 control or RNAi-treated BG3 cells were collected by centrifugation (1000 g for 5 min), suspended in 5 to 10 mL Phosphate Buffered Saline (PBS) pH 7.0, collected by centrifugation, suspended in 5 mL Buffer W [10 mM Tris-HCl pH 7.5,10 mM KCl, 150 mM sucrose 5 mM MgCl2, 0.5 mM CaCl2, 0.5 mM dithiothreitol (DTT)], and collected by centrifugation. The cells were suspended in 5 mL Buffer P (10 mM Tris-HCl, pH 7.5, 10 mM KCl, 250 mM sucrose, 5 mM MgCl2, 1 mM EGTA, 0.05% Tween-20, 0.5 mM DTT), the suspension was adjusted to 0.14% NP-40, and then incubated on ice for 3 min. The nuclei were washed twice in 5 mL Buffer W, suspended in 1 mL Buffer W, and transferred to a siliconized 1.5 mL microcentrifuge tube. The nuclei were collected by centrifugation at 1000 g for 5 min, suspended in 0.5 mL Buffer F (50 mM Tris-HCL pH 8.3, 40% glycerol, 5 mM MgCl2, 0.1 mM EDTA, 0.5 mM DTT), and counted using a hemacytometer. The nuclei were then suspended in Buffer F to concentration of 40 to 50×105 per microliter, distributed into 100 microliter aliquots in siliconized 1.5 mL tubes, snap frozen in liquid nitrogen, and stored at −80°. The PRO-seq data has been deposited in the GEO database (accession no. GSE42399).
PRO-seq reads for each duplicate sample were summed over the promoter regions and gene bodies of nearly 17,000 annotated transcription units and normalized to the total reads for each sample. Mathematical and statistical analysis of the samples was conducted using Microsoft Excel, R ([49], http://www.R-project.org), and custom programs. After confirming high correlations between the duplicate samples (Table S1), the values for the two duplicates for each condition (Mock, Rad21 depleted, Nipped-B depleted) were averaged (Table S2). PRO-seq active genes were defined as those in which there were an average of at least 1 read per million in both the 200 bp promoter region and the gene body in control samples. PRO-seq changes in the promoter regions, gene bodies, and pausing index upon cohesin depletion were calculated and plotted using R. To rank genes according to cohesin-binding levels, the Rad21 ChIP-chip MAT scores over the promoter regions of PRO-seq active genes were integrated, and the genes broken into four categories ranging from low to high mean cohesin levels, using a geometric distribution (Figure 1A, lower right panel). The lowest group had mean ChIP MAT scores between 0 to 1 in the 400 bp region surrounding the transcription start site, the next highest group had mean scores between 1 to 2, then 2 to 4 and the highest group was greater than 4. This method allowed finer distinction between cohesin-binding levels than quartiles.
To measure the fraction of PRO-seq active genes or putative CRMs that bind or do not bind cohesin (Rad21, Smc1, Nipped-B), bed files showing binding of Rpb3, Ser2P Pol II and CycT at p≤10−3 were generated using MAT software. Binding of Rpb3, Ser2P Pol II and CycT to PRO-seq active genes was determined using BEDTools software [50] to detect overlaps of the bed files with 200 bp promoter regions and gene bodies of PRO-seq active transcription units, and putative active enhancers, with a 1 bp minimum overlap. Existing Smc1 and Nipped-B ChIP-chip data for BG3 cells ([9], GEO accession no. GSE9248) was used to determine which genes and putative enhancers bind cohesin.
For some analyses, the differences in the ChIP enrichment (MAT scores) for Pol II or Pol II kinases were calculated at each of the nearly 2.8 million points measured across the genome. The distributions of the differences, means, medians and standard deviations of the differences were determined using R. In all cases, there was minimal skew in the distribution of differences, and both the mean and median differences were nearly identical and close to zero. The thresholding tool of the Affymetrix Integrated Genome Browser (IGB; http://www.affymetrix.com/partners_programs/programs/developer/tools/download_igb.affx) was used to generate bed files indicating where the Rad21 RNAi sample enrichment differs from the enrichment in control cells by at least two standard deviations from the median genome-wide difference over at least 105 bp (three microarray features, example in Figure 2). BEDTools was used to detect overlaps between these intervals and the 200 bp promoter regions or gene bodies of the PRO-seq active genes, or predicted extragenic CRMs to identify those with significant changes. The rare instances in which a feature scored positive for both a decrease and an increase in ChIP signal were resolved by visual inspection. In most cases these reflect both a small increase and a small decrease, and the genes were rescored as having no significant change. This method agrees with changes in Pol II occupancy after Rad21 depletion previously measured at multiple genes by quantitative real-time PCR ChIP in independent experiments [12].
To measure mRNA production efficiency we used expression data for 13,132 genes in BG3 cells previously measured by Affymetrix Drosophila GeneChip 2.0 for mRNA levels in control and Rad21 depleted BG3 cells ([13], GEO accession no. GSE16152). For those genes represented by multiple probes, we summed the total signals for all probes, and used the total to compare to the gene body PRO-seq signals.
|
10.1371/journal.ppat.1006478 | Treatment with integrase inhibitor suggests a new interpretation of HIV RNA decay curves that reveals a subset of cells with slow integration | The kinetics of HIV-1 decay under treatment depends on the class of antiretrovirals used. Mathematical models are useful to interpret the different profiles, providing quantitative information about the kinetics of virus replication and the cell populations contributing to viral decay. We modeled proviral integration in short- and long-lived infected cells to compare viral kinetics under treatment with and without the integrase inhibitor raltegravir (RAL). We fitted the model to data obtained from participants treated with RAL-containing regimes or with a four-drug regimen of protease and reverse transcriptase inhibitors. Our model explains the existence and quantifies the three phases of HIV-1 RNA decay in RAL-based regimens vs. the two phases observed in therapies without RAL. Our findings indicate that HIV-1 infection is mostly sustained by short-lived infected cells with fast integration and a short viral production period, and by long-lived infected cells with slow integration but an equally short viral production period. We propose that these cells represent activated and resting infected CD4+ T-cells, respectively, and estimate that infection of resting cells represent ~4% of productive reverse transcription events in chronic infection. RAL reveals the kinetics of proviral integration, showing that in short-lived cells the pre-integration population has a half-life of ~7 hours, whereas in long-lived cells this half-life is ~6 weeks. We also show that the efficacy of RAL can be estimated by the difference in viral load at the start of the second phase in protocols with and without RAL. Overall, we provide a mechanistic model of viral infection that parsimoniously explains the kinetics of viral load decline under multiple classes of antiretrovirals.
| Antiretroviral therapy in HIV-1 infection leads to characteristic viral load decay profiles. Interpretation of this profile with mathematical models has provided important insights into the dynamics of the viral lifecycle (e.g., the turnover of infected cells). Here we develop a unified model and analyze viral load data from HIV-infected participants under treatment with and without an integrase inhibitor. Our model offers a new explanation for the observed differences in the decay profiles, and strongly supports the hypothesis that the second phase of viral decay is due to the loss of long-lived resting CD4+ T-cells with slow proviral integration. We estimate parameters of this decay and estimate the efficacy of an integrase inhibitor.
| Viral dynamics analysis is a powerful tool to probe the lifecycle of viral infections [1–3]. In the case of HIV-1, treatment with reverse transcriptase inhibitors (RTI) and protease inhibitors (PI) resulted in a two-phase decline in plasma viral load [4–6]. Mathematical modeling attributed the first phase to the loss of productively infected cells with a short half-life (t1/2~0.7 days), presumably activated CD4+ T-cells [7]. The second phase was attributed to the loss of “long-lived” infected cells with a slower loss rate (t1/2 ~14 days) [4]. The nature of the cells responsible for this second phase is still debated, and possibilities include, among others, macrophages or resting CD4+ T cells [4,8,9]. In these early studies, the first phase lasted for about 6–11 days, and the second phase started after a viral load drop of 93%-99% from baseline, implying that the long-lived infected cells contributed ~1–7% of the total virus before treatment [4].
Different classes of drugs may lead to different patterns of viral decline, depending on where in the viral lifecycle the drug acts [10,11]. Modeling of one of the first clinical trials of HIV-1 treatment with an integrase strand transfer inhibitor (InSTI) showed that the viral load decline had first and second phase rates similar to those seen with InSTI-free regimens, with half-lives of ~1.2 days and ~15.5 days, respectively [12]. Although the data was sparse for the second phase, a unique characteristic of the viral load decline was that this phase started at much lower viral load levels than previously seen with InSTI-free regimens [12]. A recent study, with more frequent data, comparing InSTI-free and InSTI-containing regimens found that the rates of these two phases of decay were somewhat slower in the presence of an InSTI [13]. This study also confirmed that for InSTI-based regimens the viral load at the start of the second phase is ~1 log below the viral load at the start of the second phase for InSTI-free regimens [13] (Fig 1A). Another characteristic of viral decay under an InSTI is the separation of the first phase of decline into two sub-phases (phase 1a and phase 1b, Fig 1B) [10,14,15]. Analyzing the viral decay during the first 10 days of therapy with InSTI-containing regimens, we estimated [14] the half-life of the cells contributing to phase 1a as ~0.8 days, similar to the previously estimated half-life of short-lived infected cells (~0.7 days) [16]. In the same study, the half-life of the population of cells contributing to phase 1b was estimated as ~4.6 days, which is much shorter than the previously estimated half-life of the cells contributing to the second phase (~14 days) [4].
Several mechanistic models were proposed to explain the differences in the profile of viral decay between InSTI-free and InSTI-containing regimens [9,10,12,14,15,17], but a thorough comparison with appropriate data is still lacking. Here, we analyze in detail several data sets under InSTI-based and InSTI-free therapies using a unified approach. Our experimental data includes more frequent sampling throughout the first month of the viral load decline under InSTI-based therapy than any previous modeling study. We develop mathematical models that separate the pre-integration and post-integration state in both short-lived and long-lived infected cells. We then use a mixed-effect modeling approach to simultaneously fit the viral load data in InSTI-free combination drug therapy and therapies including raltegravir (RAL), an InSTI [13,14,16,18].
We discovered that the best explanation for all the available kinetic data is that short-lived infected cells have a fast integration rate and a short viral production period, while long-lived infected cells have slow integration, but an equally short viral production period. Thus, we attribute the long life span of some infected cells to slow proviral integration. We propose that both types of cells (short- and long-lived) represent mostly infected CD4+ T-cells but in activated and resting states, respectively.
We analyzed a total of 47 participants in three HIV-1 treatment data sets (see Methods) [13,16,18]. Eight participants were treated with a 4-drug combination without InSTI (“quad-therapy) [16]; eleven participants were treated with a combination of RAL and two RTIs (“RAL-combination”); and twenty-eight people were treated with RAL monotherapy for 9 days (“RAL-monotherapy”).
Fig 1A presents the plasma viral load of all the participants from the quad-therapy and RAL-combination protocols, as well as the respective medians. This allows a clear visualization of the different patterns of decay under the two regimens (Fig 1B). In supplementary Figs B-D in S1 Text, we present the viral load data for all 47 individuals analyzed. Examination of these data shows that the second phase of viral load decline in individuals treated without RAL starts before than in participants under RAL-therapy (~ 5 days post-treatment initiation vs. ~10 days). At the start of the second phase, the viral load level in the RAL-combination participants is ~1-log lower than the viral load at the start of the second phase in the quad-therapy group (denoted by Δ in Fig 1A). We observe also that participants treated with RAL seem to have an intermediate phase of decline (phase 1b), which is slower than the first phase in the quad-therapy group, but faster than their second phase (Fig 1B). Finally, the slope of the last phase (phase 2) for RAL-combination therapy seems similar to the second phase slope for quad-therapy.
Thus, these data sets with more frequent measurements confirm the differences in viral load decay profiles between InSTI-containing and InSTI-free regimens presented before [12–14]. These observations of the differences between the two types of drug regimens, lead to four questions: i) Are these empirical observations borne out by rigorous statistical analyses? ii) Could the early second phase in the quad-therapy be due to the same mechanisms (i.e., decay of the same population of infected cells) as phase 1b in the InSTI-based treatment, which starts also at about ~5 days? iii) If not, what populations of infected cells contribute to each phase of decay in each type of treatment? And iv) Why do the second phases of decay seem have the same slope, but occur at very different viral load levels in the two types of regimens? We now answer these questions in turn.
To put the observations of the previous section in a rigorous footing, we used a mixed-effects approach [19] using two empirical models, allowing for two or three exponential decays, respectively (see Methods). Two exponentials correspond to the two phases of decay (as it seems to be the case in the quad-therapy) and three exponentials correspond to three phases of decay (phase 1a, 1b and 2). We fitted these two models to the viral load of the three data sets simultaneously to assess which model is more consistent with the data.
We found that the three-exponentials model did not fit the quad-therapy statistically better than the two-exponentials model (p = 0.89), while for the RAL-containing regimens the reverse is true (p<0.0001). Moreover, the initial decay rate was not significantly different between the treatment protocols, but the second decay rate was significantly different between the RAL-containing regimens (phase 1b) and the quad-therapy (phase 2) (p = 0.0007), with estimates of ~0.15 day-1 and ~0.05 day-1, respectively (Table 1). The estimate for the third decay rate in the RAL-combination arm (phase 2) was ~0.045 day-1 (Table 1), which is very similar to the phase 2 decay rate in the quad-therapy (~0.05 day-1).
These statistical results, thus, indicate that it is likely phase 1b for the RAL-containing treatments has a different viral source from phase two in the treatment without RAL. To explore the potential different populations of infected cells contributing to the viral load, we next analyzed the data in the context of mechanistic models of viral dynamics. We generalized a model developed to analyze the effects of RAL [14] by including the effect of reverse transcriptase and protease inhibitors and a compartment of long-lived infected cells, as illustrated in Fig 2A (details in Methods).
To understand whether phase 2 (without RAL) and phase 1b (with RAL) have the same origin, we used the mechanistic model described by Eq (1) in Methods, without long-lived cells [10,14] (see section 2 in S1 Text).
The model shows that there are two conditions that must be satisfied to observe two early phases of decay (viz. phase 1a and 1b): i) RAL must be present and ii) the efficacy of RAL must be above a critical threshold (ω>ξ(δ1+k)/(δ2+k), where ω is the efficacy of RAL, ξ is the combined efficacy of RTIs and PIs, δ1 and δ2 are the loss rates of short-lived infected cells before and after integration respectively, and k is the integration rate in these cells), but less than 100%. In the case of RAL monotherapy, i.e., ξ = 0, the value of the critical threshold is 0, and thus with monotherapy one should always see two early phases of decay (section 2 in S1 Text) [15]. In the case of therapy without an InSTI, ω = 0 and the model predicts (see section 2.2 in S1 Text) that after a short delay (so-called shoulder phase) only one exponential-like decay is observed early on, i.e., during phase one [9].
The interpretation of these results is that in the presence of RAL, the loss of cells in the pre-integration stage is uncovered (phase 1b), whereas without RAL, integration proceeds so fast that it is not possible to observe the loss of those cells (i.e., there is no phase 1b). Thus, to explain the second phase one needs another source of infected cells (e.g. long-lived HIV-infected cells).
In the RAL-containing regimens, phase 1 (phases 1a and 1b together) lasts longer (~10 days) than phase 1 during quad-therapy (~5 days). Consequently, for RAL-containing regimens phase 2 starts at a later time and at a lower viral load (Fig 1 and [12,13]). But why does this happen?
To investigate this issue, we used the model presented in Eq (1) (Fig 2A) and fitted it to the three datasets simultaneously. We found for the best fits that the rate of integration in the “long-lived cells” was very slow (k1<0.05 day-1) when compared to that of “short-lived cells” (k = 2.6 day-1). In addition, the estimate of the loss rate of long-lived cells producing virus (δM2 ~0.90 day-1) was very close to the corresponding rate in short-lived cells (δ2 = 0.86 day-1). This suggests the hypothesis that “long-lived cells” have slow integration leading to their long lifespan, but after becoming productively infected they are lost at a rate equal to that of short-lived cells, i.e., δM2 = δ2. Indeed making this assumption gave the best fitting model (with lowest AIC, see Table C in S1 Text). Thus, in the best model to describe the data, we have one population of productively infected cells with a short lifespan, which is generated from two populations, one with fast integration events and another with slow integration events (Fig 2B). In what follows we present analyses of the data using this model, which we call the SRI model (Slow and Rapid Integration model) to differentiate from the standard model in Fig 2A.
The SRI model has three compartments (Fig 2B), corresponding to cells with fast integration (I1) that quickly become productively infected (I2) and cells with slow integration (M1), which eventually also become productively infected (I2). In this new interpretation, the previous M2 compartment in the standard model is simply the subpopulation of productively infected cells (I2) that were generated from slow integration events (i.e., from M1), but it does not correspond to a physiologically separate population of cells.
Fig 3 shows the predicted population viral load profiles for each treatment protocol resulting from fitting the SRI model to all the data simultaneously, and Table 1 summarizes the corresponding parameters. In Fig 4, we show best fits of the SRI model to the viral load data for representative individuals from the three treatments protocols. Individual best fits for all participants are shown in Figs B-D in S1 Text and the estimated parameters in Table E in S1 Text.
Overall, we found that the first phase of decline, which is phase 1 or 1a without or with RAL, respectively, corresponds to the loss of short-lived productively infected cells (~δ2, t½~ 20 h). Phase 1b only exists in RAL-based therapy and corresponds to the decay of short-lived cells pre-integration (at rate ~δ1+(1-ω)k, t½~ 1.8 days). Phase 2 corresponds to the decay of long-lived cells pre-integration (at rate ~δM1+(1-ω)k1), which in the absence of RAL (ω = 0) is slightly faster (t½ ~ 19 days) than in the presence of RAL (t½ ~ 33 days), as RAL slows the loss of these cells by integration. Note that this estimated half-life in the absence of RAL is consistent with previous estimates (~14 days, range 6 to 24 days) [4,13]. Also, the difference in the second phase slope between the treatment regimens in the absence and presence of RAL (0.037 day-1 vs. 0.021 day-1) is sufficiently small that over 30 days of treatment it can’t be discerned in the figures (Fig 3).
From the best-fit parameters, we can calculate the relative proportion of infection events leading to infected cells with fast and slow integration (see section 3.1 in S1 Text). We estimate that ~96% of infections result in cells with fast integration (I1) and only 4% generate infected cells with slow integration (M1). Yet, due to the long-lived nature of the latter population, before treatment the pool of cells with slow integration (M1) represents about 40% of total infected cells in the SRI model (i.e., M1 is 40% of I1+M1+I2).
These results offer for the first time a consistent picture for the three treatment scenarios based on the same mechanistic model. However, we still need an explanation for the observed lower viral load during the second phase decline in InSTI-based therapies.
Based on the SRI model, we can directly calculate the reduction in viral load during the second phase with RAL treatment with respect to the viral load in the second phase with quad therapy. We find that this reduction is approximately 1−(1−ω)eωk1t (section 3.2 in S1 Text). For example, at 12.5 days post-treatment, i.e. the beginning of phase two in the RAL-containing regimen, the viral load in the RAL-combination therapy is reduced by ~93% with respect to the viral load in the quad-therapy group (Fig 3). Interestingly, if we linearly extrapolate the second phase of decay for therapy without and with RAL back to time 0, then this reduction in viral load is given by ~ω (Fig 3). Therefore, our model provides a way to directly estimate the efficacy of RAL in blocking integration, ω, when compared to a highly-potent therapy without RAL, such as the quad-treatment. Thus, the SRI model intrinsically explains this difference in viral loads, based on the dynamics of different infected cell populations, and without the need for any extra assumptions.
Still, it would be important to have some intuitive understanding why there is a lower level of virus at the beginning of phase 2 in RAL-based regimens. In the SRI model, productively infected cells (I2) generate the observed virus. The different phases of virus decay correspond to loss of different cell populations contributing to generating those productively infected cells. In more detail and as shown in Movie S1, viral decay is observed as follows. Initially productively infected cells (I2) are lost quickly (rate δ2), corresponding to phase 1 or phase 1a, depending on the type of treatment. At the same time, cells progressing through integration (I1 and M1) replenish the productively infected cells (I2). In the presence of RTIs and PIs without InSTIs, the pool of cells with fast integration (I1) is quickly exhausted. In this case, the second phase can be observed when the main contribution to productively infected cells (I2) comes from the conversion of slowly integrating cells (M1) at rate k1 into productively infected cells. The productively infected cells then have an effective decay rate given by (~δM1+k1), which is the rate limiting step. Paradoxically, in the presence of an InSTI, the pool of cells that can integrate fast (I1) is exhausted more slowly, because integration is blocked/slowed but not prevented completely. This results in phase 1b with slope (~δ1+k(1-ω)) due to the slow decay of cells in I1. Together phases 1a and 1b last longer than phase 1, because it takes longer to lose most cells in I1. Again, as in the quad-therapy, the second phase is observed when the main contribution to productively infected cells comes from the slow integrating cells (M1), which in the presence of an InSTI occurs later, because integration is slowed in this compartment too. This explains the lower viral load level at the start of the second phase with an InSTI regimen. All the infected cell populations are decaying from early on at the start of treatment, but the observed viral load reflects the decline of different populations in turn, as the populations with faster turnover are lost. However, it raises another question, why does the second phase in the quad-treatment start so early (~5 days), even earlier than what had been observed before in RTI+PI treatments?
In the quad-therapy data set the first phase lasted only ~5 days, whereas early studies with less potent RTIs and PIs indicated that this phase lasted for about 6 to 11 days [4]. Using the SRI model, we found that the higher the effectiveness of the RTI+PI regimen (i.e., the larger the effectiveness of the RTI, η and of the PI, ε), the earlier the switch from the first to the second phase (Fig E(a) in S1 Text). In fact, we can calculate (Eq. S15) that the difference in the times of transition from ~8 days in the earlier studies to 5 days with the quad-regime implies that the latter had an efficacy ~1.6-fold higher, i.e., early therapies had an efficacy ξ ~ 0.62. This result compares favorably with a previous study that indicated that early therapies had an efficacy ~75% of the quad-treatment protocol [20]. In the quad regimen with higher efficacy, the transition from the first to second phase occurs earlier and at a slightly lower viral load, because of faster viral decay (Fig E(b) in S1 Text), as discussed in [16]. Thus, the SRI model clearly explains the early transition to the second phase in the quad-therapy regimen is due to the high efficacy of the combination treatment.
Analysis of viral dynamics under different treatment regimens provides insight into the possible virus sources for the observed phases of plasma viral load decay. Under RAL-containing regimens the original phase 1 decay described for PI and RTI containing regimes [4,5] is replaced by two phases, called 1a and 1b [10,14,15], which are then followed by a second phase that starts at lower viral load levels than the start of the second phase with InSTI-free regimens due to the presence of phase 1b. Here we presented for the first time a detailed quantification of the dynamics of cell subpopulations contributing to the observed viral kinetics over the first month of treatment. There are other slower phases of decay at later times, which are not studied here [3,21,22].
The mechanistic reasons for the peculiar viral load decay profile under InSTI regimens have been the subject of previous studies. Murray et al. [12] first developed models to analyze the viral load decay kinetics under RAL. The models presented in [12] are different from our SRI model, they were not fit to individual patient data, and they cannot recapitulate the phase 1b observed in the data. However, the models in [12] that could qualitatively explain the data had the interesting characteristic that the second phase virus was produced from the same cell population as the first phase virus. The difference was in the infection kinetics (e.g., proviral integration), which could be fast (in short-lived cells) or slow (in long-lived cells). This idea echoes our final model where we found that the death rate of short-lived cells was the same as the death rate of long-lived cells after they both start producing virus (i.e., δM2 = δ2).
Sedaghat et al. [9,17,23], Gilmore et al. [10] and Wang et al. [15] presented detailed mathematical analyses of models similar to ours. However, they did not analyze clinical data and/or had only very sparse viral load data, and, thus, had to make various assumptions about parameters and their relationships. For example, Sedaghat et al. [9,17,23] studied without fitting data the features of the second phase of viral decline. They concluded that in the most likely scenario the loss rate of long-lived infected cells pre-integration (~δM1+k1) should be greater than the loss rate of long-lived productively infected cells (~δM2); and that the latter should be the slope of the second phase (equal) in both treatment cases (with and without integrase inhibitors). By quantitative analyses of viral load data and fitting, we rather found that δM2>δM1+k1 even when using the conclusions in Sedaghat et al. [9,17,23] as initial guesses in the fitting procedure. Thus, we now show that the second phase slope is the loss rate of long-lived cells pre-integration, i.e., ~δM1+(1-ω)k1, and that this slope is slightly different with and without an integrase inhibitor, as also found previously by simple linear regression [13]. Interestingly, Sedaghat et al. briefly considered an “intriguing possibility” consistent with our results that the loss rate of long-lived productively infected cells is the same as that for short-lived productively infected cells, but dismissed it [17].
Our detailed data-driven comparison between treatment regimens with and without RAL leads to a novel interpretation of the profile of HIV-1 decay under treatment. The existence of an early (fast) and a late (slow) phase of viral decay is explained, as before, by decay of two “different” cell populations. However, the best fit of the SRI model supports the idea that the dynamics of proviral integration distinguishes these two cell populations, but they both could be CD4+ T-cells. The early fast decay in viral load (including phase 1a and 1b) is most likely due to infection of activated CD4+ T-cells, which quickly progress to viral production. The slow late decay (phase 2) is likely due to infection of resting CD4+ T-cells, which then progress slowly to provirus integration. This interpretation is consistent with data indicating that in resting CD4+ T-cells integration proceeds slowly [24–26]. In the SRI model a fraction of these cells eventually complete integration and become productively infected cells. The resting cells have a slow loss rate before integration, thus they are long-lived, but we found that after becoming productively infected they have a high loss rate, similar to short-lived productively infected cells. One explanation for this result could be that some stimulus activates these cells, which then quickly integrate and start producing virus. Although this interpretation of the data may be surprising, with hindsight, one recognizes that the differences observed in viral load decay profiles between treatments with and without RAL must be due to some process involving proviral integration, which is the step of the lifecycle blocked by the drug. We note that this interpretation is unlikely to depend on the assumptions of the SRI model, which we tested by examining other models and parameter ranges without finding any important effect (section 4.4 in S1 Text). In particular, alternative models, where the second phase is due to cells with virus already integrated that are lost slowly, does not generate viral decay profiles consistent with those observed. Indeed, we started our analysis with this assumption (Fig 2A) and found that the resulting differences in viral loads in the second phase between treatments with or without RAL were too small compared with the data (see section 4.3 and Fig A in S1 Text).
We estimate that before treatment, the vast majority of virus is produced by short-lived cells [4], which is indicated by the large decay in viral load during the first phase, and only a small fraction of virus is produced from cells with slow proviral integration. Several recent reports have indicated that non-integrated HIV DNA in resting CD4+ T-cell is capable of integration and generating productive infection upon activation [24,27–32]. Indeed, this phenomenon has been dubbed “pre-integration latency” [7]. Here we also find that the loss of cells in this pre-integration stage in the absence of InSTI therapy is very slow (t1/2 ~ 19 days), consistent with a study measuring the decay of unintegrated HIV DNA in chronically infected patients under successful HAART that reported a half-life of 26 days [33]. Some studies have described a very rapid decay of the pre-integration complex (PIC) with half-life of ~1 day in resting cells in vitro [24,28]. However, there are other studies that report stable accumulation of HIV DNA transcripts over longer timespans [34] with inducible pre-integration latency over times up to 28 days in vitro [27,30]. It has also been suggested that in vivo the half-life of pre-integration latent cells can be longer than in vitro [35]. Moreover, recent evidence has shown that 2-LTR circles, which are very stable, can be cleaved by the viral integrase forming a viable substrate for integration and rescuing viral production [29]. Altogether, it is possible that in vivo resting infected cells in the pre-integration stage have a broad distribution of half-lives. Thus, although the half-life of “pre-integration latency” is somewhat controversial, we think that our results add a new piece of evidence to indicate that in vivo on average this half-life is long and physiologically relevant, since these cells can be activated into producing virus.
An alternative explanation invoked for the second phase of viral decay is that the long-lived cells are infected macrophages [4,9,23]. Some results indicate that HIV DNA integration is slow in primary blood monocytes [36], however other studies have measured fast integration (~3.4 hours) [23]. In any case, it is thought that infected macrophages are long-lived even after starting to produce virus [7], which is inconsistent with our findings. Moreover, it is unlikely that infected monocytes with unintegrated HIV DNA form as large a pool of infected cells as estimated here [37]. However, it is possible that in addition to the cell compartments in the SRI model, a compartment of long-lived productively infected cells, such as macrophages (Fig 2A), exists. The contribution of these cells to the second phase in viral load should be minimal and, thus, it would not affect the results presented by the SRI model.
Altogether, the most parsimonious interpretation of the analyses of these detailed data sets is that the slow second phase viral decay corresponds to the loss of infected cells with unintegrated provirus (possibly resting CD4+ T-cells) by death, by loss of the pre-integration intermediates, or by integration of the HIV DNA to generate productively infected cells. This description is valid both in regimens with and without an InSTI, and thus alters the interpretation of the second phase of viral decay presented in previous studies [4,9,17]. Furthermore, our results explain why the second phase decay is slightly slower in the RAL containing regimen than in the RAL-free regime, as found previously [13]. This difference is small because it depends on the integration rate k1, which is slow.
In conclusion, we have proposed a new model that explains the main differences of viral load decline under InSTI therapy (as prototyped here by RAL) when compared with InSTI-free regimens: i) the two early phases 1a and 1b of viral decay; ii) the ~1 log lower viral load at the start of the second phase; and iii) the nearly identical second phase slope with and without an InSTI. Fitting this model to frequently sampled viral load data indicates that the second phase of viral decay is due to a subset of cells completing integration very slowly, which most likely are infected resting CD4+ T-cells that eventually complete integration to produce virus and then die quickly.
The first set of data is from a trial with a highly active antiretroviral quad-regimen [16], where nine chronically HIV-1-infected individuals were treated with the boosted PI lopinavir-ritonavir (1,066 and 266 mg/day, respectively), and the RTIs efavirenz (600 mg/day), lamivudine (300 mg/day), and tenofovir DF (300 mg/day). Plasma viral loads were measured by RT-PCR (Roche Amplicor Ul- trasensitive Cobas 1.5) with a limit of quantification (LoQ) of 50 copies/ml at 6 h intervals during the first 72 h, then daily until day 10, and then weekly until day 28 [16]. One participant could not be analyzed, because the precise time of each viral load quantification was not available. We refer to this data set as the “quad treatment” data.
The second data set is from 28 HIV-1-infected participants treated with RAL monotherapy for 9 days [18]. Participants received 100, 200, 400, or 600 mg RAL twice daily, and had plasma HIV RNA measured before the first dose, 6 and 12 hours post-dose, and on days 1, 2, 3, 4, 7, 9. Plasma was assayed for HIV-1 RNA using the Amplicor HIV-1 Monitor Assay, version 1.5 with a LoQ = 400 copies/ml. If the result was below that level, the UltraSensitive HIV-1 Monitor Assay (LoQ = 50 copies/ml) was used (Roche Molecular Diagnostics, Alameda, CA). We treated plasma HIV RNA data below the limits of quantification as censored at the corresponding values. Because no difference in the pattern of decay or efficiency between the different doses was found [14,18], we analyze all doses together. We refer to this data set as the “RAL-monotherapy” data.
The third set of data was obtained from 11 HIV-1-infected participants treated with a combination of FTC/TDF 200 mg/300 mg daily plus RAL 400 mg twice daily enrolled in the intensive viral dynamics A5249s substudy of ACTG protocol A5248 [13]. Plasma HIV-1 RNA was measured at baseline, and at 2, 4, 6, 12, 18, 24, 30, 36, 42, and 48 hours and days 3, 4, 7, 10, 14, 21 and 28 after treatment initiation. Plasma was assayed for HIV-1 RNA using the Amplicor HIV-1 Monitor, version 1.5, UltraSensitive protocol (LoQ = 50 copies/mL; Roche Molecular Systems, Branchburg, NJ). We refer to this data set as the “RAL-combination therapy” data.
The study protocols were approved by the respective institutional review board at each of the participating clinical research sites, and all subjects provided signed informed consent (see details in [13,16,18]). All data analyzed were anonymized.
We first analyzed the data with double and triple exponential decay curves, corresponding to sums of (j = 2 or j = 3) terms of the form Cje−αjt, to see if the data was consistent with two or three phases of decay and what the respective slopes (αj) were.
We then generalized our model developed to analyze the effects of RAL [14] by including the effect of protease inhibitors and a compartment of long-lived infected cells, to account for the long-term (up to 30 days) follow-up. The model is a modification of the standard model of virus dynamics [1,3] separating the compartment of infected cells in pre-integration and post-integration states in both short-lived and long-lived infected cells. A schematic of the model is shown in Fig 2A, and the full system of differential equations describing the dynamics is given in the S1 Text (section 1). We assumed that each drug has the same efficacy in short and long-lived infected cells and that the dynamics of virus is much faster than that of infected cells. Both of these assumptions have been widely used before [4,5,9,14,15,17], and result in the simplified model
dI^1dt=(1−ξ)T^V−δ1I^1−k(1−ω)I^1dM^1dt=(1−ξ)M^V−δM1M^1−k1(1−ω)M^1dVIdt=kc(1−ω)I^1−δ2VIdVMdt=k1c(1−ω)M^1−δM2VM
(1)
where VI and VM denote virus produced by short- and long-lived infected cells, I and M, respectively, total virus is V = VI + VM, ω is the effectiveness of the InSTI and ξ is the combined effectiveness of the PIs (ε) and RTIs (ε) given by (1-ξ) = (1-η)(1-η). See section 1 in S1 Text for further details about the definitions of the variables and parameters. In particular, note that without loss of generality, we rescaled the variables so that the rate of infection, which in our model includes reverse transcription, and the viral production rate no longer appear in the equations, although they can be different between short-lived, I, and long-lived, M, cells.
We fitted the model to the plasma HIV-1 RNA data using non-linear mixed effects (NLME) models. In this approach, we represent the parameters for each individual (i) as μi = θeφi, where θ is the median value of the parameter in the population, and φi the random terms, which are normally distributed with zero mean and a variance to be estimated. We fitted the model in Eq (1) simultaneously to the three sets of data using the software MONOLIX (www.lixoft.eu), to estimate the population parameters (by maximum likelihood) and the variances of the random effects. In total, we fit 613 data points of 47 participants simultaneously to estimate between 4 and 6 parameters, and their corresponding variances in various versions of the model. For each model fit, we estimate the log-likelihood (log L) and compute the Akaike Information Criteria (AIC = -2log L+2m, where m is the number of parameters estimated) [38]. We compared the AIC for different model assumptions and computed ΔAIC, the difference in AIC between the best model and other model scenarios analyzed. We assumed models had similar support if their ΔAIC<2 [38]. P-values were based on the log-likelihood ratio test and significance was assessed at the α = 0.05 level.
To implement the fitting, we first explored the possible range of values for the new parameters in our model, δM1, δM2 and k1 describing long-lived infected cells (see Fig 2A), keeping fixed the remaining parameters based on estimates from previous studies [14,16,39], and obtaining the values of T^ and M^ from steady state assumptions (see Eq. S.4). We performed Latin-hypercube sampling resulting in 64,000 (40 values for each parameter) simulations of the model in Eq (1) for different values of the parameters [40]. We looked for combinations of parameters that showed viral decay properties consistent with the observations in the data (section 4.1 in S1 Text). Namely, a reduction in the viral load greater than 70% at the start of the second phase of decline under RAL-combination therapy relative to start of the second phase in the quad-based treatment; and similar second phase slopes. We found that these biological conditions are only satisfied for small values of δM1 (<0.07 day-1) and k1 (<0.08 day-1), along with values of δM2 greater than 0.15 day-1 (Fig A in S1 Text). These results provided starting guesses for these parameters in the viral load fits. To simplify the fitting procedure and obtain convergence, we fixed five parameters (ξ, k, ω, c, and the fraction of virus produced from long-lived infected cells, see section 4.2 in S1 Text) at values previously estimated by us and others [14,39]. In addition, the values of T^ and M^ in Eq (1) were obtained from the other parameters based on pre-therapy steady state assumptions (see Eq. S.4 in S1 Text). Still, it was difficult to estimate independently both δM1 and k1, since when one increased the other tended to decrease (see Tables A and B in S1 Text). Therefore, we performed the fitting procedure for different fixed values of δM1, and found the best fits when δM1 = 0.01–0.02 day-1, based on AIC (see Table B in S1 Text). We then estimated the remaining parameters: V0, δ1, δ2, δM2 and k1.
|
10.1371/journal.pntd.0005985 | The cost and cost-effectiveness of rapid testing strategies for yaws diagnosis and surveillance | Yaws is a non-venereal treponemal infection caused by Treponema pallidum subspecies pertenue. The disease is targeted by WHO for eradication by 2020. Rapid diagnostic tests (RDTs) are envisaged for confirmation of clinical cases during treatment campaigns and for certification of the interruption of transmission. Yaws testing requires both treponemal (trep) and non-treponemal (non-trep) assays for diagnosis of current infection. We evaluate a sequential testing strategy (using a treponemal RDT before a trep/non-trep RDT) in terms of cost and cost-effectiveness, relative to a single-assay combined testing strategy (using the trep/non-trep RDT alone), for two use cases: individual diagnosis and community surveillance.
We use cohort decision analysis to examine the diagnostic and cost outcomes. We estimate cost and cost-effectiveness of the alternative testing strategies at different levels of prevalence of past/current infection and current infection under each use case. We take the perspective of the global yaws eradication programme. We calculate the total number of correct diagnoses for each strategy over a range of plausible prevalences. We employ probabilistic sensitivity analysis (PSA) to account for uncertainty and report 95% intervals.
At current prices of the treponemal and trep/non-trep RDTs, the sequential strategy is cost-saving for individual diagnosis at prevalence of past/current infection less than 85% (81–90); it is cost-saving for surveillance at less than 100%. The threshold price of the trep/non-trep RDT (below which the sequential strategy would no longer be cost-saving) is US$ 1.08 (1.02–1.14) for individual diagnosis at high prevalence of past/current infection (51%) and US$ 0.54 (0.52–0.56) for community surveillance at low prevalence (15%).
We find that the sequential strategy is cost-saving for both diagnosis and surveillance in most relevant settings. In the absence of evidence assessing relative performance (sensitivity and specificity), cost-effectiveness is uncertain. However, the conditions under which the combined test only strategy might be more cost-effective than the sequential strategy are limited. A cheaper trep/non-trep RDT is needed, costing no more than US$ 0.50–1.00, depending on the use case. Our results will help enhance the cost-effectiveness of yaws programmes in the 13 countries known to be currently endemic. It will also inform efforts in the much larger group of 71 countries with a history of yaws, many of which will have to undertake surveillance to confirm the interruption of transmission.
| Yaws is a non-venereal treponemal infection. The disease is targeted by WHO for eradication by 2020. Testing is envisaged for diagnosis to confirm of clinical cases during treatment campaigns and for surveillance to certify the interruption of transmission. However resources available to the global eradication programme are severely limited and the cost of testing must be contained. Testing requires simultaneous detection of antibodies to both treponemal and non-treponemal antigens for diagnosis of active infection. Currently, there is one commercially available rapid diagnostic test for yaws that can do just that. However, it is considerably more expensive than the available syphilis tests detecting treponemal antibodies only. We evaluate the cost and cost-effectiveness of a sequential testing strategy (using the treponemal test first, before the combined test), relative to a combined testing strategy (using only the combined test). We consider the two use cases: individual diagnosis and community surveillance. We find that the sequential strategy is cost-saving for both diagnosis and surveillance in most relevant settings. Yaws eradication programme should consider adopting the sequential strategy. Still, a cheaper trep/non-trep RDT is needed, costing no more than US$ 0.50–1.00. Our results will help enhance the cost-effectiveness of yaws programmes in the 13 countries known to be currently endemic. It will also inform efforts in the much larger group of 71 countries with a history of yaws, many of which will have to undertake surveillance to confirm the interruption of transmission.
| Yaws is a non-venereal treponemal infection caused by Treponema pallidum subspecies pertenue affecting primarily the skin in the early stages and the bone and cartilage in the late stages. In 1950, WHO estimated that 160 million people were infected with yaws. Between 2008 and 2012 more than 300 000 new cases were reported to the World Health Organization (WHO). The disease is now targeted by WHO for eradication by 2020. One or two rounds of mass treatment at high levels of population coverage have been shown to reduce prevalence of yaws near to elimination levels.[1] This approach is known as total community treatment (TCT)–treatment of an entire endemic community irrespective of the number of active clinical cases. A second important element of the WHO strategy is 6 monthly Total Targeted Treatment (TTT)–treatment of all active clinical cases and their contacts—to mop-up cases missed in TCT rounds.
Confirmation of clinical cases during TTT programs may be carried out using a rapid diagnostic test (RDT) for the dual detection of treponemal and non-treponemal serological markers at or near to point-of-care. Serological testing is also envisaged for certification of the interruption of transmission of T. p pertenue.
Yaws and syphilis treponemes differ in less than 0.2% of the genome sequence.[2] Yaws is serologically indistinguishable from syphilis, caused by T pallidum subspecies pallidum.[3] Serological tests developed for syphilis may therefore be used to diagnose yaws, especially among children, since its clinical manifestation and epidemiology differ from that of syphilis and may allow a differentiation of the two conditions. Serological diagnosis of clinically active yaws requires the detection of two distinct sets of antibodies: one against treponemal antigens and one against non-treponemal antigens. Treponemal in vitro diagnostics (IVDs), including T. pallidum particle agglutination assay (TPPA), T. pallidum hemagglutination assay (TPHA), and fluorescent treponemal antibody absorption test (FTA-ABS) are highly sensitive and specific but antibodies remain detectable for life following any treponemal infection even after successful treatment. A reactive treponemal test result can therefore indicate either current or past infection and may not be sufficient to indicate no new disease in people with clinical symptoms that look like yaws.
Non-treponemal IVDs, including Rapid Plasma Reagin (RPR) and Venereal Disease Research Laboratory (VDRL) assays, are less specific but since titers rise during active disease and fall following treatment, current and past infection can be distinguished. Titers refer to how many serial dilutions you can perform on the sample and still get a positive result. False positive results can occur when using non-treponemal assays alone due to acute viral infections, malaria, and connective tissue diseases which may also cause non-treponemal assays to be reactive. As a result, testing for yaws requires both treponemal and non-treponemal assays to give an accurate diagnosis of current yaws infection.
The most widely recommended yaws screening tool is the laboratory-based RPR followed by a treponemal test. RPR requires laboratory capacity, trained laboratory personnel, refrigeration for storage of reagents, and electricity to run equipment such as the refrigerator, centrifuge, and shaker. Because such facilities are generally not available in the remote areas where yaws is commonly endemic, diagnosis is often made on the basis of clinical findings only which may not be adequate for surveillance purposes. In places where laboratories are able to do the RPR, serum specimens have to be transported to centralized laboratories for testing and results are available in days or weeks. This delay may result in delayed treatment and continued transmission of the disease.
Rapid syphilis tests detecting treponemal antibodies (treponemal RDTs) are now commercially available, meeting minimum defined standards for quality, safety and performance for use at point-of-care. Treponemal RDTs have been introduced into national antenatal care programmes but these are not commonly used for yaws, as results of treponemal RDTs alone correlate poorly with presence of current infection, as explained earlier.
Currently, one commercially available RDT exists that is based on the simultaneous detection of antibodies to both treponemal and non-treponemal antigens. The DPP Yaws Trep & N.Trep Assay (Chembio, Medford, NY, USA) is designed for use in resource-limited settings where there is limited access to laboratory facilities. For brevity, we refer generically to the assay as a treponemal/non-treponemal RDT or “trep/non-trep RDT”. The dual components of the assay allows clinicians to both screen and confirm the serological status within 15 minutes and allows for differentiation of current and past yaws.
In 2014, the use of trep/non-trep RDT for diagnosis of yaws infection was evaluated and compared with T. pallidum particle hemagglutination assay (TPHA) and RPR as reference standards for treponemal and non-treponemal antibodies detection, respectively.[4] In the low-resource setting of Papua New Guinea, the treponemal test line demonstrated a sensitivity of 88.4% and a specificity of 95.2%; the non-treponemal test line demonstrated a sensitivity of 87.9% and a specificity of 92.5%. A number of evaluations of a trep/non-trep RDT for the diagnosis of yaws infection have now been conducted, as synthesized in a recent meta-analysis.[5]
It is expected that the simpler trep/non-trep RDT should improve access to yaws diagnosis relative to the RPR test. However, use of the trep/non-trep RDT alone may not be the most economical option, especially in low treponemal test positive prevalence settings. In yaws elimination pilot projects, WHO had negotiated a price of US$ 2.50 per trep/non-trep RDT and US$0.45 per treponemal RDT. For surveys where large number of people are non-reactive to the treponemal test, such as in low endemicity settings, a combination of two rapid tests (treponemal RDT for screening, and trep/non-trep RDT for diagnosis) could be cost-saving.
Studies have reported that antenatal syphilis screening and treatment is highly cost-effective in low and middle income countries.[6] Some have modelled the cost-effectiveness of different screening strategies.[7][8][9][10] Terris-Prestholt et al. (2015) were the first to compare the full range of possible screening and treatment strategies for syphilis in multiple countries, including Peru, Tanzania and Zambia. This range included a sequential strategy using a treponemal RDT followed by a dual trep/non-trep RDT. They found that the dual-only strategy was significantly higher cost than the sequential strategy in all three countries, but resulted in more true cases being detected and treated, with the result that cost-effectiveness was about the same in two out of three countries, namely Tanzania and Zambia, where prevalence was highest.
No such economic evaluation of testing strategies has been done for yaws.
We therefore evaluate a two-assay sequential testing strategy in terms of both its cost and cost-effectiveness relative to a single-assay testing strategy. In the sequential strategy, a treponemal RDT is used as the screening assay of the testing strategy, followed by reflex testing with a trep/non-trep RDT for only the reactive treponemal specimens, as depicted in Fig 1. This strategy avoids unnecessary dual treponemal/non-treponemal testing of individuals with no past or current yaws infection (i.e. treponemal negative). The sequential testing strategy is compared to a single-assay testing strategy using the trep/non-trep RDT on the entire testing population.
We aim to establish the conditions (namely, prevalence of past/current infection and relative prices of the treponemal and trep/non-trep tests) under which the sequential strategy would be cost-saving or cost-effective relative to the combined strategy, for the purposes of 1) individual diagnosis and 2) community surveillance. By diagnosis, we mean confirmation of clinically suspected cases in individuals before TCT or during TTT; by surveillance we mean screening of communities (mostly asymptomatic individuals) for the purpose of verification of the interruption of transmission in population after TCT and in countries of historic endemicity.
We use cohort decision analysis to examine the diagnostic and cost outcomes. We estimate cost and cost-effectiveness of the alternative testing strategies in a hypothetical testing population of 1000 people at different levels of past or current prevalence. We place these results in the context of treponemal positive and dually positive prevalences in Ghana, Papua New Guinea, Solomon Islands and Vanuatu—four endemic countries in which population serosurveys were undertaken in the years 2013–2014. These surveys were administered both pre- and post-TCT.
In estimating costs, we take the perspective of the global yaws eradication programme and national health systems. We include the cost of commodities to be funded in large part by the global yaws eradication programme, and the cost of other inputs such as labor to be supplied by the national health system.
We apply a unit cost of US$ 2.50 for each trep/non-trep RDT. For the sequential strategy, we apply a unit cost of US$ 0.45 for each treponemal RDT, and US$ 2.50 for each trep/non-trep RDT. We also add the cost of alcohol swabs ($3 for 100), sterile lancets ($375 for 2000) and non-sterile gloves ($3 for 50 pairs). These prices are consistent with the UNICEF supply catalogue.[11] These ancillary costs increase the unit cost of each trep/non-trep RDT and treponemal RDT to US$ 2.78 and US$ 0.73 respectively.
We consider that every test requires 2–5 minutes of a district-level laboratory technician’s time (depending on experience, this is the time it takes to collect the sample, execute the test, read and report the result). It takes 10–15 minutes between execution of the test and reading of its results, but technicians can attend to other patients during that time. We asked national yaws eradication programmes to provide estimates of the wage of a district-level laboratory technician (in US$). It ranged from US$ 210–510 per month in 11 of the 13 endemic countries, and US$ 1500–1585 per month in two small island developing states (Solomon Islands and Vanuatu).
In the sequential strategy, the trep/non-trep RDT is applied only to treponemal test positives (true and false positives). Total costs (and savings) therefore depend not only on the unit costs described above, but on the sensitivity and specificity of the treponemal RDT for yaws testing. All else equal, a less sensitive (specific) treponemal RDT will result in a smaller (larger) number of trep /non-trep RDTs required in the sequential testing strategy. We use sensitivity and specificity of the treponemal RDT from the Jafari et al. (2013) metanalysis.[12] Sensitivities and specificities are reported in Table 1. In probabilistic sensitivity analysis (PSA), we use the 95% confidence intervals for the sensitivity and specificity results.
Using the Jafari et al (2013) data, we calculate sensitivity and specificity of the treponemal RDT for two relevant subgroups: clinical syphilis cases to be confirmed at sexually transmitted infection clinics, and (asymptomatic) pregnant women to be screened at ante natal care (ANC) clinics. Unfortunately, the results from Jafari et al. (2013) relate to syphilis testing only—there is no evidence of the performance of the treponemal RDT for yaws testing. Marks et al. (2016) found that the sensitivities of both components of the trep/non-trep RDT were higher in patients with syphilis than in patients with yaws at low titers, but not at high titers.[5] It is possible, if not probable, that the sensitivity of the treponemal RDT may therefore be worse for yaws than for syphilis.
We therefore adjust (downward) the sensitivity of the treponemal RDT by the ratio of the sensitivity of the trep/non-trep RDT for yaws to the sensitivity of the trep/non-trep RDT for syphilis. This adjustment, while crude, allows for the possibility that the sensitivity of the trepenomal RDT could be inferior to that of the treponemal line of the trep/non-trep RDT. In PSA, we allow the sensitivity of the treponemal RDT to vary between this adjusted number and that of the treponemal line of the trep/non-trep RDT. We assume that the specificity of the treponemal RDT for yaws is the same as that of the treponemal line of the trep/non-trep RDT. The hypothetical performance of the treponemal RDT for yaws is reported in Table 1.
We multiply unit costs by the total number of each test required. From total costs, we calculate the cost savings associated with sequential testing strategy. We then calculate the so-called threshold unit cost of the trep/non-trep RDT at which the sequential strategy would no longer be cost-saving, assuming a fixed price for the treponemal RDT. That is, we calculate the unit cost of the trep/non-trep RDT such that:
Cd×P<Ct×P+Cd×P×{Tp×Set+(1−Tp)×(1−Spt)}
And where:
Cd is the unit cost of the dual trep/non-trep RDT, including the price of the assay as well as ancillary costs;
P is the population to be tested;
Ct is the unit cost of the treponemal RDT, including the price of the assay as well as ancillary costs;
Tp is the prevelance of past/current infection in the testing population;
Set is the sensitivity of the treponemal RDT; and
Spt is the specificity of the treponemal RDT.
Simplifying and re-arranging, the sequential strategy is no longer cost-saving when:
Cd<Ct1−{Tp×Set+(1−Tp)×(1−Spt)}
Or:
(Cd−Ct)Cd<Tp×Set+(1−Tp)×(1−Spt)
That is, when the percentage difference in unit cost of the treponemal RDT relative to the trep/non-trep RDT is less than the percentage of cases that will test positive using the treponemal RDT, which includes both true and false positives. This reactivity rate is determined by the treponemal positive prevelance (Tp) and sensitivity (Set) and specificity (Spt) of the treponemal RDT. At current prices of the trep/non-trep and treponemal RDTs, the reactivity rate would have to be more than about 74%. Of course, a low reactivity rate of the treponemal RDT, while leading to cost-savings, may not be cost-effective if it results in fewer correct diagnoses.
We calculate the total number of correct diagnoses for each strategy over the full range of prevalences. Decision trees depicting the possible pathways to correct diagnosis are depicted for both strategies in Figs 2 and 3.
We assume that the percentage of past/current infections that are current is the same in the subset of true past/current infection positives identified by the treponemal RDT as it is in the total population of past/current infections (Fig 3). This assumption is thought to be reasonable; in Marks et al (2016), 74% of TPHA positive people had positive RPR; 75% of people with a positive treponemal RDT had a positive RPR, and 77% had a positive non-trep RDT. We also assume that the prevalence of current infection among false past/current infection negatives is (at most) equal to the prevalence of current infection among past/current infections; in any case, in an eradication programme, the number of false negatives will tend towards zero.
We use sensitivity and specificity of the trep/non-trep RDT from the Marks et al. (2016) meta-analysis. Performance characteristics depend on the use case of the trep/non-trep RDT: yaws diagnosis or yaws surveillance. In confirmation of clinical cases, more people will have high titres (where the test performs better) while in confirmation of the interruption of transmission more people will have low titres (where the test performs less well). We therefore consider performance characteristics for primary and secondary disease, or asymptomatic cases (Table 1). The former is applied to populations with clinical symptoms requiring diagnosis, while the latter is applied to populations requiring surveillance.
We calculate the cost per correct diagnosis (true current infection positive or negative) under each strategy (sequential or combined strategy) and use case (diagnosis or surveillance). However, we also report the cost per true positive diagnosis, considering that true positive and negative diagnoses may not be equivalent in their benefits. In the context of individual diagnosis for eradication, for example, true positive diagnosis may be more important than a true negative diagnosis, at least from the perspective of the health system. The incremental cost of treating a false positive is relatively trivial, even considering the cost attributed to any side effects. There are very few and minor side effects associated with azithromycin and indeed, many collateral benefits for diarrheal and other diseases. From the perspective of patients, however, there may be psychosocial costs associated with false positive results.
We then calculate the incremental cost-effectiveness ratio (ICER) of the higher cost combined strategy, for the range of treponemal and dually positive prevalences over which it is not dominated by the sequential strategy. By not dominated, we mean that while the cost is higher, the number of correct diagnoses is also higher. We present cost savings and cost-effectiveness of the alternative testing strategies in the context of survey population prevalences obtained in four countries: Ghana, Papua New Guinea (PNG), Solomon Islands, and Vanuatu.[13–15]
Pre-TCT survey population prevalences obtained using trep/non-trep RDTs are provided in Supporting Information S1 Table. Treponemal positive prevalence varied from 22% in Vanuatu to 51% in PNG. Among those testing treponemal positive, non-treponemal positives were between 21% in Solomon Islands and 71% in Vanuatu. Out of the total population tested, dually positive prevalences were between 7% in Solomon Islands and 18% in PNG.
Post-TCT survey population prevalences obtained using trep/non-trep RDTs are presented for four countries in Supporting Information S2 Table. The treponemal positive prevalence decreased to between 15% in Ghana and 42% in Solomon Islands. Among those testing treponemal positive, non-treponemal positives were between 5% in Solomon Islands and 49% in Vanuatu. The dually positive prevalence decreased, as a percentage of the population tested, to between 1% in Solomon Islands and 8% in Vanuatu.
We are not in this paper attributing these reductions in prevalence to TCT. We are simply using pre- and post-TCT prevalence as a proxy for the prevalence that one might encounter in community surveillance and individual diagnosis settings, respectively.
Use cases and prevalences of the testing population are not independent. In particular, prevalences will be higher when doing individual diagnosis than when doing community surveillance. We therefore focus on the following plausible ranges of prevalence: for individual diagnosis, current/past infection prevalence of 20–55%, of which 20–75% is currently infected; for community screening, current/past infection prevalence of 15–45%, of which 5–50% are currently infected.
We report best estimates using the median of 1000 simulations and the 95% confidence intervals using the 2.5th and 97.5th centiles. All data analysis and visualization were done using R (Foundation for Statistical Computing, Vienna, Austria).[16] All the necessary code is provided as Supporting Information.
We present results separately for the two use cases: individual diagnosis and community surveillance. The differences in results between use cases are driven by different values of sensitivity and specificity in populations with different clinical presentations (stages of disease).
We visualize results over the full range of possible prevalences of past and/or current infection of the testing populations but focus on the plausible ranges determined by treponemal and dually trep/non-trep positive prevalences in Ghana, Papua New Guinea (PNG), Solomon Islands, and Vanuatu.
At the current price of the treponemal RDT and a high prevalence of past/current infection of 51% (the treponemal positive prevalence in pre-TCT Papua New Guinea), we obtain a threshold unit cost for the trep/non-trep RDT of US$ 1.38 (1.31–1.46), including ancillary costs (i.e. swabs, lancets and gloves) and laboratory technician time, or US$ 1.08 (1.02–1.14) for the price of the assay alone. This is the unit cost below which the sequential strategy would no longer be cost-saving for individual diagnosis in a testing population where about one in two are or have been infected.
More generally, costs savings of the sequential strategy in diagnosing 1000 individuals are presented in Fig 4 (top row) across all scenarios of prevalence. At current prices of the treponemal and trep/non-trep RDTs, the sequential strategy is cost-saving if the prevalence of past/current infection of the testing population is less than 85% (81–90). Within the plausible range of prevalence (20–55%), the savings are US$ 1079 (703–1448) per 1000 people tested. Above 85%, it is the combined strategy that is cost-saving.
The number of correct diagnoses of current infection (true positives and true negatives) is presented in Supporting Information S1 Fig (top two rows). Based on our assumptions about the relative performance of the treponemal RDT for yaws, the number of correct diagnoses is somewhat higher under the combined strategy than under the sequential strategy. However, in the plausible range of prevalences, more than 900 correct diagnoses are made for every 1000 people tested under both strategies. It is only at higher prevalences that differences between the strategies become non-trivial. The number of true current infection positives is presented in Supporting Information S2 Fig.
Given our assumptions about the relative performance of the treponemal RDT, there is a range of prevalences over which a higher cost and (hypothetically) more sensitive combined strategy could be more cost-effective than the sequential strategy. Incremental cost-effectiveness ratios (ICERs; ratio of incremental costs over incremental benefits or incremental cost per correct diagnosis gained) are presented in Fig 5 (top row), across different scenarios of prevalence. At prevalence of past/current infection of 51% and current infection of 18% (again, the trep/non-trep RDT positive prevalences in pre-TCT Papua New Guinea), the ICER is US$ 58 (42–103) per correct diagnosis gained.
At very high prevalence of past/current infection, where it becomes cost-saving, a more sensitive combined strategy may dominate the sequential strategy. At very low prevalence of either past/current infection or current infection, it is specificity that matters more for the number of correct diagnoses, and even a more sensitive combined strategy may be dominated by the sequential strategy. In theory, there is a combination of prevalences (very high prevalence of past/current infection and very low prevalence of current infection) where a more sensitive combined strategy could produce fewer correct diagnoses of current infection (this is the area depicted by a black rectangle in Fig 5). In practice, however, this combination is unlikely.
In Supporting Information S3 Fig (top row), we present the same figure, but using only true positive diagnoses in the denominator of the ICER. Here, the ICER is US$ 38 (32–48) at prevalence of past/current infection of 51% and current infection of 18%. Given our assumptions about the relative performance of the treponemal RDT, the combined strategy is nowhere dominated by the sequential strategy when considering only true positive diagnoses; the combined strategy dominates the sequential strategy wherever it is cost-saving.
The cost-effectiveness plane is presented in Supporting Information S4 Fig (top row), at the lower and upper limits of the plausible range of prevalence: for individual diagnosis, the current/past infection prevalence ranges from 20% (lower limit) to 55% (upper limit), of which 20% (lower limit) or 75% (upper limit) are currently infected. It shows that at the lower limit, the combined testing strategy is less effective in spite of being more costly; at the upper limit it results in somewhere between 20–60 additional correct diagnoses (per 1000 tested) for somewhere between US$ 500–600.
At the current cost of the treponemal RDT and a low prevalence of past/current infection of 15% of the testing population (similar to Ghana post-TCT), we obtain a threshold unit cost for the trep/non-trep RDT of US$ 0.84 (0.81–0.88), including ancillary costs and laboratory technician time, or US$ 0.54 (0.52–0.56) for the price of the assay alone.
At current prices of the treponemal and trep/non-trep RDTs, the sequential strategy is cost-saving in surveillance at all levels of prevalence of past/current infection—see Fig 4 (bottom row). Within the plausible range of prevalence (15–45%), the savings are US$ 1527 (1279–1748) per 1000 population.
The number of correct diagnoses of current infection (true positives and true negatives) is presented in Supporting Information S1 Fig (bottom two rows). Again, based on our assumptions about the relative performance of the treponemal RDT for yaws, the number of correct diagnoses is somewhat higher under the combined strategy than under the sequential strategy. Again, under both strategies, in the plausible range of prevalences, more than 900 correct diagnoses are made for every 1000 people tested.
ICERs are presented in Fig 5 (bottom row). At a prevalence of past/current infection of 42% and prevalence of current infection of 6% (similar to post-TCT Papua New Guinea), the best estimate is US$ 355 per correct diagnosis gained by the combined strategy. However, the low estimate is in an area of the plot where the combined strategy is dominated by the sequential strategy. Again, at very low prevalence of either past/current infection or current infection, it is specificity that matters more for the number of correct diagnoses.
In Supporting Information S3 Fig (bottom row), we present the same figure, but using only true positive diagnoses in the denominator of the ICER. Here, the ICER is US$ 117 (90–155) at a prevalence of past/current infection of 42% and prevalence of current infection of 6%. Given our assumptions about the relative performance of the treponemal RDT, the combined strategy is nowhere dominated by the sequential strategy when considering only true positive diagnoses; but, unlike in the diagnosis use case, the combined strategy is never cost-saving and nowhere dominates the sequential strategy.
The cost-effectiveness plane is presented in Supporting Information S4 Fig (bottom row), again at the lower and upper limits of the plausible range of prevalences: for community surveillance, current/past infection prevalence ranges from 15% (lower) to 45% (upper), of which 5% (lower) to 50% (upper) are currently infected.
In summary, this study finds that, at current prices, a sequential strategy is cost-saving relative to use of a combined strategy for individual diagnosis, at a prevalence of past/current infection less than 85% (81–90); it is cost-saving for community surveillance at a prevalence of less than 100% (i.e. always). The threshold prevalence for community surveillance is so high because when titres are low, the reactivity rate of the treponemal RDT is so low and so few people will need a non-treponemal result.
It turns out that the sequential strategy is no longer cost-saving for individual diagnosis in testing populations with high prevalence of past/current infection (i.e. 51%) when the price of the trep/non-trep RDT is less than US$ 1.08 (1.02–1.14). Likewise, the sequential strategy is no longer cost-saving for community surveillance in populations with low prevalence of past/current infection (i.e. 15%) when the price of the trep/non-trep RDT is less than US$ 0.54 (0.52–0.56).
In the absence of evidence assessing relative performance (sensitivity and specificity), the cost-effectiveness of a hypothetically more sensitive combined strategy is uncertain. However, the conditions under which it might be cost-effective are fairly limited. This finding is true even under fairly pessimistic assumptions about the performance of the treponemal RDT for yaws.
In addition to its relatively high cost, a major limitation of the current trep/non-trep RDT is its reduced sensitivity for low titer yaws, at least in the Solomon Islands where it was tested. Further research is required to determine whether available treponemal RDTs (for syphilis) perform any better for low titer yaws. Reduced sensitivity is likely to be a greater problem when using the test as part of yaws surveillance; a higher sensitivity assay will be needed to confirm interruption of transmission, such as RPR or even polymerase chain reaction (PCR) as PCR positive and trep/non-trep RDT negative cases have been observed. Criteria for eradication of yaws in the Morges strategy of 2012 are: 1) absence of new indigenous cases for 3 consecutive years; 2) absence of evidence of transmission for 3 continuous years measured with sero-surveys among children aged 1–5 years (for example, no young children with RPR sero-reactivity); and 3) negative PCR for Treponema pallidum subspecies pertenue in suspected lesions.[17]
There are several limitations to this study.
Serology does not result in identification of all cases of current yaws where early infection may be seronegative, and seropositive patients could have persisting antibodies after successful treatment. Therefore PCR is now considered the gold standard for the diagnosis of active yaws. The sensitivity and specificity of both the treponemal RDT and trep/non-trep RDT have not been assessed relative to PCR. However, there is no reason to believe that the bias favours the sequential testing strategy, as both the treponemal and trep/non-trep RDTs have been assessed against the same standard.
As described in the methods, treponemal RDTs have not been assessed for yaws, and we have therefore had to infer sensitivity and specificity from test performance for syphilis, as reported by Jafari et al (2013). Performance in syphilis is likely to be better than it is in yaws, as the trep/non-trep RDT also performs better in syphilis than in yaws. Although titres are often higher in syphilis compared with yaws (especially asymptomatic disease), it is unclear why Marks et al (2016) found that trep/non-trep performance was worse for yaws even when controlling for titre. Again, yaws and syphilis treponemes differ in less than 0.2% of the genome sequence.[2] Notwithstanding, that the specificity for yaws will be equal to or lower than that reported for syphilis should possibly be further assessed.
More generally, it should be noted that reported sensitivities and specificities can depend upon contextual factors, at least partially, and therefore the results of the meta-analyses of both Jafari et al (2013) and Marks et al (2016) may not fully reflect test performance in all settings, which underscores the need of interpreting our results in the light of the sensitivity analysis we performed.
Our probabilistic sensitivity analysis was focused on uncertainty around the relative performance of the tests, and to a lesser extent on costs. We assumed that the cost of traded commodities, procured by the global yaws eradication programme from international markets, was deterministic. Furthermore, we had only one estimate per country for the wage of laboratory technicians at the district level. In settings where either the commodity or labor costs are highly uncertain and/or their distribution highly skewed, a more sophisticated analysis of costs could be warranted.
We have not considered the time and other indirect costs incurred by the tested populations. The treponemal RDT produces results after 10 minutes; the trep/non-trep RDT requires 15 min. Under the sequential strategy, therefore, treponemal negatives wait 5 fewer minutes and treponemal positives wait 10 more minutes. Had we taken these costs into account, the results might have been less favourable to the sequential testing strategy in higher treponemal positive settings. Therefore, from a patient’s perspective too, there is a case to be made for negotiating lower prices for the trep/non-trep RDT in settings with a high prevalence of past/current infection.
A cheaper trep/non-trep RDT is needed, costing no more than US$0.50–1.00, depending on the use case.
However, other strategies are theoretically possible. RPR is already available and the centralized execution and availability of results may not be a major problem in some settings. Furthermore, a non-trep point of care RDT (alone, without the treponemal RDT) is technically feasible but has not yet been developed. An alternative strategy could involve the treponemal RDT followed by either RPR or the non-trep RDT. A reverse sequential strategy (non-trep test followed by the trep test) could also be possible.
Of course, these alternative sequential strategies not considered in our analysis would only be cost-saving relative to our original sequential strategy as long as the price of the RPR or non-trep RDT did not exceed the cost of the trep/non-trep RDT. Unfortunately, the cost of RPR, including transport to centralized or even international laboratories, will be prohibitively high in most of the settings in question, and we know of no plans to manufacture a non-trep point of care RDT.
Diagnosis and surveillance are essential to the yaws eradication effort. However, the yaws eradication effort is yet to be funded.[18] There are two situations of particular relevance in which savings could be substantial if the sequential testing strategy was implemented: first, during mass screening campaigns, before and after TCT; second, during final screening campaigns, including verification of the interruption of transmission. Cost savings from the sequential strategy could be reallocated to other essential interventions, such as sensitization to increase treatment coverage.
Our results will help enhance the cost-effectiveness of yaws programmes in the 13 countries known to be currently endemic. It will also inform efforts in the much larger group of 71 countries with a history of yaws, many of which will have to undertake surveillance to confirm the interruption of transmission.
|
10.1371/journal.pntd.0004811 | Concomitant Immunity Induced by Persistent Leishmania major Does Not Preclude Secondary Re-Infection: Implications for Genetic Exchange, Diversity and Vaccination | Many microbes have evolved the ability to co-exist for long periods of time within other species in the absence of overt pathology. Evolutionary biologists have proposed benefits to the microbe from ‘asymptomatic persistent infections’, most commonly invoking increased likelihood of transmission by longer-lived hosts. Typically asymptomatic persistent infections arise from strong containment by the immune system, accompanied by protective immunity; such ‘vaccination’ from overt disease in the presence of a non-sterilizing immune response is termed premunition or concomitant immunity. Here we consider another potential benefit of persistence and concomitant immunity to the parasite: the ‘exclusion’ of competing super-infecting strains, which would favor transmission of the original infecting organism.
To investigate this in the protozoan parasite Leishmania major, a superb model for the study of asymptomatic persistence, we used isogenic lines of comparable virulence bearing independent selectable markers. One was then used to infect genetically resistant mice, yielding infections which healed and progressed to asymptomatic persistent infection; these mice were then super-infected with the second marked line. As anticipated, super-infection yielded minimal pathology, showing that protective immunity against disease pathology had been established. The relative abundance of the primary and super-infecting secondary parasites was then assessed by plating on selective media. The data show clearly that super-infecting parasites were able to colonize the immune host effectively, achieving numbers comparable to and sometimes greater than that of the primary parasite.
We conclude that induction of protective immunity does not guarantee the Leishmania parasite exclusive occupation of the infected host. This finding has important consequences to the maintenance and generation of parasite diversity in the natural Leishmania infectious cycle alternating between mammalian and sand fly hosts.
| Transmission is an essential aspect in the life cycle of obligate pathogens, and the ability of a pathogen to be transmitted while simultaneously limiting the chances for other pathogens of the same species could provide an important selective advantage. One mechanism whereby a pathogen could accomplish this is by co-opting its host’s immune response to prevent super-infecting pathogens from becoming established, thus effectively gaining exclusive transmission rights from that host. Several pathogen species have evolved the ability to persist indefinitely within their hosts despite the host’s acquisition of protective immunity against subsequent infection by that pathogen, a condition known as concomitant immunity. We asked whether ‘exclusivity’ was a force underlying the evolution of concomitant immunity using the protozoan parasite Leishmania major as a model. Using genetically marked parasite lines derived from the same Leishmania strain, we show that prior infection does not prevent the entry and survival of subsequently infecting parasites, even though the pathology induced by the secondary infections was markedly reduced. Thus, while persistent Leishmania parasites may vaccinate their hosts against pathology, they do not confer protection against superinfection. This finding has important consequences to the maintenance and generation of Leishmania genetic diversity, especially through sexual processes.
| Persistent host/pathogen relationships are often characterized by a ‘stalemate’ in which the host neither succumbs to disease nor is able to completely achieve sterile cure. Persistent infections can show varying degrees of pathology, ranging from chronic overt disease to asymptomatic infections, reflecting different mechanisms of disease tolerance [1,2,3,4,5,6]. For asymptomatic persistent infections, often a key component is a strong immune response on the part of the host, which is required to keep pathogen numbers in check. In some cases, this immune response also serves to protect against pathology resulting from subsequent re-infection by the same pathogen, a process known as premunition or concomitant immunity [7,8,9].
Long-term host/pathogen relationships carry benefits and risks to both partners, and have been the subject of considerable study from an evolutionary perspective [10,11,12]. In the case of concomitant immunity, the host benefits by its immune system’s ability to control the infection and minimize pathology, as well as protection from disease arising from new infections. However, this comes at the cost of increased risk of disease reactivation, typically following immunosuppression or stress [1,4,13,14,15]. From the pathogen’s perspective, while concomitant immunity decreases microbial numbers, it may improve the likelihood of transmission due to the increased longevity of the infected host.
A second potential benefit to the pathogen of concomitant immunity is ‘exclusivity’, in that the pathogen may use its host’s immune response to gain a competitive advantage by reducing the invasion of the host by other strains or species. For Schistosoma mansoni, concomitant immunity may limit intraspecific competition for limited resources [7,16]. This question has been less studied in microbes, where potentially, concomitant immunity could completely or partially preclude secondary colonization of the infected host, and thereby favor transmission of the primary infecting strain. In some respects concomitant immunity might act as a barrier to superinfection in a manner analogous to the mechanisms employed by lysogenic bacteriophages which generally render their bacterial host resistant to super-infection with closely related phage [17]
The protozoan parasite Leishmania major provides an excellent model for investigating forces of concomitant immunity and persistence. L. major is transmitted to mammalian hosts by the bite of phlebotomine sand flies, and in laboratory mice a range of pathology ensues depending on both the particular parasite and mouse strain [18]. Infections of genetically susceptible mice (such as BALB/c) with most L. major strains yields a progressive and fatal infection [18]. In contrast, infection of genetically resistant mice (such as C57BL/6) initially gives rise to a progressive parasitemia and lesion pathology at the site of inoculation similar to that seen in BALB/c mice, but after 4–6 weeks an immune response develops which controls both parasitemia and pathology [18,19]. Notably, the healed mice are effectively vaccinated and resistant to disease pathology from subsequent infections. Following healing, and for the remainder of the host’s life, a small number of parasites often persists in the skin at the site of inoculation and in the regional lymph node draining that site [20]. In keeping with concomitant immunity/premonition paradigm, these persistent parasites appear to be important for the maintenance of an anti-Leishmania immune response, as treatment resulting in sterile cure is associated with the loss of immunity [21,22]. Indeed, the strong protective immunity induced by persistent Leishmania is the basis for the ancient practice of leishmanization, in which live, virulent parasites are intentionally inoculated in inconspicuous sites of the body to protect against natural infection and pathology at other sites [23]. However, persistent Leishmania are likewise the source for reactivation following immunosuppression [14,15].
Asymptomatic persistent Leishmania infections of C57BL/6 mice fit several criteria relevant to understanding of the benefits and tradeoffs of concomitant immunity. The animals are healthy, and despite the small numbers (< 1000 / mouse), persistent parasites can be efficiently transmitted to sand flies [24,25,26]. Several previous studies exploring the immune response induced by persistent parasites inoculated L. major into a primary site, waited for the lesion pathology to resolve, and inoculated a challenge at a secondary site [22,27,28,29,30]. Each time, viable parasites were recovered from the secondary site, the assumption being that these arose from the secondary challenge. However, L. major is known to traffic to sites distant from the site of inoculation [20]. Thus, parasites isolated at the secondary inoculation site may have actually originated from the primary infection, perhaps accentuated by the transient reactivation of parasites at the primary infection site as reported by Mendes et al [27].
To unambiguously establish the question of secondary colonization and exclusivity, we generated parasites derived from the same strain of L. major of comparable virulence but bearing independent drug resistance markers (PHLEO/phleomycin and SAT/nourseothricin). These were then used in the classic infection/challenge persistence model, using one strain as the primary infection, which gave rise to the expected lesion/healing/persistence phenomenon, followed by injection with the second strain in the opposite foot. The results show clearly that under these conditions Leishmania persistence is not accompanied by ‘exclusivity’, in that similar numbers of both ‘primary’ and ‘secondary’ parasites persisted at their respective sites of inoculation. These data suggest that while persistent L. major vaccinates its host from disease pathology, it does not confer exclusivity to the acquisition of secondary infecting Leishmania. This finding has important consequences to the maintenance and generation of Leishmania genetic diversity, including that arising through sexual processes [31,32].
The generation of both the phleomycin resistant parasites (SSU:IR1PHLEO-YFP; referred to here as LmjF-PHLEO) and the nourseothricin-resistant parasites (SSU:SAT-TK-LUC; referred to here as LmjF-LUC-SAT) used in this study was described previously [33,34]. Parasites were grown at 26˚C in M199 medium (US Biologicals) supplemented with 40 mM 4-(2-hydroxyethyl)-1-piperazine-ethanesulfonic acid (HEPES) pH 7.4, 50 μM adenosine, 1 μg ml−1 biotin, 5 μg ml−1 hemin, 2 μg ml−1 biopterin and 10% (v/v) heat-inactivated fetal calf serum [35]. Nourseothricin (Jena Bioscience, Jena, Germany) was used at a concentration of 100 μg/ml and phleomycin (Sigma, St. Louis, MO) was used at a concentration of 20 μg/ml. Infective metacyclic-stage parasites were recovered using the density gradient centrifugation method [36].
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the United States National Institutes of Health. Animal studies were approved by the Animal Studies Committee at Washington University (protocol #20090086) in accordance with the Office of Laboratory Animal Welfare's guidelines and the Association for Assessment and Accreditation of Laboratory Animal Care International.
Female C57Bl/6J mice (Jackson Labs) were injected subcutaneously in a hind footpad with 105 metacyclic stage parasites. Naïve mice (6–8 weeks old) were injected in the left hind footpad. Secondary injections took place in the right hind footpad at a time point >1 month after primary lesions had resolved. Footpad lesion thickness was measured using a Vernier caliper (Mitutoyo). Lesion size was calculated as the difference in thickness between the infected and uninfected footpads. Luciferase activity was determined as described elsewhere [34]. Briefly, mice were given a dose of D-luciferin (150 μg gram-1 body weight; Biosynth) in PBS 10 minutes prior to imaging with an IVIS 100 imaging system (Xenogen Corp). In this study, values less than 105 p/s fall into the background range. Limiting dilution assays were performed as described previously [37], with the addition of phleomycin or nourseothricin as indicated. Reconstruction experiments suggest that the limit of detection was about 14 parasites/footpad.
Data are presented as the arithmetic mean ± the standard deviation. P values were calculated by the Student’s t-test.
We used two L. major Friedlin V1 parasites expressing genes conferring resistance to the antibiotics nourseothricin (SAT) or phleomycin (PHLEO). The nourseothricin resistant parasites also express firefly luciferase, and will be referred to hereafter as “LmjF-LUC-SAT”, while the phleomycin resistant parasites will be referred to as LmjF-PHLEO. To confirm that the LmjF-LUC-SAT and LmjF-PHLEO parasites were of comparable virulence in mice, 105 metacyclic-stage parasites were inoculated into the footpads of naïve C57BL/6 mice (5 mice/group), and the lesion pathology was monitored over time (Fig 1A). Both lines exhibited disease progression typical of untransfected L. major / C57BL/6 infections, with lesions developing between 10–17 days post infection and reaching their maximum (~1.4 mm increased footpad thickness) around 30 days post infection [38,39]. Thereafter the lesions declined, and were completely resolved by 130 days post-infection (Fig 1A). While there was some tendency for the LmjF-LUC-SAT line to show smaller lesion sizes, at no point was this difference statistically significant. After resolution, mice were sacrificed and the parasite titers in the infected feet were enumerated by limiting dilution analysis (Fig 1B). The number of persistent parasites recovered for both lines was in agreement with what is expected in this experimental system (typically 100–1000 parasites, with substantial variability amongst mice and experiments) [20,27,28,30,40]. Importantly, we found no significant difference in the number of persistent parasites between the two lines, with LmjF-LUC-SAT and Lmj-PHLEO showing a similar range (Fig 1B) and mean (25 and 32 parasites / foot; P > 0.45 by Student’s t-test). We judged these lines to be of comparable virulence and suitable for subsequent experiments.
Two experiments were performed in which naïve mice (4–5 mice per experiment) were inoculated with 105 purified metacyclic-stage LmjF-LUC-SAT parasites in the left hind footpad primary infection site. A lesion formed at that site and resolved in accordance with the data shown in Fig 1A. At a time point >1 month after resolution (1.5 and 3 months for experiments 1 and 2 respectively), 105 metacyclic LmjF-PHLEO parasites were inoculated into the right hind footpad secondary infection site. Footpad swelling of both the primary (L) and secondary (R) injection sites was then measured over time. We also used in vivo imaging of luciferase activity to visualize LmjF-LUC-SAT parasites, as a second probe of whether transient reactivation of primary parasites occurred [27].
As expected, in both experiments the mice showed good protection, as evidenced by a reduction in lesion pathology at the secondary ‘challenge’ site. Although with some variation, in both experiments the lesions generated by the secondary LmjF-PHLEO parasites were significantly smaller and resolved more rapidly than those in naïve mice (Fig 2A). We saw no evidence of reactivation of the “primary” LmjF-LUC-SAT parasite, as judged by either lesion measurement (Fig 2B) or in vivo imaging of parasite luciferase (Fig 2C, left), the latter yielding values in the background range, and orders of magnitude less than what is seen following infection of naïve mice by these parasites at the peak of parasitemia (Fig 2C, right).
Having established the classic Leishmania paradigm of vaccination following resolution of a primary challenge for the genetically marked lines in our study, we then measured the occurrence of both the primary- and secondary- infecting parasites, in both infection sites. This was performed by limiting dilution assays at day 87 (experiment 1) or day 139 (experiment 2) post-infection. Total parasites were assessed by growth in the absence of drug, while LmjF-LUC-SAT (primary) was estimated from growth in media containing nourseothricin and LmjF-PHLEO (secondary) from growth in media containing phleomycin. The results from individual mice from both experiments as well as the global averages are shown in Fig 3.
Parasites were recovered from all primary infection sites, ranging from 14 to 504 parasites/foot, with an average of 282 ± 158 parasites recovered per foot (N = 9). These parasites were exclusively the primary LmjF-LUC-SAT parasite, as they were unable to grow in the presence of phleomycin. In one animal parasites expressing the SAT marker were apparently lost; similar results have been reported in L. tarentolae and attributed to the genetic plasticity of the ribosomal RNA locus [41], and we have seen this occasionally in other experiments in L. major.
Parasites were also recovered from the secondary infection site from 8 of the 9 mice, ranging from 14 to 785 parasites/foot, with an average of 119 ± 156 parasites/foot. Importantly, nearly all of the parasites recovered from the secondary infection site were the LmjF-PHLEO parasite inoculated there (99 ± 3%). In only one mouse (#2–5) was colonization of the secondary site by ‘primary’ infection site LmjF-LUC-SAT parasites found, suggesting that metastasis of parasites from the primary to the secondary sites occurs infrequently. Importantly, the numbers of ‘primary’ infection site LmjF-LUC-SAT parasites were not significantly different from that seen for the ‘secondary’ infection site LMjF-PHLEO parasites (P > 0.08, Student’s T-test). These data show that despite successful ‘vaccination’, as defined by reduction in lesion pathology, this immunity was not ‘sterilizing’ against secondary infection and did not preclude efficient colonization of the infected mouse significantly.
A number of factors have been proposed to contribute to the maintenance of pathogens for long periods of time in the host, including an insufficient immune response and the benefits accruing to the pathogen from residing within a longer-lived host thereby increasing the likelihood of transmission [10,12]. In many cases this relationship has progressed to the point where the pathogen infection is asymptomatic, thereby fulfilling the evolutionary dictum that a ‘successful pathogen does not kill its host too quickly”. Often this asymptomatic persistence is accompanied by protection from disease induced by further infections of the same or related pathogens, a process termed concomitant immunity [7]. Such a relationship provides benefits to both the pathogen and the host through increased longevity of the latter (albeit with some risk of reactivation), and increased transmission of the former.
Leishmania provides an attractive system for the study of concomitant immunity [20,21,27,42,43,44] and here we have used this to consider another potential benefit to the pathogen, one of ‘exclusivity’. Exclusivity would favor transmission of the primary infecting pathogen due to reduction in the ability of secondary infecting parasites to becoming established in a previously infected host. However, our data show clearly that despite induction of a protective immune response able to mitigate disease pathology (Fig 2), secondary Leishmania major infections are nonetheless able to establish themselves effectively in a previously infected host (Fig 3). While this result may have been anticipated from prior studies [22,28], this is the first time this has been established rigorously for Leishmania using genetically marked parasites able to distinguish primary from secondary infections and bioluminescent imaging to assess reactivation. Our studies also provide limited support for the prior assumption that in general parasites are not frequently transferred from the primary to the secondary site of infection, although we did observe transfer in one mouse (Fig 3, mouse #2–5).
Consistent with prior studies, the immunity generated by persistent parasites was not always sterilizing and the average number of “secondary” parasites was not statistically different from that of the “primary” parasites (Fig 3). Nonetheless, the average number of parasites recovered from the secondary site was about 2-fold less than from the primary site, similar to the findings of Mendez et al (2004) [27]. Thus, it is possible that secondary infecting parasites may experience a modest quantitative disadvantage, which over evolutionary time could provide a strong positive selective force on the parasite favoring the induction of concomitant immunity. This phenomenon may warrant further study in the future.
In our studies an inoculum of 105 purified metacyclic parasites was used. While most sand flies transmit less than 600 parasites to mice, some transmit up to 105 [25]. Thus, the infecting dose used here falls on the high side of the biologically relevant range. Studies using low-dose infections with 100 metacyclics also recovered parasites from the site of secondary infection, although genetic markers were not available to confirm their identity [22,27,28,29]. Undoubtedly there are a number of experimental variables that could be pursued in future studies, including infecting dose, the relative timing of the primary and challenging infections, or sites of inoculation other than the footpad that may be potentially relevant, such as the ear, snout or tail. Another important variable is the extent of genetic identity between the primary and secondary infections; we purposefully chose to study isogenic parasite lines here to maximize the likely efficacy of concomitant immunity, the efficacy of which might be expected to decrease with heterologous strains or event species. Lastly, while our experiments were carried out in an ‘orderly’ manner with primary and secondary infections in separate feet, nature is decidedly less so, and indeed infections may occur at the same location [45], thereby increasing the likelihood of transmissible mixed infections.
An important question is the relevance of ‘needle’ infections performed here to natural sand fly transmission, where parasites are deposited along with immunomodulatory factors of both sand fly and parasite origin. These factors include saliva and secreted parasite molecules such as proteophosphoglycan, both of which typically act to facilitate primary infections [46,47,48,49] but which can also engender various protective responses [50,51] and thus have the potential to either favor or hinder the entry of the secondary ‘invading’ Leishmania. In several studies examining challenge by sand fly bite of mice which had healed from primary infections, sterilizing immunity was seen in 33/64 mice tested in challenge infections (52%), while the remainder showed minimal pathology accompanied by parasite numbers ranging from 100 to 10,000 at the challenge bite site [48,52]. Assuming that these parasites arise primarily from the challenge parasite (as shown here), both natural sand fly and ‘needle’ challenge can yield infections with robust parasite survival at the secondary challenge site at significant frequencies.
That concomitant immunity induced by primary L. major infections protects against pathology can occur at significant frequencies without sterilization, instead leading to ‘mixed’ infections of the host, has important implications for the generation and maintenance of Leishmania diversity. In regions where Leishmania is endemic, mammalian hosts are likely subjected to many bites by infected sand flies [53,54], which over time could result in the host being persistently infected with several genetically distinct parasite lines. There are numerous reports documenting the recovery from infected animals and humans of Leishmania stabilates exhibiting mixed genotypes, using a variety of molecular taxonomic methods [55]. Some fraction of these represent true mixed infections, while others may arise from the presence of intra- or inter-specific hybrids [31,55,56,57,58,59,60]. In several studies the incidence of mixed populations exceeded 10% [54,61,62,63]. Moreover, concerns have been raised about the efficiency of detection of mixed infections, ranging from technical analysis to problems associated with differential outgrowth during adaptation to culture [62,63,64], suggesting that the true incidence may be greater than presently appreciated. Notably, human infections showing overt pathology have been most highly sampled, and even for humans the situation in the more prevalent ‘asymptomic’ infections (primary or secondary) is largely unknown. Thus while it is difficult to say with any certainty what fraction of natural Leishmania infections are truly ‘mixed’ in human or animal reservoir populations, they are far from rare, and potentially quite common.
Once established, mixed infections have the potential to be passed on to sand flies, which have recently been shown to be the site of both intra-specific and interspecific genetic exchange [32,65,66,67,68]. Since the frequency of sand flies bearing Leishmania in natural populations is relatively low (often just a few percent) [69,70,71], the accumulation and maintenance of mixed populations over time in persistent mammalian infections would act to increase the frequency at which sand flies acquire mixed infections, which thereafter undergo genetic exchange and generate diversity. While genetic exchange occurs relatively infrequently on a per Leishmania cell basis (<10−4; [32]), Leishmania numbers in sand flies are sufficient to yield hybrid parasites at high frequencies (25% or greater per fly; [32,65]. Thus, the lack of ‘exclusivity’ even in the presence of protection against disease pathology may result in increased opportunities for genetic exchange and the emergence of new disease phenotypes in nature [72].
Our data also have some consequences to vaccination strategies. Currently the ‘healed’ mouse is considered a ‘gold standard’ for the maintenance of effective immunity against disease pathology, and the generation of live-attenuated parasite lines that persist without pathology while immunizing against virulent challenge has been a priority in vaccine research [40,73]. Our data suggest that such an approach would likely allow virulent parasites from subsequent natural infections to establish their own persistent infections, which could then pose a risk of reactivation and/or transmission. This may provide further impetus for the development of vaccines conferring sterilizing, long-lasting protection against both pathology and parasitemia.
|
10.1371/journal.pntd.0002391 | Differential Epidemiology of Salmonella Typhi and Paratyphi A in Kathmandu, Nepal: A Matched Case Control Investigation in a Highly Endemic Enteric Fever Setting | Enteric fever, a systemic infection caused by the bacteria Salmonella Typhi and Salmonella Paratyphi A, is endemic in Kathmandu, Nepal. Previous work identified proximity to poor quality water sources as a community-level risk for infection. Here, we sought to examine individual-level risk factors related to hygiene and sanitation to improve our understanding of the epidemiology of enteric fever in this setting.
A matched case-control analysis was performed through enrollment of 103 blood culture positive enteric fever patients and 294 afebrile community-based age and gender-matched controls. A detailed questionnaire was administered to both cases and controls and the association between enteric fever infection and potential exposures were examined through conditional logistic regression. Several behavioral practices were identified as protective against infection with enteric fever, including water storage and hygienic habits. Additionally, we found that exposures related to poor water and socioeconomic status are more influential in the risk of infection with S. Typhi, whereas food consumption habits and migration play more of a role in risk of S. Paratyphi A infection.
Our work suggests that S. Typhi and S. Paratyphi A follow different routes of infection in this highly endemic setting and that sustained exposure to both serovars probably leads to the development of passive immunity. In the absence of a polyvalent vaccine against S. Typhi and S. Paratyphi A, we advocate better systems for water treatment and storage, improvements in the quality of street food, and vaccination with currently available S. Typhi vaccines.
| Enteric fever, caused by ingestion of bacteria Salmonella Typhi or Salmonella Paratyphi A, is common in regions with poor water quality and sanitation. We sought to identify individual-level risks for infection in Kathmandu, Nepal, a region endemic for enteric fever. In this study, we enrolled patients presenting to hospital who were blood-culture positive for enteric fever and a series of community controls matched for age, gender and residential ward. Our findings suggest that while some risks for infection with S. Typhi and S. Paratyphi A overlap, these organisms also have distinctive routes of infection in this setting; poor water and socioeconomic status seemed more influential in infection with S. Typhi, whereas food consumption habits and migratory status were shown to play a larger role in infection with S. Paratyphi A. Additionally, serological evaluation of IgG levels against the Vi (Salmonella Typhi) and the O:2 (Salmonella Paratyphi A) antigens demonstrated high titers against both antigens throughout life, suggesting frequent and constant exposure to these organisms in Kathmandu. As major improvements in sanitation infrastructure are unlikely in this setting, we recommend water treatment and storage-based prevention strategies, as well as street food quality regulation, and the promotion of vaccination with existing typhoid vaccines.
| The human systemic disease enteric fever is most commonly caused by the Salmonella enterica serovars Typhi (S. Typhi) and Paratyphi A (S. Paratyphi A) [1], [2]. The disease is found in areas with poor sanitation and hygiene [3], and has an estimated global burden of 27 million new cases and 200,000 deaths annually [4]. The causative bacteria are transmitted fecal-orally. After ingestion, a 7 to 14 day symptomatic period ensues whereupon bacteremia presents as a persistent non-focal fever with malaise. The bacteria can induce a protracted illness that lasts several weeks, and while rarely fatal, the disease can result in life threatening complications including hypotensive shock and intestinal perforation [2], [5].
Enteric fever is endemic in Nepal and S. Typhi and S. Paratyphi A are the most commonly isolated organisms from the blood of febrile patients in our Kathmandu-based healthcare setting [6], [7]. A retrospective analysis highlighted a substantial burden of enteric fever within the local population, particularly in school-age children and males aged 15 to 25 years [8]. Furthermore, we have shown that indirect transmission through contaminated drinking water may play a more important role in maintaining the endemicity of infection in Kathmandu than close contact with symptomatic or asymptomatic individuals [9], [10]. However, there are several gaps in our knowledge of the transmission of enteric fever and specific behavioral risk factors for infection in Kathmandu have not been identified.
Several case-control studies have investigated risks for enteric fever; the majority implicate water and food as important transmission routes [11]–[21]. Additional risk factors include previous contact with an enteric fever case [12], [19], [22], recent antimicrobial treatment [15], local flooding [19], poor hygiene [13], [14] and poor socioeconomic status [13], [19]. The identification of tractable risk factors and probable routes of transmission for the agents of enteric fever are necessary for the development of targeted interventions to reduce disease burden. In this study, a case-control investigation and age-stratified serology were performed to identify risk factors for, and measure expose to, enteric fever in Kathmandu.
This study was approved by the institutional ethical review boards of Patan Hospital and The Nepal Health Research Council. All enrollees were required to provide written informed consent for the collection and storage of all samples and subsequent data analysis. In the case of those under 18 years of age, a parent or guardian was asked to provide written informed consent.
Patan Hospital is a 318-bed government hospital providing emergency and elective outpatient and inpatient services located in Lalitpur Sub-metropolitan City (LSMC) within the Kathmandu Valley (Figure 1). Enteric fever is common at the outpatient clinic at Patan Hospital, which has approximately 200,000 outpatient visits annually. The population of LSMC is generally poor, with most living in overcrowded conditions and obtaining their water from stone spouts or sunken wells (Figure 1).
All febrile patients attending the outpatient or emergency department between April and October 2011 with a non-focal fever lasting 3 or more days, aged between 2–65 years and providing informed consent were eligible for this study. All individuals received a blood culture and only those with a blood culture positive for S. Typhi or S. Paratyphi A were enrolled.
Community-based controls were matched for age, sex and residential ward and enrolled at a ratio of 3∶1. Controls were identified in households neighboring cases. If the case lived in a stand-alone house, the household to the right of the “case-household” was approached by a community medical assistant (CMA) within 2 weeks of the case enrollment. If this control refused, the household to the left was approached, followed by the house parallel across the street. If the case lived in multi-story building, the household above the “case-household” was approached. If this control refused, a household a storey either below or two stories below was approached. Controls were required to be within 5 years of the age of the case and must not have had fever, gastrointestinal disturbances or history of enteric fever in the month before administration of the questionnaire. If an approached control failed to meet the enrollment criteria or refused participation, the house to the left or the storey below the “case-household” was approached for enrollment. The trained CMAs administered a 129-question questionnaire to each enrolled individual.
Anti-coagulated blood samples were collected from all febrile patients upon arrival in the outpatient department. For those over the age of 12 years, 10 ml of blood sample was collected; 5 ml was collected from those aged 12 years or less. The blood samples were inoculated into tryptone soya broth and sodium polyethanol sulphonate up to 50 ml. The inoculated media was incubated at 37°C and examined daily for bacterial growth over seven days. On observation of turbidity, the media was sub-cultured onto MacConkey agar. Any bacterial growth presumptive of S. Typhi or S. Paratyphi was identified using serotype specific antisera (Murex Biotech, Dartford, UK).
ELISAs to measure IgG against the Vi and O:2 antigens (NVGH, Siena, Italy) [23], [24] were performed on 795 plasma samples that were age-stratified and randomly selected from a serum bank comprised from the blood of patients attending Emergency Department of the Patan Hospital for reasons other than typhoid treatment or those relating to a febrile illness between January 2009–December 2011. Ages of enrollees ranged from 0–65 and were resident in the same demographic area as the case/control enrollees. All plasma samples were subjected to ELISA detecting IgG antibodies to both Vi and O:2. Briefly, plasma was diluted into 1∶200 and aliquoted into ELISA plates (Nunc, Sigma-Aldrich Co, UK) independently coated with Vi and O:2 antigens. After washing, bound IgG was detected using an alkaline phosphatase–conjugate antiserum (Sigma-Aldrich Co, UK). Antibody levels were quantified using standard curves. The standard curve for Vi-antibody was created using a prepared anti-human Vi-antibody standard. The standard curve for O:2-antibody was created using a pool of plasma from S. Paratyphi A confirmed patients who had high levels of antibody to O:2, which had been screened prior to the serum bank. The cutoff value of the ELISAs was defined as the optical density of blank control wells plus two standard deviations.
Data were imported into STATA v9.2 (College Station, TX, USA). The association between the outcome of enteric fever (defined as infection with either S. Typhi or S. Paratyphi A) and each exposure of interest was examined using a matched univariate analysis through conditional logistic regression. A conceptual model was generated to develop a biologically plausible set of covariates a-priori that were thought likely to influence the outcome, including those related to poor socioeconomic status and poor water quality. Variables were included in a matched multivariate analysis through a-priori selection or were associated (p<0.25) with enteric fever in the univariate analysis [25]. Model fit was assessed through log likelihood and relative AIC value. Co-linearity among variables was assessed but no strong associations were found between the variables included in the final model. The presence of effect modification was evaluated between each of the salient exposures and confounders of interest through the χ2 test for homogeneity. Several sets of interactions were found to be significant. Each interaction term was included in the final model of interest and model fit was assessed. However, due to small numbers of patients in many of the strata, inclusions of these interactions led to an unstable estimate and were thus discarded. One interaction, however, household size affecting the risk of water storage on outcome of enteric fever, improved model fit and led to stable estimates so was included. The categorical variable household size was generated through assessing whether the house was the same size or smaller than the median number of people (12) in this dataset or larger than the median. All P values are two-sided.
During the period of investigation 103 febrile patients with culture confirmed enteric fever were enrolled, 48% (49/103) were positive for S. Typhi and 52% (54/103) were positive for S. Paratyphi A. Baseline characteristics of the enteric fever cases are described in Table 1. Briefly, those with culture confirmed enteric fever were more often male (64%; 66/103) and had a median age of 18 years (interquartile range (IQR) 10–23 years). Males and females with S. Paratyphi A did not differ significantly in age (median: 18 years, IQR: 5–55 and 20 years, IQR: 6–28, respectively), but female S. Typhi cases were significantly older (median: 21.5 years, IQR: 8–50) than male S. Typhi cases (median: 16 years, IQR: 7–32) (P = 0.03, Mann-Whitney U test). A total of 294 controls were enrolled, leading to a final case-control ratio of 1∶2.85.
The clinical presentations of S. Typhi and S. Paratyphi A were largely indistinguishable, with most cases exhibiting a progressive fever (77%; 79/103), nausea (50%; 51/103), and a limited number having an abdominal rash (2%; 2/103) or constipation (6%; 6/103). Patients with S. Typhi were more likely to present with abdominal pain (61%; 30/49) than those with S. Paratyphi A (33%; 18/54) (P = 0.005, χ2 test) and to have diarrhea (25%; 12/49 and 9%; 5/54, respectively (P = 0.038, χ2 test)).
To identify important associations between various exposures and the outcome of enteric fever due to either S. Typhi or S. Paratyphi A, we performed a series of univariate analyses and, to control for confounding, built a multivariate model. As shown in Table 2, a variety of exposures were found to be protective against enteric fever in a univariate analysis. These protective variables included, an awareness of enteric fever, reporting a previous enteric fever episode, recent contact with an enteric fever case, using a metal cover on a household water storage container, and the consumption of pani puri. The counterintuitive protective effects of recent contact with an enteric fever case and consumption of pani puri were considered to be a result of study design. As cases and controls were geographically matched, it is likely that controls were aware of local cases. Additionally, as pani puri has the potential to be fecally contaminated [26], the protective association is likely to be explained by uncollected information. These exposures were not included in the final model. Enteric fever risks from the univariate analysis included a household monthly income of <$125, the duration of residence in Kathmandu, the use of stone spout water, storing water in the household, the use of a household latrine, and the recent consumption of street food. From the multivariate model, an awareness of enteric fever (AOR: 0.25, 95%CI: 0.1–0.5, p<0.001) and the use of a metal cover on household water storage (AOR: 0.28, 95%CI: 0.1–0.7, p = 0.006) remained significantly protective against infection. The use of a household latrine, compared to a community latrine (AOR: 4.10, 95%CI: 1.4–12.4, p = 0.013), as well as recent street food consumption (AOR: 2.85, 95%CI: 1.4–6.0, p = 0.006) remained as strong risk factors for infection. Notably, the risk of storing water was found to vary by household size. Those living in households with more than 12 people were at significant risk of infection if they stored water (AOR: 9.35, 95%CI: 1.7–51.0, p = 0.010), while those who lived in smaller houses were unaffected by water storage habits.
From the series of univariate analyses performed on S. Paratyphi A cases and matched controls only, many of the identified protective and risk factors, as shown in Table 3, were similar to the overall enteric fever analysis shown in Table 2. However, the final multivariate model demonstrated that the use of a metal cover for water storage (AOR: 0.19, 95%CI: 0.1–0.7, p = 0.014) remained strongly protective against infection with S. Paratyphi A, whereas the consumption of street food within the two weeks preceding illness remained a significant risk factor (AOR: 2.95; 95%CI: 1.1–7.8, p = 0.028). Furthermore, residing in Kathmandu for less than two years (AOR: 2.29, 95%CI: 0.9–5.7, p = 0.077) remained a weakly significant risk factor for S. Paratyphi A infection only.
Many of the important protective and risk factors for S. Typhi infection only were again similar to the overall enteric fever analysis, however some S. Typhi-specific exposures emerged. Firstly, the use of stone spout water and household water storage were strong risk factors for S. Typhi infection in the univariate analysis, although the use of stone spout water was only mildly significant in the multivariate model (AOR: 4.17, 95%CI: 0.9–20.5, p = 0.078). Water storage was a risk but was not significantly associated with S. Typhi infection in the multivariate model (AOR: 2.56, 95%CI: 0.8–8.6, p = 0.129). Living in a house with more than 12 people was also a risk that did not remain strongly significant in the multivariate model (AOR: 2.11, 95%CI: 0.8–5.5, p = 0.127), although reporting a monthly income of less than $125 was strongly and independently associated with S. Typhi infection (AOR: 5.21, 95%CI: 1.5–18.4, p = 0.010). Awareness of enteric fever was strongly protective, specifically against infection with S. Typhi (AOR: 0.28, 95%CI: 0.1–0.8, p = 0.014).
To assess exposure to S. Typhi and S. Paratyphi A in the local population, we measured IgG against Vi antigen (S. Typhi) and O:2 antigen (S. Paratyphi A) in 795 age-stratified (0–65 years) serum samples derived from the same population as our case/control enrollees (Figure 2). The resulting data demonstrated a consistently high level of IgG against Vi and O:2 in all age groups. IgG against S. Typhi was highest at birth and then declined with a secondary peak at the age of 17–18 years. Antibody against S. Paratyphi A was lowest at birth and peaked at the age of 11–12 years and subsequently declined; yet persisted into old age. There was a weak correlation between levels of IgG to Vi and to O:2 (Spearman's ρ: 0.30, P<0. 001), this was most apparent within the group aged 11–20 years (Spearman's ρ: 0.50, P<0.001) (Figure 2).
The aim of this study was to elucidate risk factors for, and protective behavior against, S. Typhi and S. Paratyphi A infection in Kathmandu. These analyses were performed in order to more clearly define the epidemiology of enteric fever for developing appropriately targeted control measures. We identified several risk factors and protective variables that could be targets for future intervention studies; in addition, our data show some differences in epidemiology between the two enteric fever serovars and highlight continued exposure and enteric fever risk as a function of daily life in this location.
Compared to S. Paratyphi A, enteric fever due to S. Typhi is historically thought to be more common, have a more severe clinical course, and result in more frequent and severe sequelae [27]. However, recent studies have suggested that infections caused by S. Paratyphi A are now more prevalent in areas endemic for enteric fever, and that infections caused by S. Typhi and S. Paratyphi A are clinically indistinguishable [8], [19], [28]–[33]. Our data in the present study support this trend. Whether this increase in S. Paratyphi A is a consequence of the decline of enteric fever due to S. Typhi or due to an absolute increase in the incidence of S. Paratyphi A is still not clear [19], [34], [35]. The relative increase in S. Paratyphi A has important implications for public health efforts to control the burden of disease. Particularly, the oral Ty21a and the parenteral Vi typhoid vaccines offer limited or no protection, respectively, against S. Paratyphi A [29], [36]. From our serological data, we suggest that vaccination with the currently available vaccines in many age groups may not dramatically impact the rates of disease in this population due to sustained exposure and high levels of pre-existing antibody. The dawn of Vi and O:2 conjugate vaccines may substantially reduce the burden of disease, but selecting the pivotal target population is paramount for the success of these vaccines.
When examining all enteric fever cases and matched controls, several factors that were protective against infection emerged. Notable protective variable included, an awareness of enteric fever, and the use of a metal cover on household water storage containers. The strong protective effect of a metal covering for water storage (as opposed to plastic) is likely explained by uncollected information associated with water treatment practices. We understand that households that filter their water prior to drinking are likely to have more permanent water storage containers with a metal covering; therefore it is likely that use of a ceramic filter explains the protective association of a metal covering on water storage units. We identified the use of a household latrine, apposed to a community latrine, as a significant risk for enteric fever in this analysis. Proximity of a household latrine to the kitchen or sleeping area of a family may explain this risk factor. In times of water shortage people are not be as likely to flush after each use, thereby exposing residents to contamination. The community latrines, however, are located at considerable distances from the living area of families in this setting and thus present a lesser risk of household fecal contamination. Further investigation into household toilet usage and behavior is warranted. Additionally, storage of water in houses with a large number of residents was found to be a risk for infection. Contamination of water stored in a household is common in regions where municipal water supply is unsafe [37]. Large numbers of people using one supply of stored water increases the opportunity for contamination due to increased contact with hands and utensils.
It has been suggested that S. Paratyphi A and S. Typhi may follow different transmission routes and the former requires a higher infectious dose for clinical disease [19]. Our exposure analysis supports this notion and suggests that while the two serovars share some risk factors and protective effects, the two organisms may have some independent epidemiology. In comparing various exposures between those with S. Typhi and S. Paratyphi A, S. Typhi cases were more likely to be associated with a poor-quality water source than the S. Paratyphi A cases. Water contamination with S. Typhi has been demonstrated previously in Nepal [16], [21]. The traditional stone spouts found throughout Kathmandu are supplied by a system of waterways with high risk of sewage contamination due to poor maintenance. High concentrations of S. Typhi and S. Paratyphi A DNA, in addition to fecal coliforms, have been reported from water samples from these stone spouts [10]. The fact that a majority of water-related and neighborhood condition exposures did not remain significantly associated with S. Typhi infection in the analysis is likely a result of our study design as both cases and controls within a particular neighborhood were likely to report using the same water sources. Finally, an additional risk factor specific for S. Typhi included a household income <$125/month; poor socioeconomic status is a well-known risk factor for infection with S. Typhi [13], [19].
For S. Paratyphi A induced enteric fever, two factors were found to be associated with risk of infection: residing in Kathmandu for less than years and eating street food in the two weeks preceding illness. Migration to Kathmandu from surrounding rural areas in search of economic prosperity is common and thus recent arrival into the city may correlate with immunologic naiveté. As S. Paratyphi A is an emerging pathogen in this setting [28], individuals from rural areas may not be exposed to the bacteria until arriving in Kathmandu. As S. Typhi has traditionally been the dominant serovar in this region, it is possible that transmission occurs throughout urban and rural Nepal [38]. Additionally, eating street food has been implicated in a previous study from Indonesia as a risk factor for infection with S. Paratyphi A as S. Paratyphi A reaches the required threshold to cause disease within certain food products [19]. As there is no hygiene legislation for street vendors in Kathmandu, such conditions and lengthy incubation periods are feasible.
There are some limitations with this study. Firstly, matching controls for ward of residence influenced the analysis, as the ward of residence is likely to be a correlate of various exposures. This altered the interpretation of the results, as many of the case/control pairs were concordant with regard to exposures, and did not contribute to the overall estimate of effect. Additionally, as shown by the serological data, identifying an appropriate group of controls in an endemic area poses a significant challenge. Given these challenges, we were still able to identify risks for enteric fever within this population that may allow for targeted interventions for the reduction of transmission. Such information is valuable as informed control and prevention strategies could provide a palatable and feasible method of reducing overall disease burden in this area of high endemicity.
Historical surveillance data suggest that enteric fever rates decrease in parallel with the introduction of treatment of water supplies, and the exclusion of human feces from food production [39], [40]. Improvements in the infrastructure of the municipal water delivery system, in addition to the provision of a combined vaccine against both S. Typhi and S. Paratyphi A, would be optimum for eliminating enteric fever in Kathmandu. In the short to mid-term absence of these interventions, we advocate safer household water supplies through the use of small water filtration and storage systems [41]. Additionally, our data suggest that improvements in the quality of street food, as well as promotion of enteric fever and toilet hygiene awareness through campaigns educating the population on risks, symptoms and preventive measures would have the largest impact on the burden of enteric fever in Kathamandu.
|
10.1371/journal.pntd.0000263 | Relationship between Transmission Intensity and Incidence of Dengue Hemorrhagic Fever in Thailand | Dengue is the most prevalent mosquito-borne virus, and potentially fatal dengue hemorrhagic fever (DHF) occurs mainly in secondary infections. It recently was hypothesized that, due to the presence of cross-immunity, the relationship between the incidence of DHF and transmission intensity may be negative at areas of intense transmission. We tested this hypothesis empirically, using vector abundance as a surrogate of transmission intensity.
House Index (HI), which is defined as the percentage of households infested with vector larvae/pupae, was obtained from surveys conducted on one million houses in Thailand, between 2002 and 2004. First, the utility of HI as a surrogate of transmission intensity was confirmed because HI was correlated negatively with mean age of DHF in the population. Next, the relationship between DHF incidence and HI was investigated. DHF incidence increased only up to an HI of about 30, but declined thereafter. Reduction of HI from the currently maximal level to 30 would increase the incidence by more than 40%. Simulations, which implemented a recently proposed model for cross-immunity, generated results that resembled actual epidemiological data. It was predicted that cross-immunity generates a wide variation in incidence, thereby obscuring the relationship between incidence and transmission intensity. The relationship would become obvious only if data collected over a long duration (e.g., >10 years) was averaged.
The negative relationship between DHF incidence and dengue transmission intensity implies that in regions of intense transmission, insufficient reduction of vector abundance may increase long-term DHF incidence. Further studies of a duration much longer than the present study, are warranted.
| An infection with dengue virus may lead to dengue hemorrhagic fever (DHF), a dangerous illness. There is no approved vaccine for this most prevalent mosquito-borne virus, which infects tens of millions (or more) people annually. Therefore, health authorities have been putting an emphasis on reduction of vector mosquitoes, genus Aedes. However, a new mathematical hypothesis predicted, quite paradoxically, that reducing Aedes mosquitoes in highly endemic countries may “increase” the incidence of DHF. To test this hypothesis based upon actual data, we compared DHF incidence collected from each of 1,000 districts in Thailand to data of Aedes abundance, which was obtained by surveying one million households. This analysis showed that reducing Aedes abundance from the highest level in Thailand to a moderate level would increase the incidence by more than 40%. In addition, we developed computer simulation software based upon the above hypothesis. The simulation predicted that epidemiological studies should be continued for a very long duration, preferably over a decade, to clearly detect such a paradoxical relationship between Aedes abundance and incidence of DHF. Such long-term studies are necessary, especially because tremendous efforts and resources have been (and perhaps will be) spent on combating Aedes.
| Dengue is the most prevalent vector-borne viral disease, the distribution of which has been expanding continually [1]. Dengue virus is transmitted by Aedes mosquitoes [2]–[4], which breed predominantly in water-holding containers within human habitats. Infections with dengue virus may manifest as dengue fever (DF), or the potentially more fatal dengue hemorrhagic fever (DHF). There are four serotypes of dengue virus, among which transient cross-protection exists [5]. Dengue virus is unique in that viral amplification in a primate host is enhanced dramatically in the presence of pre-existing immunity to a heterogeneous dengue serotype(s). This phenomenon, called antibody-dependent enhancement (ADE), had been reported initially in other arthropod-borne virus infections [6],[7]. In terms of dengue, ADE was demonstrated both by in vitro [8] and animal experiments [9]. Subsequently, pre-existing hetero-serotypic antibodies were shown to be associated with elevated risk for development of DHF [10].
Although the periodicity of highly oscillatory DHF outbreaks has been under intensive study [11],[12], determinants of the absolute magnitude of DHF incidence remain poorly understood. It would be understandable if the incidence of DF or DHF were affected positively by transmission intensity (measured either as force of infection or basic reproductive number). However, this intuitive thinking may be too naive in terms of dengue illness. As an example, increases in DF observed in Singapore were thought to be due to insufficient vector reduction [13],[14]. This paradox may be explained as follows, at least to some extent, by the age-dependent manifestation of DF [15],[16]. Under more intense transmission, infections occur at earlier ages [17]. Primary infections of younger children often result in no symptoms or mild illness [16],[18]. As a result, many infections do not manifest as clinical DF under high transmission intensity, and consequently, the incidence of DF decreases. This state of low incidence of clinical illness under intense transmission is known as “endemic stability” [15].
In contrast to DF, children seemed to be more prone to manifest DHF than are adults [19]–[21]. However, these studies, which did not fully consider the immunological status of the hosts, cannot be compared easily. This lack of reliable information about age-dependency in the manifestation of DHF has made it difficult to predict whether endemic stability occurs for DHF. On the other hand, a mathematical model recently predicted that, due to the presence of transient cross-serotype immunity, the incidence of DHF and transmission intensity will be correlated negatively at high transmission intensities [22]. This model hypothesized that a cross-protected individual will be seroconverted to an infecting viral serotype, while he/she is protected from manifesting severe illness. Under this assumption, which is consistent with results from experiments on monkeys [23], the individual would acquire immunity to nearly all serotypes while being cross-protected from clinical illness, at very intense transmission. As a result, the incidence of DHF could be correlated negatively to transmission intensity at areas of intense transmission, while the correlation is positive only at low levels of transmission. In the present study, such a complex correlation structure mixed with positive and negative correlations will be called “non-monotonic”, hereafter. To the contrary, correlation structure, which is simply either positive or negative, is referred to as “monotonic”.
The present study aims to provide an empirical example of this non-monotonic relationship between the incidence of DHF and transmission intensity, with transmission intensity represented by vector abundance. Vector abundance is one of the major determinants for transmission intensity of a vector-borne disease [24]. Accordingly, the WHO recommends that vector abundance be quantified in regions highly infested with Aedes through breeding site surveys and/or adult mosquito collections [25]. In developing countries, breeding site survey is preferred over mosquito collection, since the former is less labor-intensive. These surveys measure the number of houses or water containers infested by Aedes larvae/pupae through standard larval indices, such as House index (HI) and Breteau index (BI). HI is defined as the percentage of all surveyed houses in which Aedes larvae or pupae are present, while BI represents the number of infested containers in 100 houses. BI was shown to be relatively sensitive in predicting transmission [26]. Since HI and BI are strongly positively correlated [26], HI also may reflect transmission intensity to some extent. Although the absolute number of pupae is thought to reflect transmission intensity more directly than do larval indices [27]–[29], Southeast Asian households often possess many large water containers [30], that are irregularly shaped and partially sealed, making it difficult to obtain precise estimates of the absolute number of pupae. For this reason, absolute pupal counts have not been used in large-scale surveys in Thailand.
Here, we describe the empirical relationship between DHF incidence and transmission intensity, as represented by HI. The age-specific structure of this relationship also was characterized to support findings obtained for the entire population. The epidemiological characteristics of DHF were compared with predictions made by simulation of an individual-based model based upon the above mentioned mathematical modelling study. Our findings have major implications for future epidemiological surveys and dengue control programs.
In Thailand, the highest incidence of DHF occurs between June and August. Hence, entomological surveys mainly are conducted in the pre-epidemic season (e.g., April), with the assumption that vector abundance in this season will serve as an indicator of disease incidence later in the year. Between 2002 and 2004, a large-scale national Aedes survey was conducted in all 914 districts of Thailand. The survey was intended partly for community education and was implemented by investigators dispatched from 302 vector control units and community volunteers, under the supervision of the five regional vector-borne disease control offices. The administrative central village or municipality of each sub-district was surveyed because of their accessibility. A total of 40 houses were visited in each village/municipality. Prior appointments with the residents were not made, so that the residents did not clean their houses in advance. This design, taken with incomplete house registry in rural areas, made genuine randomization impossible. Surveys were conducted in April of each year, with 9,483 villages surveyed in 2002, 9,763 in 2003, and 7,482 in 2004. The HI values were averaged for each district for comparison with district-level DHF data. The average population of a district in Thailand is 67,500.
The Bureau of Epidemiology, Ministry of Public Health, provided the annual number of cases of DHF (including Dengue Shock Syndrome) in nine age categories (0–4, 5–9, 10–14, 15–24, 25–34, 35–44, 45–54, 55–64, ≥65 years) for each district, for the years between 1994 and 2004. Age-stratified population data, based upon five yearly censuses/surveys and yearly projections, were obtained from the National Statistics Office of Thailand (http://www.nso.th.go) to calculate DHF incidence. The incidence of DHF in an entire district population was adjusted to the national age-population structure of 2000 by using the Direct Method, to eliminate possible interference by the heterogeneity in demographic structure. The mean age of DHF cases was calculated as the average of the mid-point of the different age categories (2.5, 7.5, 12.5, 20, 30, 40, 50, 60, and 75 years) weighted by the number of cases in each category.
Subsequent statistical analyses were performed using R 2.6.2 and Stata 9.2. We used non-parametric statistical methods, Spearman's rank correlation analysis and the generalized additive model (GAM), so that analyses did not have to assume any fixed distribution a priori. Akaike's Information Criteria (AIC) inversely represented the goodness of fit, or predictability, for a regression model obtained from GAM [31]. Deviance around the prediction also was presented, although this measurement is not adjusted for degree of freedom (df) used in a regression model.
To ensure that HI could be used as a reliable surrogate of transmission intensity, we compared the mean age of DHF cases to HI using rank correlation analysis. A high mean age of DHF cases was used as an indicator of low transmission intensity, because the mean age of infected individuals generally is negatively correlated with the transmission intensity of an acute infectious disease [17]. Since each district was surveyed three times (2002, 2003, and 2004), the possible bias from this repeated measurement was adjusted by simply aggregating records from three years for each district. Among the all 914 districts, this analysis incorporated 909 districts that reported at least one case of DHF between 2002 and 2004.
We examined the quantitative relationship between incidence of DHF and HI using GAM. Logarithm was used as the link function. First, we tested this relationship by incorporating only HI as the independent variable (univariate analysis). Then, we adjusted for possible confounding by socioeconomic and climatic variables. Socioeconomic factors may affect reported incidence in diverse fashions. For example, incidence may be biased by (a) the prevalence of health offices, which are responsible for DHF case reporting in each district. Abundance of breeding places is affected by local water storage practices (reviewed by [32]). Our analysis incorporated the following socioeconomic factors that were reported to be associated with dengue transmission intensity [33]: (b) per capita number of public large water wells, (c) that of public small wells, (d) that of private small wells, (e) annual birth rate per 1,000 individuals, (f) proportion of households owning land, and (g) proportion of villages in which high schools are present. These seven socioeconomic variables (a–f), censused every other year, were obtained from the Information Processing Centre of Thammasat University, Bangkok, and were interpolated linearly to the intervening years. On the other hand, dengue transmission intensity is influenced by climatic factors as well. Temperature affects critically the rate of viral amplification in mosquitoes [34]. In addition, extremely high or low temperatures are rate-determining factors for the growth and survival of mosquitoes [35]. Atmospheric vapor pressure is known to affect dengue transmission [36]. Aridity, which is likely to reflect the scarcity of underground water, may be associated with increased use of household water containers. To adjust for these possible confounders, the following climatic variables were obtained from the University Cooperation for Atmospheric Research [37]: (a) temperature averaged between January and February, the coolest months in Thailand (“winter temperature”, °C), (b) temperature averaged between April and May, the hottest months (“summer temperature”, °C), (c) average vapor pressure (AVP, hPa), and (d) average pan evapo-transpiration (APET, mm/day). These climatic variables were obtained from 89 weather stations in Thailand and its adjacent countries, averaged for each year, and interpolated to the geographic centroid of each district by using Inverse Distance Weighting method. We confirmed that multiple interpolation methods generated comparative results, perhaps because these weather stations constituted a sufficiently exhaustive dataset [38].
Collectively, these socioeconomic/climatic variables were averaged for the period for which the dependent variable, incidence, was averaged. We enrolled districts from which socioeconomic and climatic variables have been available from 1994 to 2004. Consequently, 785 districts were enrolled. This dataset (incidence linked with covariates) is available on request from the corresponding author. Multivariate analyses were conducted using the following procedure. First, HI and all socioeconomic/climatic factors were incorporated as independent variables, with df of each variable set to 2. Next, independent variables that remained significant (P<0.05) in a stepwise elimination procedure were selected, generating the “smallest regression model for df = 2”. Finally, df of each of the remaining six variables was replaced with df = 3, generating 26 combinations of df. Among these, the combination that exhibited the smallest AIC was adopted as the “final regression model”.
The relationship between DHF incidence and HI was examined within different age classes for which original age categories were aggregated into the following three age classes: 0–4, 5–24, and ≥25 years. GAM was applied similarly to these age-class specific incidences.
We employed computer simulations to see whether (and to what extent) the observed epidemiological pattern could be explained based upon a theoretical framework. The assumption of the above mentioned mathematical model was expressed equivalently by an individual-based model (see Protocol S1, Section I). This model is summarized as follows. The cross-protective period was assumed to be of a fixed duration (“C” years). Inoculation by a virus, which occurred during this cross-protective period, does not develop into DHF, but induces seroconversion. As the cross-protective period expires, the individual is predisposed to the risk of manifesting DHF in a subsequent inoculation by a secondary (or later) serotype. An individual could manifest DHF after secondary, tertiary or quaternary infections.
In addition, this individual-based model can incorporate the age-dependency in the probability to manifest DHF (categorical parameter “A”, defined in Figure S1B in Protocol S1). Transmission intensity is represented by basic reproductive number (R0) of dengue virus. The present study parameterized simulations with the following three scenarios. (I) Cross-immunity scenario: the duration of cross-serotype protection (“C”) was set to two years, while the probability to manifest DHF was assumed to be independent of age (A = 0). We selected this duration of cross-immunity based upon the results of sensitivity analysis (see Protocol S1). (II) Age-dependency scenario: the probability to manifest DHF in secondary or later infections was assumed to increase in accordance with the age of the individual (A = 2), while no cross-immunity was assumed (C = 0). (III) Control scenario: no cross-immunity or age-dependency was assumed (C = 0, A = 0). R0 was selected by extrapolating the mean age of DHF obtained between 2002 and 2004 from each of the 785 districts, through the relationship between R0 and mean age of DHF (Figure S6 in Protocol S1). This set of R0 values was used as the input for all three scenarios. Each simulation was run for 150 years.
At different durations for averaging (W), the goodness of fit was compared between the statistical models that explained the incidence in simulations versus those that explained the actual incidence. Incidences of DHF generated from simulations were averaged from the last W years [W = 3, 4 … 40] (for example, 148th, 149th and 150th years were averaged for W = 3). Subsequently, the averaged incidence was regressed against R0 using GAM.
On the other hand, actual incidences were averaged for the recent W years [W = 3, 4 … 11] (for example, W = 3 corresponds to 2002–2004; W = 11 corresponds to 1994–2004). Then, the averaged actual incidence was regressed against HI obtained from the 2002–2004 survey, and socioeconomic/climatic variables averaged for the recent W years.
The national-level mean age of DHF cases was 16 years during 2002 to 2004. The mean HI recorded each April during 2002 to 2004 was 23. As shown in Figure 1, the mean age was negatively correlated with HI at the district level (Spearman's R = −0.35, P<0.0001, N = 909).
During 2002 to 2004, the annual DHF incidence was 83 per 100,000 individuals. HI showed a statistically significant contribution to the log incidence of DHF, both in univariate and multivariate regression models (Table 1). Univariate analysis of GAM revealed that the correlation between HI and incidence was positive below about HI = 30, while the correlation was negative above this HI value (Figure 2B; Figure 3). As HI decreases from 70 to 30, for example, the log incidence would increase by 0.35 (Figure 3), which is equivalent to an increase of 40% in incidence, since exp (0.35) = 1.4.
In multivariate analysis, the following six variables remained in the final regression model (Table 1; Figure 4): HI, winter temperature, summer temperature, APET, public large wells, and birth rate. The best predictability (or lowest AIC) was achieved by the final regression model which assigned df = 3 only to public large wells, and df = 2 to other covariates. Multivariate analysis estimated that, as HI decreases from 70 to 30, log incidence would increase by 0.6 (Figure 4A), which corresponds to an increase of 80%.
Although incorporation of socioeconomic/variables improved the goodness of fit, this multivariate predictive model still failed to reproduce the very wide variation in the observed incidence (compare Figure 2B vs Figure 5).
Further analysis of the age-specific associations between incidence of DHF and HI was conducted, as shown in Table 2. Univariate analysis revealed that incidence and HI were positively correlated in the youngest age class (Figure 2C); whereas, DHF incidence and HI were negatively correlated in the oldest age class (Figure 2E). A non-monotonic relationship between DHF incidence and HI was detected within the intermediate age class (Figure 2D). When the socioeconomic/climatic variables were incorporated, the statistical significance of positive correlation among the youngest age class diminished (Table 2).
As shown in Figure 6, averaging only the last three years of each simulation resulted in a negligibly detectable relationship between DHF incidence and R0, which greatly resembled the empirical relationship (compare with Figure 2B). As the window for averaging increased, the relationship generally became more apparent. The incidences generated by simulations with cross-immunity were much more dispersed than those generated by other simulations, at any window lengths.
GAM was applied to examine the relationship between incidences generated by simulations and R0 (Figure 7). As a result, GAM detected a non-monotonic relationship in the simulations with cross-immunity (Figure 7A), a negative relationship in those with age-dependency (Figure 7E), and a slightly positive relationship in the control simulations (Figure 7I). Age-stratification of the simulation results generated a similar trend in the empirical data, regardless of the presence of cross-immunity or age-dependency (Figure 7B–D, F–H, J–L). That is, a positive correlation was observed between DHF incidence and transmission intensity in the younger population, and a negative correlation was present in the older population.
The goodness of fit in predicting incidence by R0 showed remarkable differences between simulation with cross-immunity and those without cross-immunity (Figure 8). The predictability was much worse in simulation with cross-immunity than in other simulations. In addition, the response to the window length was more complex in the presence of cross-immunity than in other simulations. That is, in simulations without cross-immunity, the predictabilities improved continuously as W increased. In contrast, the predictability in the presence of cross-immunity deteriorated as the window for averaging increased from W = 3 to W = 4, then improved up to W = 6. With a small setback at W = 7, it improved again thereafter. Such a complex response of predictability to W was reproduced at diverse durations of cross-immunity (Figure 9), which were sufficiently long to generate dominant supra-annual periodicities (see Section II and Figure S7 in Protocol S1).
The goodness of fit in predicting actual incidence, either by HI only or by HI and covariates, showed a similarly complex response to W (Table 3). The predictability attained solely by HI was much inferior to that attained in any simulations (Figure 8). However, the predictability using the multivariate regression model was as good as that in simulations with cross-immunity, up to W = 8.
Our analysis demonstrates that HI is a reliable indicator of transmission intensity, at least at the district level. The usefulness of HI is evident by its highly significant, inverse relationship to mean age, otherwise equivalent to a positive correlation between HI and transmission intensity. Our findings are consistent with observations from Singapore, where an increase in the mean age of patients with dengue infection was preceded by a substantial reduction in HI [13],[14].
Analysis of DHF incidence among the entire Thai population revealed that incidence rose up to HI of about 30 and gradually declined thereafter. This non-monotonic relationship appears to be consistent with a state of endemic stability. However, the age-dependency in the probability to manifest DHF may not simply satisfy the condition for endemic stability, because DHF occurs more frequently in children than in adults. On the other hand, cross-immunity explains not only this non-monotonic relationship (Figure S2 in Protocol S1), but also the wide variation in incidence of DHF, as well as its complex response to the duration of window for averaging (Figure S8 in Protocol S1). Of note, the regression model comprised of HI and socioeconomic/climatic variables predicted actual incidence to the same goodness of fit, with which R0 predicted incidence from simulations with cross-immunity (Figure 8). This finding may support the validity of the multivariate regression model, and that of our assumption for cross-immunity, simultaneously.
Stratification of data according to age revealed a positive association between DHF incidence and HI among the youngest population. In contrast, a negative association was observed in the oldest population. These contrasting correlations may be explained as follows. Under low transmission intensity, the majority of individuals in the youngest age class do not possess antibodies against any serotype and are relatively resistant to DHF. As the transmission intensity increases, a larger number of individuals in this age class possess antibodies to only one serotype, making them predisposed to DHF. Therefore, the correlation between DHF incidence and transmission intensity becomes positive in the youngest age class, as observed here. In contrast, when transmission intensity is low, many in the oldest age class possess antibodies against only one serotype and are predisposed to DHF. As transmission intensity increases, more members of this age class possess antibodies against almost all serotypes, conferring resistance to DHF. Importantly, these age-stratified relationships could be reproduced by simulations of any scenarios examined. Therefore, this analysis did not differentiate whether cross-immunity or age-dependency determined the epidemiological characteristics of DHF.
The negative response of incidence to transmission intensity at areas of intense transmission has important public health implications, regardless of its underlying mechanism. The incidence of DHF is affected by the dominant virus serotype, which shifts from period to period [39],[40]. In addition, HI measured in one country cannot be compared with HI in another country. Since our analysis was based on a single three-year period in one country, the stability of our estimated HI value at the maximum (“turning”) point should be treated with some caution. However, with these caveats, our results indicate that insufficient reduction of vector abundance in highly endemic areas could result in an increased incidence of DHF. As the HI decreases from the current highest level in Thailand, the incidence of DHF could increase by more than 40%. Any medical/public-health intervention that causes a foreseeable increase of illness should be subject to ethical discussion.
Theoretically, sufficiently radical reduction of vector mosquitoes can achieve a decrease of the entire incidence of DHF. However, it is unclear whether such radical vector control is possible at a nation-wide scale in developing countries. Instead, reduction of the vector population may become stagnant as the vector abundance decreases. Furthermore, even substantial vector reduction (for example, from HI = 60 to10) would not necessarily decrease the final incidence (extrapolate the HI values to incidence in Figure 3 and Figure 4A), but would result most likely in a greater number of DHF cases accumulated over the course of time. This calculation suggests that it is extremely difficult for vector control alone to achieve the ultimate goal of control program– reduction of incidence. |
10.1371/journal.ppat.1000665 | Identification of Host Cytosolic Sensors and Bacterial Factors Regulating the Type I Interferon Response to Legionella pneumophila | Legionella pneumophila is a gram-negative bacterial pathogen that replicates in host macrophages and causes a severe pneumonia called Legionnaires' Disease. The innate immune response to L. pneumophila remains poorly understood. Here we focused on identifying host and bacterial factors involved in the production of type I interferons (IFN) in response to L. pneumophila. It was previously suggested that the delivery of L. pneumophila DNA to the host cell cytosol is the primary signal that induces the type I IFN response. However, our data are not easily reconciled with this model. We provide genetic evidence that two RNA-sensing proteins, RIG-I and MDA5, participate in the IFN response to L. pneumophila. Importantly, these sensors do not seem to be required for the IFN response to L. pneumophila DNA, whereas we found that RIG-I was required for the response to L. pneumophila RNA. Thus, we hypothesize that bacterial RNA, or perhaps an induced host RNA, is the primary stimulus inducing the IFN response to L. pneumophila. Our study also identified a secreted effector protein, SdhA, as a key suppressor of the IFN response to L. pneumophila. Although viral suppressors of cytosolic RNA-sensing pathways have been previously identified, analogous bacterial factors have not been described. Thus, our results provide new insights into the molecular mechanisms by which an intracellular bacterial pathogen activates and also represses innate immune responses.
| Initial detection of invading microorganisms is one of the primary tasks of the innate immune system. However, the molecular mechanisms by which pathogens are recognized remain incompletely understood. Here, we provide evidence that an immunosurveillance pathway (called the RIG-I/MDA5 pathway), thought primarily to detect viruses, is also involved in the innate immune response to an intracellular bacterial pathogen, Legionella pneumophila. In the response to viruses, the RIG-I/MDA5 immunosurveillance pathway has been shown to respond to viral RNA or DNA. We found that the RIG-I pathway was required for the response to L. pneumophila RNA, but was not required for the response to L. pneumophila DNA. Thus, one explanation of our results is that L. pneumophila RNA may access the host cell cytosol, where it triggers the RIG-I/MDA5 pathway. This is unexpected since bacteria have not previously been thought to translocate RNA into host cells. We also found that L. pneumophila encodes a secreted bacterial protein, SdhA, which suppresses the RIG-I/MDA5 pathway. Several viral repressors of the RIG-I/MDA5 pathway have been described, but bacterial repressors of RIG-I/MDA5 are not known. Thus, our study provides novel insights into the molecular mechanisms by which the immune system detects bacterial infection, and conversely, by which bacteria suppress innate immune responses.
| The intracellular bacterium Legionella pneumophila has become a valuable model for the study of immunosurveillance pathways. L. pneumophila is a motile gram-negative bacterium that is the cause of a severe pneumonia called Legionnaires' Disease [1]. In the environment, L. pneumophila is believed to replicate in various species of freshwater amoebae. In humans, L. pneumophila causes disease by replicating within alveolar macrophages in the lung [2]. Replication in macrophages and amoebae requires a type IV secretion system that the bacterium uses to inject effector proteins into the host cell cytosol [3]. These effectors are believed to orchestrate the creation of an intracellular vacuole in which L. pneumophila can replicate. Interestingly, there appears to be considerable redundancy among the effectors, and there are few examples of single effector mutations that have a large effect on intracellular replication of L. pneumophila. One L. pneumophila effector required for intracellular replication is SdhA [4], but the mechanism by which SdhA acts on host cells remains uncertain [4].
A variety of immunosurveillance pathways that detect L. pneumophila infection have been described [5],[6],[7],[8]. The best characterized cytosolic immunosurveillance pathway requires the host proteins Naip5 and Ipaf to detect the cytosolic presence of L. pneumophila flagellin, leading to activation of caspase-1, rapid pyroptotic macrophage death, and efficient restriction of bacterial replication [9],[10],[11],[12],[13]. L. pneumophila has also been observed to induce transcriptional activation of type I interferon (IFN) genes in macrophages and epithelial-like cell lines by a mechanism that remains incompletely characterized [14],[15]. Induction of type I IFNs by L. pneumophila is independent of the flagellin-sensing pathway [16], but also appears to contribute to restriction of bacterial replication in macrophages [16],[17] and epithelial-like cell lines [14].
Type I IFNs are an important class of cytokines that orchestrate diverse immune responses to pathogens [18]. Encoded by a single IFNβ gene as well as multiple IFNα and other (e.g., IFNε, κ, δ, ζ) genes, type I IFNs are transcriptionally induced by a number of immunosurveillance pathways, including Toll-like receptors (TLRs) and a variety of cytosolic sensors [19]. For example, cytosolic RNA is recognized by two distinct helicase and CARD-containing sensors, RIG-I and MDA5 [20], that signal through the adaptor IPS-1 (also called MAVS, CARDIF, or VISA) [21],[22],[23],[24],[25]. The cytosolic presence of DNA also induces type I IFNs, but this phenomenon is less well understood [15],[26]. Studies with Ips-1-deficient mice have indicated that cytosolic DNA can signal independently of Ips-1 in many cell types, including macrophages [25]. However, cytosolic responses to DNA appear to require IPS-1 in certain cell types, including 293T cells [26],[27]. Indeed, two recent reports have described a pathway by which AT-rich DNA can signal via IPS-1 [28],[29]. In this pathway, DNA is transcribed by RNA polymerase III to form an RNA intermediate that can be sensed by RIG-I. The RNA Pol III pathway appears to be operational in macrophages, but is redundant with other DNA-sensing pathways in these cells. A couple of reports have proposed that DAI (also called ZBP-1) is a cytosolic DNA-sensor [30],[31], but Zbp1-deficient mice appear to respond normally to cytosolic DNA [32], consistent with the existence of multiple cytosolic sensors for DNA. Other small molecule compounds, such as cyclic-di-GMP and DMXAA, can also trigger cytosolic immunosurveillance pathways leading to induction of type I IFNs, but these remain to be fully characterized [33],[34],[35].
Type I IFNs are typically considered antiviral cytokines that act locally to induce an antiviral state and systemically to induce cellular innate and adaptive immune responses [19]. Mice deficient in the type I IFN receptor (Ifnar) are unable to respond to type I IFNs, and are highly susceptible to viral infections. Interestingly, most bacterial infections also trigger production of type I IFNs, but the physiological significance of type I IFNs in immune defense against bacteria is complex. Type I IFN appears to protect against infection with group B Streptococcus [36], but this is not the case for many other bacterial infections. For example, the intracellular gram-positive bacterium Listeria monocytogenes induces a potent type I IFN response [37],[38], but Ifnar-deficient mice are actually more resistant to L. monocytogenes infection than are wildtype mice [39],[40],[41]. Many bacterial pathogens, including Francisella tularensis, Mycobacterium tuberculosis, Brucella abortus, and group B Streptococcus, induce type I IFN production by macrophages via a cytosolic TLR-independent pathway [42],[43],[44],[45], but the bacterial ligands and host sensors required for the interferon response of macrophages to these bacteria remain unknown.
It was demonstrated that induction of type I IFN by L. pneumophila in macrophages did not require bacterial replication or signaling through the TLR-adaptors MyD88 or Trif, but did require the bacterial Dot/Icm type IV secretion system [15]. Because the IFN response could be recapitulated with transfected DNA [15],[26] and because Dot/Icm system has been shown to conjugate DNA plasmids to recipient bacteria [46], it was proposed that perhaps L. pneumophila induced type I IFN via a cytosolic DNA-sensing pathway [15]. Another report used RNA interference to implicate the signaling adaptor IPS-1 (MAVS) in the IFN response to L. pneumophila in human A549 epithelial-like cells [14]. However, the significance of this latter finding is unclear since RNAi-mediated knockdown of RIG-I and MDA5, the two sensor proteins directly upstream of IPS-1, did not have an effect on induction of type I IFN by L. pneumophila [14]. Moreover, the A549 response to L. pneumophila may be distinct from the macrophage or in vivo response.
Recently, one report proposed that L. pneumophila DNA was recognized in the cytosol by RNA polymerase III [29], resulting in the production of an RNA intermediate that triggered IFN production via the IPS-1 pathway. Apparently consistent with this proposal, Ips-1-deficient mouse macrophages did not produce type I IFN in response to L. pneumophila [29]. Moreover, since Pol III acts preferentially on AT-rich substrates, it is plausible that Pol III would recognize the L. pneumophila genome, which has a high proportion (62%) of A:T basepairs. However, the response to L. pneumophila DNA was not investigated [29]. In addition, the same report, as well as others [28],[34], observed that the type I IFN response to AT-rich (or any other) DNA is not Ips-1-dependent in mouse cells. Thus, if L. pneumophila DNA was reaching the cytosol, the simplest prediction would be that the resulting type I IFN response would be independent of Ips-1, instead of Ips-1-dependent, as was shown [29]. Thus, the mechanism of IFN induction by L. pneumophila remains unclear.
In the present study, we sought to define bacterial and host factors controlling the macrophage type I IFN response to L. pneumophila. In agreement with previous studies [14],[29], we find that Ips-1 is required for optimal induction of type I IFN in response to L. pneumophila infection in vitro. We extend this observation by demonstrating that Ips-1 also contributes to the type I IFN response in an in vivo model of Legionnaires' Disease. Furthermore, we provide the first evidence that two RNA sensors upstream of Ips-1, Rig-i and Mda5, are involved in the macrophage interferon response to L. pneumophila. Importantly, however, we did not observe a role for the Pol III pathway in the type I IFN response to L. pneumophila. Instead, we found that L. pneumophila genomic DNA stimulates an Ips-1/Mda5/Rig-i-independent IFN response in macrophages, which contrasts with the Ips-1-dependent response to L. pneumophila infection. On the other hand, we found that L. pneumophila RNA stimulated a Rig-i-dependent IFN response. Thus, our data are consistent with a model in which L. pneumophila RNA, or host RNA, rather than L. pneumophila DNA, is the primary ligand that stimulates the host IFN response. We also investigated whether bacterial factors that modulate the host type I IFN response. Although numerous viral proteins that interfere with IFN signaling have been described [19], similar bacterial proteins have not been documented. It is therefore interesting that we were able to identify a secreted bacterial effector, SdhA, as an inhibitor of the Ips-1-dependent IFN response to L. pneumophila. Taken together, our findings provide surprising evidence that cytosolic RNA-sensing pathways are not specific for viral infections but can also respond to bacterial infections, and moreover, our data provide a specific example of a bacterial factor that suppresses the host IFN response.
We hypothesized that a cytosolic innate immune sensing pathway controls the type I IFN response to L. pneumophila. To test this hypothesis, we determined whether macrophages deficient in known cytosolic RNA and DNA sensing pathway components can induce type I IFNs in response to L. pneumophila. Macrophages were infected with L. pneumophila at a multiplicity of infection (MOI) of 1 and induction of interferon beta (Ifnb) message was analyzed by quantitative RT-PCR after 4 hours (Figure 1A–D). As previously reported [29], Ips-1−/− macrophages showed a significantly reduced induction of Ifnb in response to infection with wild type L. pneumophila compared to Ips-1+/+ macrophages (p<0.05; Figure 1A). Induction of Ifnb was not completely eliminated in Ips-1−/− macrophages, however, as Irf3−/− macrophages exhibited an even lower induction of Ifnb compared to Ips-1−/− (p<0.05; Figure 1A). Consistent with previous reports [15], we found that the Dot/Icm type IV secretion system was required to elicit the macrophage type I interferon response since Δdot L. pneumophila did not induce a robust type I interferon response (Figure 1A). These results suggest that L. pneumophila induces type I IFN via a cytosolic RNA immunosurveillance pathway that involves the adaptor Ips-1.
We hypothesized that a cytosolic RNA sensor that functions upstream of Ips-1 could be involved in the type I interferon host response to L. pneumophila. However, knockdown experiments in A549 cells previously failed to reveal a role for the known sensors (MDA5 and RIG-I) upstream of IPS-1 [14]. Therefore, we tested Mda5−/− knockout macrophages (Figure 1B) and found reduced induction of Ifnb message as compared to control Mda5+/+ macrophages. Importantly, however, Dot-dependent induction of type I IFN was not completely abolished in Mda5−/− macrophages, implying that other redundant pathways are also involved.
Rig-i knockout mice die as embryos, so we were unable to obtain Rig-i−/− knockout macrophages. To circumvent this problem, we stably transduced immortalized macrophages with a retrovirus expressing an shRNA to knock down Rig-i expression. Quantitative RT-PCR demonstrated that the knockdown was effective, even in infected macrophages (Figure 1C), and that Rig-i knockdown had a significant effect on the induction of type I interferon by L. pneumophila (Figure 1D). In the experiments in Figures 1C and 1D we used the ΔflaA strain of L. pneumophila, but similar results were obtained with wildtype, and it was previously shown that flagellin is not required for the IFN response to L. pneumophila [14],[16]. It is unusual, but not unprecedented, that a pathogen would stimulate both the RIG-I and MDA5 RNA-sensing pathways [47].
At present, only one candidate cytosolic DNA sensor involved in the IFN response has been described [30],[31]. To determine whether this sensor, called Dai (or Zpb1), is involved in the type I interferon response to L. pneumophila, we tested whether Zbp1−/− macrophages respond to L. pneumophila. We observed similar levels of Ifnb induction in Zbp1+/+ and Zbp1−/− macrophages (Figure 1E). Taken together, these results imply that the RNA sensors Rig-i and Mda5, but not the DNA sensor Zbp1, are involved in sensing L. pneumophila infection.
We tested whether loss of signaling through the RNA sensing components Ips-1 or Mda5 could mimic the previously observed permissiveness of Ifnar−/− macrophages [16]. However, neither Ips-1−/− nor Mda5−/− macrophages were permissive to L. pneumophila, suggesting that the low levels of IFNβ produced in the absence of Ips-1 or Mda5 are sufficient to restrict L. pneumophila growth (Figure S1).
To identify bacterial components that modulate the type I interferon response to L. pneumophila, we conducted a transposon mutagenesis screen. The LP02 strain of L. pneumophila was mutagenized with a mariner transposon as described previously [12]. Individual transposon mutants were used to infect MyD88−/−Trif−/− bone marrow-derived macrophages at an MOI of 1, and after approximately 16 hours, supernatants were collected and overlayed on type I IFN reporter cells [48]. Induction of type I IFN was compared to wild type (LP02) and Δdot L. pneumophila controls. We tested approximately 2000 independent mutants and isolated eight mutants that were confirmed to be defective in induction of type I IFN. All these mutants harbored insertions in genes required for the function of the Dot/Icm apparatus (e.g., icmB, icmC, icmD, icmX, icmJ), thereby validating the screen.
Interestingly, a single transposon mutant, 11C11, was found that consistently hyperinduced the type I interferon response. The transposon insertion mapped to the 3′ end (nucleotide position 3421 of the open reading frame) of a gene, sdhA, that was previously shown [4] to encode a type IV secreted effector protein of 1429 amino acids (166kDa) (Figure 2A). SdhA has previously been shown to be essential for bacterial replication in macrophages [4], but a connection to type I IFNs was not previously noted.
To confirm that the hyperinduction of type I interferon was due to mutation of sdhA, the 11C11 transposon mutant was compared to an unmarked clean deletion of sdhA (Figure 2B). Both the 11C11 mutant and ΔsdhA L. pneumophila showed similar levels of hyperinduction of type I interferon. The L. pneumophila genome contains 2 paralogs of sdhA, called sidH and sdhB. A triple knockout strain, ΔsdhAΔsdhBΔsidH, was compared to single deletion of sdhA to determine if either paralog regulated the induction of type I IFNs. Similar levels of IFNβ were induced ΔsdhAΔsdhBΔsidH and ΔsdhA (Figure 2B). Similar results were obtained when induction of Ifnb was assessed by quantitative RT-PCR (Figure 2C). A role for sdhA in regulating the interferon response was further confirmed by complementing the ΔsdhA mutation with an sdhA expression plasmid [4]. As expected, the complemented strain induced significantly less type I IFN than the control ΔsdhA strain harboring an empty plasmid (Figure 2D). These results indicate that SdhA functions, directly or indirectly, to repress the induction of type I IFN by L. pneumophila.
It was possible that ΔsdhA mutants hyperinduced type I IFN via a pathway distinct from the normal cytosolic RNA-sensing pathway that responds to wildtype L. pneumophila. Therefore, to determine whether hyperinduction of type I interferon by ΔsdhA occurs through the same pathway that responds to wild type L. pneumophila, we infected Ips-1−/− and Mda5−/− macrophages with ΔsdhA L. pneumophila. Induction of Ifnb message was determined by quantitative RT-PCR. The hyperinduction of Ifnb seen in Ips-1+/+ macrophages was almost abolished in Ips-1−/− macrophages (p<0.001; Figure 3A). As a control, induction of Ifnb by poly I:C, a double-stranded synthetic RNA analog, was also Ips-1-dependent as expected. Similarly, the hyperinduction of Ifnb was also reduced in Mda5−/− macrophages (p<0.01; Figure 3B). However, the Mda5−/− macrophages still induced significant amounts of Ifnb, suggesting that the requirement for Mda5 is not complete. We also tested the ΔsdhA mutant in Rig-i knockdown macrophages. Rig-i knockdown appeared to be effective (Figure 3C) and specifically diminished Ifnb expression (Figure 3D). Thus, the residual Ifnb induction in Mda5−/− may be due to Rig-i, or to another uncharacterized pathway. As a control, Theiler's virus (TMEV) induced Ifnb in a completely Mda5-dependent manner, as expected (Figure 3B).
It was previously shown that ΔsdhA mutants induce a rapid death of infected macrophages that is dependent upon activation of multiple cell death pathways [4]. Consequently, we hypothesized that the hyperinduction of type I IFN by the ΔsdhA mutant might be due to the release of molecules from dying cells, such as DNA, that could induce Ifnb expression. To rule out this explanation, we infected Casp1−/− macrophages, which are resistant to cell death at the early timepoints examined (e.g., 4h post infection), and asked whether type I interferon was still hyperinduced in response to ΔsdhA L. pneumophila. In fact, we found that Casp1−/−macrophages infected with the ΔsdhA mutant hyperinduced Ifnb to levels above that observed in B6 macrophages (Figure 4A). We suspect that the increased Ifnb induction seen in Casp1−/− cells was an indirect consequence of the lower levels of cell death in these cells, and was not due to a specific suppression of type I interferon transcription by Casp1 activation. In any case, our results indicated that the hyperinduction of type I IFN by the ΔsdhA mutant was not due to increased cell death induced by the mutant. As a control, we confirmed that Casp1−/− macrophages were resistant to cell death at the 4h timepoint tested (Figure 4B).
Induction of Ifnb is often regulated by a positive feedback loop in which initial production of IFNβ results in signaling through the type I IFN receptor (Ifnar) and synergistically stimulates the production of additional type I IFN. We therefore examined whether the hyperinduction of Ifnb by the ΔsdhA mutant might be due to positive feedback through the type I IFN receptor. To test this possibility we examined induction of Ifnb by the ΔsdhA mutant in Ifnar−/− macrophages. We found that hyperinduction of Ifnb by ΔsdhA L. pneumophila occurs even in the absence of signaling from the type I interferon receptor, since Ifnar−/− macrophages hyperinduce Ifnb in response to infection with ΔsdhA L. pneumophila (Figure 4A).
The mechanism by which the ΔsdhA mutant induces cell death remains unclear [4]. Studies with the intracellular bacterial pathogen Francisella tularensis have demonstrated the existence of a type I IFN-inducible caspase-1-dependent cell death pathway [43]. Therefore, we sought to establish if caspase-1-dependent cell death occurred in the absence of Ifnar signaling in response to wild type and ΔsdhA L. pneumophila. Ifnar−/− macrophages were infected at an MOI of 1 and assayed for release of the intracellular enzyme lactate dehydrogenase (LDH) 4 hours post infection. Ifnar−/− macrophages exhibited similar LDH release as B6 macrophages, whether infected with WT or ΔsdhA L. pneumophila, and this LDH release was dependent upon caspase-1 activation (Figure 4B). These data demonstrate that caspase1-dependent pyroptotic death occurs independently of the type I interferon receptor during infection with wild type and ΔsdhA L. pneumophila.
Since growth of the ΔsdhA mutant is severely attenuated in macrophages [4], we hypothesized that hyperinduction of type I interferon might contribute to the restriction of replication of the ΔsdhA mutant. To test this hypothesis, we infected lfnar−/− macrophages with luminescent strains of L. pneumophila at an MOI of 0.01 and monitored bacterial replication over a 72 hour time period. As previously reported [16], lfnar−/− macrophages were more permissive to WT and ΔflaA L. pneumophila as compared to C57BL/6 macrophages (Figure S2A, C). However, the ΔsdhA or ΔflaAΔsdhA L. pneumophila strains were still significantly restricted in Ifnar−/− macrophages (Figure S2B, D). Thus, SdhA is required for bacterial replication in macrophages primarily via a mechanism independent of its role in suppressing type I IFN. As expected, Δdot L. pneumophila did not replicate in WT or Ifnar−/− macrophages (Figure S2E).
Since SdhA is a secreted effector, we hypothesized that SdhA may act in the host cell cytosol, rather than in the bacterium, to repress Ifnb induction. To test this hypothesis, we co-expressed SdhA with MDA5 or RIG-I, by transient transfection of HEK293T cells, and assessed interferon expression with an IFNβ-luciferase reporter. Expression of either MDA5 or RIG-I robustly induced the IFNβ-luc reporter upon stimulation with poly I:C (Figure S3). When SdhA was co-expressed with MDA5, a dose-dependent repression of the IFNβ-luc reporter was observed (Figure S3A). Co-expression of SdhA also resulted in a dose-dependent repression of RIG-I-dependent induction of the IFNβ-luc reporter (Figure S3B). However, SdhA co-expression did not affect TRIF-dependent induction of the IFNβ-luc reporter (Figure S3C), arguing against the possibility that SdhA expression has non-specific effects on IFNβ-luc induction. These results must be interpreted with caution since the 293T IFNβ-luc reporter system is highly artificial; moreover, we have not demonstrated a direct interaction of SdhA with signaling components in the RNA-sensing pathway. In fact, the reported effects of SdhA on mitochondria [4] suggest the effect may be somewhat indirect (see Discussion). Nevertheless, the 293T transfection results suggest that SdhA can act in the host cytosol to specifically repress induction of the RIG-I/MDA5 pathway.
Based on our observation that the host type I IFN response requires the L. pneumophila Dot/Icm type IV secretion system and was at least partly Ips-1, Rig-i, and Mda5-dependent, we hypothesized that L. pneumophila nucleic acids (RNA, DNA or both) might gain access to the macrophage cytosol via the type IV secretion system and induce a host type I interferon response. To test if L. pneumophila nucleic acids are sufficient to induce type I interferon, we transfected MyD88−/−Trif−/− macrophages with purified L. pneumophila genomic DNA or total RNA and determined the induction of type I interferons by bioassay. Poly(dA-dT):poly(dA-dT) (abbreviated as pA:T) was used as a non-CpG containing DNA control and poly I:C was used as an RNA control. Nucleic acid preparations were treated with DNase and/or RNase to eliminate contaminating nucleic acids. Both purified L. pneumophila DNA and the crude RNA preparation induced IFNβ (Figure 5A). L. pneumophila RNA treated with RNase also induced IFNβ, presumably due to (contaminating) DNA in the preparation (Figure 5A). However, L. pneumophila RNA treated with DNase induced type I interferon to a level above that induced by L. pneumophila RNA treated with both RNase and DNase, suggesting that L. pneumophila RNA alone can induce type I interferon production (Figure 5A). The induction of type I IFN by L. pneumophila RNA was modest, possibly because bacterial RNA is less stable than DNA. Nevertheless, these results suggest that both L. pneumophila RNA and DNA can induce a type I interferon host response.
Next, we determined if L. pneumophila nucleic acids could induce type I interferon in an Ips-1-dependent manner in macrophages. In certain cell types, though not mouse macrophages [34], AT-rich DNA has been shown to induce type I IFN via IPS-1 [26],[27],[28],[29]. It was important to assess whether L. pneumophila DNA, in particular, might signal in an Ips-1-dependent manner since the L. pneumophila type IV secretion system has previously been shown to translocate DNA [46]. Ips-1+/− and Ips-1−/− macrophages were transfected with pA:T and L. pneumophila DNA, as well as infected with Sendai virus, a virus previously determined to induce an Ips-1-dependent IFN response. Stimulation with pA:T or L. pneumophila DNA failed to induce Ifnb in an Ips-1-dependent manner, whereas Sendai virus induced significantly more Ifnb in Ips-1+/− versus Ips-1−/− macrophages (Figure 5B). Similar results were obtained in Mda5−/− macrophages: induction of type I IFN with pA:T or L. pneumophila genomic DNA showed no requirement for Mda5, whereas a control simulation, Theiler's Virus, showed Mda5-dependent induction of IFNβ, as expected (Figure 5C). It was possible that at high concentrations of DNA, an Ips-1-independent DNA-sensing pathway overwhelmed any putative Ips-1-dependent recognition of DNA. However, induction of Ifnb was independent of Ips-1 even when titrated amounts of pA:T or L. pneumophila genomic DNA were transfected into macrophages (Figure 5D, 5E). Thus, these results suggest that while transfected L. pneumophila DNA robustly induces type I interferon, L. pneumophila genomic DNA does not appear to induce the Ips-1-dependent IFN response that is characteristic of L. pneumophila infection.
To determine whether L. pneumophila RNA could be recognized by Rig-i, we transfected L. pneumophila RNA into macrophages in which Rig-i expression had been stably knocked down. Importantly, the Rig-i knockdown was performed in immortalized bone-marrow-derived macrophages that lack MyD88 and Trif, in order to avoid potential activation of known RNA-sensing TLRs. Knockdown of Rig-i was effective under our transfection conditions, as Rig-i message was significantly lower in macrophages transduced with a Rig-i shRNA compared to a control shRNA (p<0.05; Figure 6A). Crude L. pneumophila RNA (which also contains genomic DNA contaminants) induced Ifnb robustly in both control shRNA and Rig-i shRNA macrophages, even upon treatment with RNase A (Figure 6B). However, transfection of DNase-treated L. pneumophila nucleic acids induced significantly less Ifnb in Rig-i knockdown macrophages as compared to control knockdown macrophages (p<0.05; Figure 6B.) This result suggests that L. pneumophila RNA can induce Rig-i-dependent type I interferon. It was not possible to perform a similar experiment in the Ips-1−/− macrophages because these macrophages were MyD88/Trif+ and exhibited background interferon, presumably due to TLR3 signaling.
A recent report found that an inhibitor of RNA polymerase III, ML-60218 [49], blocked the type I IFN response to L. pneumophila [29]. It was proposed that L. pneumophila DNA is translocated into macrophages and transcribed by Pol III into a ligand that could be recognized by RIG-I [29]. In contrast, we did not see an effect of ML-60218 on induction of type I IFN by L. pneumophila in bone marrow-derived macrophages (Figure 7A). The lack of an effect does not appear to be due to redundant recognition by another DNA sensor in macrophages because the interferon induction was still largely Ips-1-dependent (Figure 7A). Because our results with the Pol III inhibitor were negative, we cannot rule out the possibility that the Pol III inhibitor fails to function in macrophages. However, we also tested 293T cells, which express only the Pol III pathway for cytosolic recognition of DNA [28],[29]. As expected, 293T cells responded to pA:T in an ML-60218-inhibitable manner, but did not respond well to L. pneumophila genomic DNA (Figure 7B), again suggesting that L. pneumophila genomic DNA is not an efficient substrate for the Pol III pathway. The Pol III inhibitor also appeared to have little effect on L. pneumophila replication in bone-marrow macrophages (Figure 7C–E). This latter result was expected, since we found that even Ips-1−/− macrophages exhibit normal restriction of L. pneumophila replication (Figure S1), despite significantly reduced IFN induction.
In order to validate our findings in vivo, we infected Ips-1−/− and littermate Ips-1+/− mice with L. pneumophila (2.5×106 LP01 ΔflaA per mouse, infected intranasally) and assayed type I interferon production in bronchoalveolar lavage fluid 20 hours post infection by bioassay. Ips-1+/− mice induced an IFN response that was statistically significantly greater than the response of Ips-1−/− mice (Student's t-test, p = 0.01; Figure 8A). The difference in IFN production was not explained by a difference in bacterial burden in the Ips-1+/− and Ips-1−/− mice, since both genotypes exhibited similar levels of bacterial colonization (p = 0.76, Student's t-test; Figure 8B). The lack of an effect of Ips-1-deficiency on bacterial replication in vivo was not surprising given that we also failed to observe an effect of Ifnar-deficiency on bacterial replication in vivo (data not shown). We suspect that type II IFN (IFNλ), which is not made by macrophages in vitro, or another in vivo pathway, may compensate for loss of type I IFN in vivo. Nevertheless, our results provided an important validation of our in vitro studies and affirm a role for Ips-1 in the in vivo type I interferon response to L. pneumophila. Since Ips-1-deficient mice still mounted a measurable IFN response in vivo, it appears that additional Ips-1-independent pathways (e.g., TLR-dependent pathways, possibly involving other cell types [50]) also play a role in vivo.
Type I interferons (IFNs) have long been appreciated as critical players in antiviral immune defense, and recent work has identified several molecular immunosurveillance pathways that induce type I IFN expression in response to viruses [18],[19]. In contrast, the roles of type I IFNs in response to bacteria, and the pathways by which bacteria induce type I IFNs, are considerably less well understood. In this study, we sought to characterize the type I IFN response to the gram-negative bacterial pathogen Legionella pneumophila. Our study focused on the type I IFN response mounted by macrophages, since this is the cell type that is believed to be the primary replicative niche in the pathogenesis of Legionnaires' Disease.
In agreement with previous work [15], we found that L. pneumophila induces type I IFNs in macrophages via a TLR-independent pathway that requires expression of the bacterial type IV secretion system. These results suggested that a cytosolic immunosurveillance pathway controls the IFN response in macrophages. In this report we identify the cytosolic RNA-sensing pathway as a key responder to L. pneumophila infection (Figure 1) and, in agreement with previous results using human A549 cells [51], we did not observe a role for Dai (Zbp1), a gene implicated in the response to cytosolic DNA [30],[31]. A previous study using RNA interference in the human A549 epithelial-like cell line also found a role for IPS-1 in the type I IFN response to L. pneumophila [14]. However, knockdown of RIG-I or MDA5 did not appear to affect the IFN response [14], so the role of the IPS-1 pathway was unclear. In our study, we used mice harboring targeted gene deletions to establish a role for Mda5 and Ips-1 in the type I IFN response to L. pneumophila in macrophages, and uncovered a role for Rig-i using an shRNA knockdown strategy. We also found that the cytosolic RNA-surveillance pathway regulated the IFN response in vivo in a mouse model of Legionnaires' Disease.
After our manuscript was submitted, a report published by Chiu and colleagues also concluded that Ips-1 is required for the macrophage type I IFN response to L. pneumophila [29]. However, the report of Chiu et al differs considerably from our current work by proposing that the type I IFN response to L. pneumophila occurs via a novel and unexpected pathway in which L. pneumophila DNA reaches the host cytosol and is transcribed by RNA polymerase III to generate an RNA intermediate that is sensed by RIG-I. Others have found that the Pol III pathway can be activated by viral and AT-rich DNA in certain cell types [28]. Our data, however, are not easily reconciled with a role for the Pol III pathway in recognition of L. pneumophila. First, and perhaps most important, is the observation that the response to DNA (in contrast to the response to L. pneumophila infection) has never been seen to be Ips-1-dependent in macrophages ([34]; Figure 5). This suggests that the response to L. pneumophila is not simply a response to DNA, regardless of the mechanisms by which potentially translocated DNA might be recognized.
We considered the possibility that L. pneumophila DNA exhibits unique properties that cause it to be a particularly efficient substrate for the Pol III pathway. Indeed, the L. pneumophila genome does contain stretches of highly AT-rich DNA, and it has been reported that only highly AT-rich DNA is an efficient substrate for the Pol III pathway [28],[29]. Therefore we tested whether L. pneumophila genomic DNA, unlike other DNA, could induce an Ips-1-dependent response in macrophages. Although L. pneumophila DNA induced a robust IFN response, the response was not Ips-1-dependent (Figure 5B, E). Indeed, even the optimal Pol III substrate poly(dA–dT):poly(dA–dT) (abbreviated as pA:T) does not appear to induce an Ips-1-dependent IFN response in macrophages (Figure 5B, D and [34]). The lack of Ips-1-dependence in the response to pA:T appears to be due to an unidentified Ips-1-independent DNA-sensing pathway that recognizes pA:T and dominates over the Pol III pathway in bone marrow macrophages [28]. Thus, if translocated DNA is the relevant bacterial ligand that stimulates the Ips-1-dependent host type I IFN response, an explanation is required for how the dominant and unidentified DNA-sensing pathway is not activated. While L. pneumophila could selectively inhibit or evade the dominant DNA-sensing pathway, there is at present no evidence to support this mechanism. Moreover, in our hands, the Pol III inhibitor used by Chiu et al (ML-60218) failed to affect IFN induction or bacterial replication in macrophages (Figure 7), in contrast to what would be predicted if the Pol III pathway was selectively activated in response to L. pneumophila infection. Therefore, our data lead us to consider alternative models.
Although Chiu et al primarily used the RAW macrophage-like cell line in their experiments with L. pneumophila, we do not believe that cell-type-specific effects can account for the discrepancy in results. Although it is possible that RAW cells express only the Pol III pathway, this would not change the fact that the proposed model of Chiu et al invokes DNA as the primary IFN-inducing ligand produced by L. pneumophila. The simplest prediction of such a model would be that the response of bone marrow macrophages to L. pneumophila would be Ips-1-independent, as is the response of macrophages to all forms of DNA that have been tested. In contrast, as documented here (Figure 1) and by Chiu et al [29], the response to L. pneumophila is Ips-1-dependent. Moreover, 293T cells, which express only the Pol III DNA-sensing pathway [28],[29], failed to respond significantly to L. pneumophila genomic DNA, despite a robust response to pA:T (Figure 7B). Therefore, our data suggest that recognition of L. pneumophila genomic DNA by Pol III is not responsible for the Ips-1-dependent IFN response to L. pneumophila.
We considered two other models to explain how L. pneumophila induces a type I interferon response. The first is that L. pneumophila translocates RNA into host cells. In support of this model, we demonstrate that L. pneumophila RNA, unlike any form of DNA tested, induced a Rig-i-dependent type I IFN response in macrophages (Figures 5A, 6). However, we did not demonstrate that L. pneumophila RNA species are translocated into host cells, and this will be important to examine in future studies. Interestingly, it was recently reported that purified Helicobacter pylori RNA stimulates RIG-I in transfected 293T cells [52]. A second model to explain type I IFN induction by L. pneumophila is that infection induces a host response that indirectly results in signaling via the MDA5/RIG-I/IPS-1 pathways. L. pneumophila secretes a large number of effectors into the host cytosol and these effectors disrupt or alter a large number of host cell processes [53]. Such disruption may either lead to the generation of host-derived RNA ligands for the RIG-I and MDA5 sensors, or may result in signaling through these sensors in the absence of specific ligands. It was previously proposed that a host nuclease, RNaseL, can generate self-RNA ligands for the RIG-I and MDA5 pathways in response to viral infection [54]. Although we could not observe a role for RNaseL in the response to L. pneumophila (K.M. Monroe, unpublished data), it is conceivable that a different host enzyme can fulfill a similar function.
Our finding that a secreted bacterial effector, SdhA, previously shown to suppress host cell death, also suppresses the IFN response to L. pneumophila, is consistent with a model in which a host cell stress response leads to direct or indirect activation of the cytosolic RNA-sensing pathway. However, the mechanism by which SdhA acts on host cells remains mysterious. Laguna and colleagues provided evidence that SdhA is critical for prevention of mitochondrial disruption that occurs when host cells are infected with the ΔsdhA mutant [4]. Given that Ips-1 localizes to mitochondria and requires mitochondrial localization for its function [21], it is tempting to speculate that SdhA acts on mitochondria in a way that both prevents their disruption and interferes with the function of Ips-1. To provide evidence that SdhA acts specifically on the RIG-I/MDA5 pathway, we used transient transfections of 293T cells. SdhA repressed induction of Ifnb when co-expressed with Mda5 or Rig-I but not Trif (Figure S3). Given these results and the evidence that SdhA is translocated into host cells [4], we favor the idea that SdhA acts within host cells. Mutation of sdhA was reported not to affect translocation of other effectors into host cells [4]; thus, we tend not to support the alternative possibility that SdhA blocks translocation of the putative IFN-stimulatory ligand through the type IV secretion system. SdhA is a large protein of 1429 amino acids, but does not contain domains of known function, except for a putative coiled coil (a.a. 1037–1068). In future studies it will be important to address whether subdomains of SdhA can be identified that are required for suppression of the IFN response. It will also be important to determine whether these subdomains are distinguishable from any putative subdomains required for suppression of host cell death. In fact, our data have suggested that suppression of cell death and the IFN response may be separable functions of SdhA. We found that cell death was not required for hyperinduction of IFN by the ΔsdhA mutant, and conversely, we also found that hyperinduction of type I IFN does not lead to increased cell death (Figure 4).
Our studies demonstrate a partial role for both Mda5 and Rig-i RNA sensors in response to L. pneumophila. Although these sensors are typically thought to respond to distinct classes of viruses, there are indications that they can also function cooperatively in response to certain stimuli, e.g., West Nile Virus [47]. Our results suggest that L. pneumophila produces ligands that can stimulate both Mda5 and Rig-i and that these two sensors cooperatively signal via Ips-1. Fitting with this model, we found that Ips-1-deficiency generally had a more severe impact on type I IFN induction than did Mda5 or Rig-i deficiency.
Cytosolic RNA-sensing pathways are believed to respond exclusively to viral infection, and it is therefore surprising that L. pneumophila appears to trigger these pathways. Other bacterial species, such as Listeria monocytogenes and Francisella tularensis, have been shown to induce an Ips-1-independent cytosolic pathway leading to type I IFN induction [25],[43],[55]. The sensor(s) required for the IFN response to Listeria or Francisella have not yet been identified, but are widely assumed to be identical to the (also unknown) sensor(s) that respond to cytosolic DNA [15],[26].
Ips-1 or Mda5-deficiency, as well as Rig-i knockdown, did not result in a complete elimination of the type I IFN response (Figure 1, Figure 3). Thus, a cytosolic DNA-sensing pathway may also be stimulated in response to L. pneumophila infection. A minor role for a cytosolic DNA-sensing pathway would be consistent with the observation that the L. pneumophila Dot/Icm type IV secretion system can translocate DNA into recipient cells [46]. However, as discussed above, our results with purified L. pneumophila DNA suggest that cytosolic sensing of L. pneumophila DNA does not account for the Ips-1-dependent induction of IFN that we observe (Figure 5). One last possibility that we cannot eliminate is that a non-DNA, non-RNA ligand is translocated into host cells and stimulates the Ips-1 pathway. In fact, in separate work, we have found that a small bacterial cyclic dinucleotide, c-di-GMP, can trigger a type I IFN response in macrophages, but importantly, this response is entirely independent of the Ips-1 pathway [34]. Nevertheless, there may be other small molecules that can be translocated by the Dot/Icm secretion system and signal in host cells via Ips-1.
Taken together, our results lead to new insights into the host immunosurveillance pathways that provide innate defense against bacterial pathogens. We demonstrate an unexpected role for a viral RNA-sensing pathway in the response to L. pneumophila, and identify a secreted bacterial effector, SdhA, that can suppress this response. Our results therefore open new possibilities for immunosurveillance of bacterial pathogens.
Animal experiments were approved by the University of California, Berkeley, Institutional Animal Care and Use Committee.
Bone marrow derived macrophages were derived from the following mouse strains: C57BL/6J (B6), Ips-1−/− [25], Mda5−/− [56], Ifnar−/− [57], Zbp1−/− [32], MyD88/Trif−/−, and Casp1−/− [58]. C57BL/6J mice were purchased from the Jackson Laboratory. Ips-1−/− mice were from Z. Chen (University of Texas Southwestern Medical Center). Ips-1−/− were obtained on a mixed B6/129 background and Ips-1−/− and Ips-1+/− littermate controls were generated by breeding (Ips-1−/− x B6) F1 mice to Ips-1−/−. Mda5−/− mice were from M. Colonna and S. Gilfillan (Washington University). L929-ISRE IFN reporter cells were from B. Beutler (The Scripps Research Institute). Viruses to immortalize MyD88−/−Trif−/− immortalized bone marrow derived macrophages were the generous gift of K. Fitzgerald, D. Golenbock (U. Mass, Worcester) and D. Kalvakolanu (U. Maryland). The complementation plasmid (pJB908-SdhA) was generously provided by R. Isberg (Tufts). Expression constructs pEF-BOS-RIG-I and pEF-BOS-MDA5 were generously provided by J. Jung (Harvard Medical School).
LP02 is a streptomycin-resistant thymidine auxotroph derivative of Legionella pneumophila strain LP01. LP02ΔsdhA and LP02ΔsdhAΔsdhBΔsidH were a generous gift from R. Isberg (Tufts University). The ΔflaAΔsdhA strain was generated by introducing an unmarked deletion of flaA in LP02ΔsdhA using the allelic exchange vector pSR47S-ΔflaA [12].
L929-ISRE and HEK293T cells were cultured in DMEM supplemented with 10% FBS, 2 mM L-glutamine, 100 µM streptomycin, and 100 U/mL penicillin. Macrophages were derived from bone marrow cells cultured for eight days in RPMI supplemented with 10% FBS, 2 mM L-glutamine, 100 µM streptomycin, 100 U/mL penicillin, and 10% supernatant from 3T3-CSF cells, with feeding on the fifth day of growth. MyD88−/−Trif−/− immortalized macrophages were cultured in RPMI supplemented with 10% FBS, 2 mM L-glutamine, 100 µM streptomycin, and 100 U/mL penicillin.
Poly I:C was from GE Biosciences, pA:T (poly(dA-dT):poly(dA-dT)) was from Sigma, and Sendai Virus was from Charles River Laboratories. Wildtype Theiler's Virus GDVII was from M. Brahic and E. Freundt (Stanford University). Pol III inhibitor (ML-60218) was from Calbiochem.
Total bacterial RNA was isolated using RNAprotect Bacterial Reagent (Qiagen) and RNeasy kit (Qiagen). Genomic DNA was isolated by guanidinium thiocyanate followed by phenol:chloroform extraction. Nucleic acids were treated with RQ1 RNase-Free DNase (Promega) and/or RNaseA (Sigma).
Bone marrow derived macrophages were plated at a density of 2×106 per well in 6 well plates and infected with an MOI of 1. Macrophage RNA was harvested 4 hours post infection and isolated with the RNeasy kit (Qiagen) according to the manufacturer's protocol. RNA was DNase treated with RQ1 RNase-Free DNase (Promega) and reverse transcribed with Superscript III (Invitrogen). Quantitative PCR assays were performed on the Step One Plus RT PCR System (Applied Biosystems) with Platinum Taq DNA polymerase (Invitrogen) and EvaGreen dye (Biotium). Gene expression values were normalized to Rps17 (mouse) or S9 (human) levels for each sample. The following primer sequences were used: mouse Ifnb, F, 5′-ATAAGCAGCTCCAGCTCCAA-3′and R, 5′-CTGTCTGCTGGTGGAGTTCA-3′; mouse Rps17, F, 5′-CGCCATTATCCCCAGCAAG-3′ and R, 5′- TGTCGGGATCCACCTCAATG-3′; mouse Rig-i, F, 5′-ATTGTCGGCGTCCACAAAG-3′ and R, 5′-GTGCATCGTTGTATTTCCGCA-3′, human Ifnb, F, 5′-AAACTCATGAGCAGTCTGCA-3′ and R, 5′- AGGAGATCTTCAGTTTCGGAG G-3′; human S9, F, 5′-ATCCGCCAGCGCCATA-3′ and R, 5′-TCAATGTGCTTCTGGGAATCC-3′.
Cell stimulants were transfected with Lipofectamine 2000 (LF2000, Invitrogen) according to the manufacturer's protocol. Nucleic acids were mixed with LF2000 in Optimem (Invitrogen) at a ratio of 1.0 µl LF2000/µg nucleic acid and incubated for 20 minutes at room temperature. The ligand-lipid complexes were added to cells at a final concentration of 3.3 µg/ml (96-well plates) and 1.0 µg/ml (6 well plates). For poly I:C, the stock solution (2.5 mg/ml) was heated at 55°C for 10 minutes and cooled to room temperature immediately before mixing with LF2000. Transfection experiments were incubated for 8 hours, unless otherwise stated. RIG-I, MDA5, TRIF and SdhA expression plasmids, along with an IFNβ-firefly luciferase reporter and TK-Renilla luciferase plasmids, were transfected with FuGENE 6 (Roche) according to the manufacturer's protocol. Nucleic acids were mixed with FuGENE 6 in Optimem at 0.5 µl/96 well and incubated for 15 minutes. Total transfected DNA was normalized to 200 ng per well using an empty pcDNA3 plasmid. Cells were stimulated 20 hours after transfection of expression plasmids.
Cell culture supernatants or bronchoalveolar lavage fluid (BALF) was overlayed on L929-ISRE IFN reporter cells in a 96-well plate format and incubated for 4 hours at 37°C and 5%CO2. L929-ISRE IFN reporter cells and HEK293T cells expressing an IFNβ-firefly luciferase reporter and TK-Renilla luciferase were lysed in Passive Lysis Buffer (Promega) for 5 minutes at room temperature and relative light units were measured upon injection of firefly luciferin substrate (Biosynth) or Renilla substrate with the LmaxII384 luminometer (Molecular Devices). For transient transfection reporter assays, luciferase values were normalized to an internal Renilla control.
Cytotoxicity of bacterial strains was determined by measuring lactate dehydrogenase release essentially as previously described [59]. Macrophages were plated at a density of 1×105 in a 96-well plate and infected with stationary phase L. pneumophila at a multiplicity of infection (MOI) of 1. Plates were spun at 400×g for 10 minutes to allow equivalent infectivity of non-motile and motile strains [12]. Plates were re-spun 4 hours post infection and cell culture supernatants were assayed for LDH activity. Specific lysis was calculated as a percentage of detergent lysed cells.
Bacterial growth was determined as previously described [16]. Bone marrow derived macrophages were plated at a density of 1×105 per well in white 96-well plates (Nunc) and allowed to adhere overnight. Macrophages were infected with stationary-phase L. pneumophila at a multiplicity of infection (MOI) of 0.01. Growth of luminescent L. pneumophila strains was assessed by RLU with the LmaxII384 luminometer (Molecular Devices). Nonluminescent bacterial strains were analyzed for colony-forming units on buffered charcoal yeast extract plates.
Transposon mutagenesis of LP02 was previously described [12]. Briefly, the pSC123 mariner transposon was mated from E.coli SM10 λpir into the L. pneumophila strain LP02. Matings were plated on buffered yeast extract charcoal plates with streptomycin (100 µg/ml) and kanamycin (25 µg/ml). Single colonies were isolated and grown in overnight cultures and used to infect bone marrow derived MyD88−/−Trif−/− macrophages. After overnight incubation, levels of type I interferon in the supernatant was determined by bioassay. The site of transposon insertion was determined by Y-linker PCR [60].
Age and sex-matched Ips-1−/− and littermate Ips-1+/− mice were infected intranasally with 2.5×106 LP01 ΔflaA in 20 µl PBS. Bronchoalveolar lavage was performed 20 hours post infection via the trachea using a catheter (BD Angiocath 18 g, 1.3×48 mm) and 800 µl PBS. Type I interferon induction was determined by bioassay. Type I interferon amounts were calculated using a 4-parameter standard curve determined by dilution of recombinant IFNβ (R&D Systems). CFUs were determined by hypotonic lysis of cells from the brochoalveolar lavage fluid (BALF). In parallel experiments, it was determined that CFU in the BALF was representative of total CFU in the lung.
Knockdown constructs were generated with the MSCV/LTRmiR30-PIG (LMP) vector from Open Biosystems. shRNA PCR products were cloned into the LMP vector using XhoI and EcoRI sites. Rig-i sequence: 5′-GCCCATTGAAACCAAGAAATT-3′, control shRNA sequence: 5′-TGACAGTGTCTTCGCTAATGAA-3′. MyD88−/−Trif−/− immortalized bone marrow derived macrophages were transduced with retrovirus as previously described [10]. GFP+ macrophages were sorted with the DAKO-Cytomation MoFlo High Speed Sorter.
|
10.1371/journal.pgen.1004693 | RNA-Processing Protein TDP-43 Regulates FOXO-Dependent Protein Quality Control in Stress Response | Protein homeostasis is critical for cell survival and functions during stress and is regulated at both RNA and protein levels. However, how the cell integrates RNA-processing programs with post-translational protein quality control systems is unknown. Transactive response DNA-binding protein (TARDBP/TDP-43) is an RNA-processing protein that is involved in the pathogenesis of major neurodegenerative diseases, including amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD). Here, we report a conserved role for TDP-43, from C. elegans to mammals, in the regulation of protein clearance via activation of FOXO transcription factors. In response to proteotoxic insults, TDP-43 redistributes from the nucleus to the cytoplasm, promoting nuclear translocation of FOXOs and relieving an inhibition of FOXO activity in the nucleus. The interaction between TDP-43 and the FOXO pathway in mammalian cells is mediated by their competitive binding to 14-3-3 proteins. Consistent with FOXO-dependent protein quality control, TDP-43 regulates the levels of misfolded proteins. Therefore, TDP-43 mediates stress responses and couples the regulation of RNA metabolism and protein quality control in a FOXO-dependent manner. The results suggest that compromising the function of TDP-43 in regulating protein homeostasis may contribute to the pathogenesis of related neurodegenerative diseases.
| TDP-43 is linked to pathogenesis of major neurodegenerative diseases, including amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD). How TDP-43 contributes to the development of these degenerative diseases remains unsolved, and the full range of TDP-43 functions has yet to be established. In the present study, we explored a conversed function of TDP-43 in regulating protein homeostasis from C. elegans to mammals. Under conditions of stress, TDP-43 translocates from the nucleus to the cytoplasm, competes with FOXO transcription factors for binding to 14-3-3 proteins, and releases FOXO for nuclear translocation and activation. These data are consistent with the ability of TDP-43 to regulate protein aggregation. Together the results provide important insight into the role of TDP-43 in stress responses and disease mechanisms. Since chronic stress is associated with neurodegenerative diseases, the TDP-43 switch could be kept in overdrive mode in these disorders, with its capacity to buffer further stress and maintain protein homeostasis being compromised. This mechanism also suggests that other RNA-processing proteins that exhibit similar stress-induced behavior may be coupled to other cellular pathways to provide coordinated reprogramming in stress responses.
| A defining feature of all living cells is the ability to adapt to stress stimuli. This adaptive response is particularly important for maintaining protein homeostasis, which is critical for cellular functions. The cell employs a variety of protein quality control mechanisms in an effort to maintain the integrity of the proteome, including those regulating protein synthesis and degradation. The regulation of protein synthesis occurs at multiple levels, including transcription [1], RNA processing [2], and translation initiation [3]. Global attenuation of protein synthesis is often part of stress responses, and RNAs and RNA-processing proteins are central players in this adaptation [4]. Meanwhile, coordinated protein quality control systems are activated to enhance the degradation of damaged proteins. For instance, endoplasmic reticulum (ER) stress activates the signal transduction pathway known as the unfolded protein response (UPR), which coordinates a general translational attenuation and a specific induction of quality control proteins, including molecular chaperones, in order to improve protein folding in the ER lumen [5]. However, the coordination between these distinct stress responses is not completely understood and may involve connected regulation at both RNA and protein levels.
TAR-DNA binding protein (TDP-43) is an RNA-binding protein that has been suggested to play a major role in the pathogenesis of amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD) [6]–[34]. Bearing features of a heterogeneous nuclear ribonucleoprotein (hnRNP), TDP-43 has well-characterized RNA-processing functions [35]–[37]. TDP-43 has been shown to regulate transcription [38], [39], RNA splicing [40], [41], mRNA stability [42], [43], and microRNA processing [44]. An increasing number of RNA-binding proteins have been implicated in ALS/FTD and related neurodegenerative diseases [45], including FUS (fused in sarcoma) [46], [47], hnRNPA2B1, and hnRNPA1 [48]. As is true for TDP-43, a pathological feature of these RNA-binding proteins is the formation of proteinaceous inclusions in patients' tissues. Another feature shared by these RNA-binding proteins is their redistribution during stress. Although primarily nuclear, they can be found in stress granules after diverse stimuli [48]–[54]. Although the role of TDP-43 in RNA processing is well established, the full range of TDP-43 function has yet to be understood. Recently, the ortholog of TDP-43 in C. elegans, TDP-1, was shown to negatively regulate proteotoxicity associated with protein misfolding, suggesting that the nematode protein plays a role in the regulation of protein homeostasis [55], [56]. This observation raises a question as to whether the RNA-processing function of TDP-43 is directly coupled to its ability to regulate protein homeostasis.
Despite extensive study of the complex pathways responsible for cellular stress responses, how the cell coordinates these different adaptive programs is not yet completely understood. In particular, the exact mechanisms by which RNA processing coordinates with other aspects of stress responses to maintain protein homeostasis remain unclear. Here, we present a mechanistic pathway through which TDP-43 couples RNA processing with active protein quality control during stress. Our results show that TDP-43 regulates the activities of FOXO transcription factors, which are orthologous to C. elegans DAF-16 and mediate expression of genes involved in longevity, stress resistance, and protein quality control [57]. The activation of FOXOs is switched on when TDP-43 responds to differential stress signals, undergoes nucleocytoplasmic translocation, and reconfigures its interacting partners. We propose that the regulation of FOXO by TDP-43 represents an important scheme for the cell to efficiently maintain protein homeostasis by exerting control at both the RNA and protein levels; compromising the function of this pathway may contribute to the pathogenesis of TDP-43-related diseases.
C. elegans lacking its sole TDP-43 ortholog, TDP-1, lives longer than wild-type (WT) controls (Figure 1A) [55], [56], and the underlying mechanism is not understood. The insulin and insulin-like growth factor (IGF) pathway is an evolutionarily conserved regulator of longevity from C. elegans to humans [58]. Reduced function of the insulin/IGF-1 receptor, DAF-2, significantly extends the lifespan by activating DAF-16, a transcription factor that controls the expression of aging-related and stress-resistance genes [59]. To determine whether TDP-1 functions in the DAF-2–DAF-16 pathway, we utilized hypomorphic or null alleles of these genes to perform an epistasis analysis. A double mutant, tdp-1(ok803lf);daf-2(e1370lf), exhibited a longer lifespan than did daf-2(e1370lf) alone (Figure 1B). There are further genetic interactions between the two genes on other phenotypes; the tdp-1;daf-2 double mutant had improved egg-laying and locomotion compared with the daf-2 mutant alone (Figure S1). These data suggest that TDP-1 acts in a pathway parallel to that of DAF-2 to influence longevity, although a crosstalk between TDP-1 and DAF-2 may still be possible. However, another double mutant, tdp-1(ok803lf);daf-16(mu86lf), completely abolished the longevity effect of tdp-1(ok803lf) alone (Figure 1A) [55], indicating that there is a genetic link between TDP-1 and DAF-16, in which DAF-16 lies downstream of TDP-1 in the regulation of lifespan (Figure 1C).
Correlating with increased lifespan in C. elegans lacking TDP-1, the mutant also shows enhanced clearance of misfolded proteins [55], [56]. Consistently, we found that the tdp-1(ok803lf) mutant reduced the aggregation of TDP-C25-YFP, a misfolded protein reporter expressed in C. elegans neurons (Figure 1D). To test whether DAF-16 is downstream of TDP-1, we generated a strain expressing TDP-C25-YFP and harboring the double mutant, tdp-1(ok803lf);daf-16(mu86lf). The daf-16 mutant reversed the reduction of protein aggregation conferred by the tdp-1 mutant (Figure 1D). This data suggests that TDP-1 regulates proteotoxicity via DAF-16.
Since DAF-16 is a transcription factor, we asked whether TDP-1 influences the expression of DAF-16's transcriptional targets. We performed quantitative RT-PCR to measure the mRNA levels of a panel of known DAF-16 target genes in tdp-1(ok803lf) mutants and WT controls. These DAF-16-regulated genes included stress-resistance genes such as the metallothioneins mtl-1 and mtl-2 as well as uncharacterized genes dao-4, dct-8, and dct-17. The results indicated that most of the tested DAF-16 target genes are significantly up-regulated in tdp-1(ok803lf) mutant C. elegans (Figure 1E).
Next we asked whether this up-regulation upon loss of TDP-1 is specific to DAF-16 target genes. Although the transcriptional profiles of tdp-1(ok803lf) mutants and WT controls indicates that there are more genes down-regulated than up-regulated in tdp-1(ok803lf) mutants (Figure S2A) [55], quantitative RT-PCR analysis shows that DAF-16 targets are specifically up-regulated (Figure S2B). Taken together, these data demonstrate that loss of tdp-1 produces a specific up-regulation of DAF-16 transcription factor activity.
Since DAF-16 is a major transcription factor of stress-resistance genes [59], and mammalian TDP-43 undergoes stress-induced localization changes [49]–[51], we hypothesized that C. elegans TDP-1 is involved in stress signaling. To determine whether TDP-1 undergoes stress-induced changes in neurons, we generated transgenic strains that expressed YFP-tagged TDP-1 under the control of the pan-neuronal snb-1 promoter. These animals exhibited severe locomotor defects similar to those that we have previously noted in transgenic C. elegans expressing human TDP-43 driven by the same neuronal promoter [60].
Since protein quality control is involved in the regulation of both lifespan and neurodegeneration, we investigated whether TDP-1 responds to proteotoxic stress. First, we observed that neuronal TDP-1 responds to heat shock stress, which is known to increase misfolded proteins. When the transgenic TDP-1-YFP strain was grown on solid or liquid medium at 20°C, the TDP-1-YFP protein was localized to neuronal nuclei (Figure 2A). However, when the strain was subjected to heat shock stress at 28°C for 16 hours, TDP-1-YFP migrated to the cytoplasm, and in a subset of neurons, the protein formed granular structures (Figure 2B). Next, we tested the effects of hypertonic stress, which has been shown to induce molecular crowding and protein damage [61]–[64]. When the strain was treated with 0.4 M NaCl in liquid medium, we again observed the nucleocytoplasmic translocation of TDP-1 and formation of granular structures (Figure 2C). To test the response of TDP-1 in a setting directly relevant to proteotoxicity-related neurodegeneration, we crossed the stable transgenic TDP-1-YFP C. elegans strain into a C. elegans model of ALS expressing human SOD1 with the G85R mutation. The SOD1-G85R mutant has a high propensity to misfold and aggregate in this model system [65]. In the double-transgenic strain expressing both TDP-1-YFP and SOD1-G85R, but not the single-transgenic strain, we observed a switch in localization of TDP-1-YFP from the nucleus to the cytoplasm, where it was localized to punctate granules similar to those observed under heat shock and hypertonic stress (Figure 2D). These TDP-1 puncta could be distinct RNA granules; alternatively, the presence of misfolded proteins, such as SOD1-G85R, may seed the aggregation of TDP-1. Taken together, these results demonstrate that TDP-1 responds to different types of proteotoxic stresses, suggesting that the regulation of lifespan by TDP-1 involves its function in stress signaling.
Next, we asked whether the observed regulation of DAF-16 by TDP-1 is conserved from C. elegans to humans. DAF-16 is the sole C. elegans ortholog of four mammalian FOXO members (FOXO1, FOXO3a, FOXO4, and FOXO6), with the first three showing a high degree of structural and regulatory similarity [66]. To determine whether TDP-43 regulates the transcriptional activity of FOXOs, we used a luciferase reporter under the control of forkhead responsive elements (FHRE-Luc) to measure the FOXO transcriptional activity in HEK293T human embryonic kidney cells. Co-expression of the FHRE-Luc reporter with a FOXO family member significantly boosted the luciferase signal, enhancing the sensitivity for measuring the activity of a particular FOXO transcription factor. Consistent with the up-regulation of C. elegans DAF-16 activity by loss of TDP-1, shRNA-mediated knockdown of endogenous TDP-43 in HEK293T cells significantly increased the FOXO transcriptional activity, as indicated by the increase in luciferase activity (Figure 3A–B). Conversely, ectopic expression of TDP-43 markedly decreased the transcriptional activity of all three FOXO family members (Figure 3C–D). Moreover, the effects of TDP-43 on FOXO transcriptional activity were dose-dependent, with increasing levels of TDP-43 causing further suppression of the FHRE-Luc reporter signal. This suppression of FOXO transcriptional activity was not due to a decrease in FOXO protein levels caused by the overexpression of TDP-43 (Figure 3E–F). Overexpression of TDP-43 alone did not significantly change the luciferase activity of the FHRE-Luc reporter, reflecting the fact that the low endogenous level of FOXOs is not sufficient for the assay (Figure S3) [67]. Given the primarily nuclear localization of TDP-43, these results suggest an inhibition of FOXO transcriptional activity by TDP-43 in the nucleus. Taken together, these results establish a regulation of FOXO by TDP-43 that is conserved from C. elegans to humans.
Since NaCl treatment induced the nucleocytoplasmic translocation and granule formation of C. elegans TDP-1, we asked whether this response to hypertonic stress is evolutionarily conserved, by examining the effect of NaCl treatment on human TDP-43 in HEK293T cells through immunofluorescent staining. Treatment of HEK293T cells with 0.2 M NaCl induced the translocation of TDP-43 from the nucleus to the cytoplasm, where it co-localized with the stress granule marker Ras GTPase-activating protein-binding protein 1 (G3BP) (Pearson's correlation coefficient>0.9) (Figure 4A). The redistribution of TDP-43 to cytoplasmic G3BP-positive structures with treatment of 0.2 M NaCl was similar to that observed after treatment with sorbitol, a known osmotic stressor and well-established inducer of stress granule formation [68]. To examine the dynamic redistribution of TDP-43 during hypertonic stress, we stained for TDP-43 after incubating HEK293T cells with 0.2M NaCl for various lengths of time (Figure S4). Changes in TDP-43 were observed within 15 min of NaCl treatment, with TDP-43 forming punctate structures in the nucleus. By 2 h, TDP-43 was observed in the cytoplasm in large stress granules co-localizing with G3BP.
Interestingly, we found that NaCl-stress-induced patterns of TDP-43 translocation and sequestration varied in a concentration-dependent manner. When HEK293T cells were treated with 0.3 M NaCl, TDP-43 underwent cytoplasmic translocation but was recruited to a previously undescribed, smaller type of granules that did not co-localize with G3BP (Pearson's correlation coefficient <0.4) (Figure 4A). We also examined the dynamic redistribution of TDP-43 with 0.3 M NaCl treatment (Figure S5). TDP-43 formed punctate granules in nucleus within 15 minutes, and by 30 minutes TDP-43 redistributed to the cytoplasm forming this type of small granules.
Next we explored whether the translocation of TDP-43 is a reversible process. After a 3-h treatment with sorbitol or NaCl, the stressors were washed off, and the cells were kept in normal medium. After 24 h, TDP-43 was completely translocated back to the nucleus, although G3BP-positive stress granules remained in the cytoplasm (Figure S6). These results indicate that TDP-43 responds to various stresses in a dynamic and reversible manner that is not always associated with stress granules.
Since redistribution of proteins in the cell is often associated with post-translational modifications, we examined how the phosphorylation of TDP-43 correlates with the changes in its localization induced by hypertonic stress. We used a phospo-TDP-43 Ser409/410-specific antibody to detect the phosphorylated TDP-43 in HEK293T cells by immunofluorescence microscopy (Figure 4B). In untreated HEK293T cells, unlike the unmodified TDP-43, the phosphorylated protein appeared in both the nucleus and cytoplasm. As has previously been observed for sorbitol [49], low hypertonic stress (0.2M NaCl) treatment resulted in the majority of the phosphorylated TDP-43 co-localizing with G3BP-positive stress granules in the cytoplasm (Pearson's correlation coefficient>0.9). In contrast, when cells were exposed to high hypertonic stress (0.3M NaCl), the phosphorylated TDP-43 was localized differently and produced an appearance similar to that of untreated cells: The phosphorylated TDP-43 was distributed throughout both the nucleus and cytoplasm and did not colocalize with the G3BP-positive stress granules (Pearson's correlation coefficient <0.4). These results indicate that although cytoplasmic translocation is a consistent feature of TDP-43 during stress responses, its localization patterns are determined by the type and strength of the stress signals.
To understand how TDP-43 could regulate the activity of FOXOs during stress responses, we investigated the protein interactions that link these two proteins. We co-transfected HEK293T cells with tagged versions of the TDP-43 and FOXO3a proteins and performed co-immunoprecipitation assays. Immunoprecipitation assays using TDP-43 as bait failed to pull down FOXO3a (Figure S7), indicating that there is no physical interaction between TDP-43 and FOXO3a.
Next we investigated whether TDP-43 and FOXOs are linked in their localization changes during stress responses. To easily visualize the localizations of FOXO proteins, we utilized two U2OS cell lines that stably expressed GFP-FOXO1 or GFP-FOXO3a. Interestingly, we observed a strong mutual exclusion in the nucleocytoplasmic compartmentalization of TDP-43 and FOXO proteins (Figure 5A–B). Under normal culture conditions, endogenous TDP-43 was primarily localized in the nucleus, and only less than 5% of cells had a significant fraction of the TDP-43 in the cytoplasm, with the percentage increasing with stress. When TDP-43 was in the nucleus, the GFP-tagged FOXOs were almost invariably in the cytoplasm. When TDP-43 was cytoplasmic, the majority of the cells showed translocation of the FOXO proteins to the nucleus. Under oxidative stress induced by H2O2 treatment, cytoplasmic translocation of TDP-43 and nuclear translocation of FOXOs both increased, and their exclusive spatial correlation was maintained.
We then asked whether the cytoplasmic translocation of TDP-43 specifically drives the nuclear translocation of FOXO proteins as a stress response. To address this question, we performed cellular fractionation assays using HEK293T cells co-transfected with tagged versions of the TDP-43 and FOXO3a proteins and treated with hydrogen peroxide, which consistently induces TDP-43 cytoplasmic translocation. Western blotting against a cytoplasmic marker (caspase-3) and a nuclear marker (PARP1) confirmed a clean separation of cytoplasmic and nuclear fractions. The expression of TDP-43 alone, without stress, did not induce any detectable change in the fractionation of FOXO3a. However, with 1 mM H2O2 treatment, which increases the cytoplasmic fraction of TDP-43, the expression of TDP-43 led to a pronounced shift in FOXO3a from the cytoplasm to the nucleus (Figure 5C–D). These results suggest that the redistribution of TDP-43 from the nucleus to the cytoplasm under stress drives FOXO proteins translocating from the cytoplasm to the nucleus, consistent with the observation of a strong mutual exclusion in the nucleocytoplasmic compartmentalization of TDP-43 and FOXO proteins. To test whether TDP-43 also regulates the function of FOXOs in a stress-dependent manner, we used the FHRE-Luc reporter to measure FOXO transcriptional activity in the absence or presence of TDP-43. In contrast to the unstressed condition in which the expression of TDP-43 significantly suppressed FOXO activities, the TDP-43 expression dramatically increased FOXO3a transcriptional activity under the H2O2 stress (Figure 5E). Thus TDP-43 plays a pronounced role in stress to promote the nuclear translocation and transcriptional activity of FOXO proteins.
To understand how TDP-43 regulates FOXO localization and activity without a physical interaction between the proteins, we asked whether there is another player that might mediate an indirect association between them. The 14-3-3 family of proteins emerges as a possible candidate because these proteins have been shown to be involved in a multitude of signaling pathways and have a diverse set of binding partners. FOXO1, FOXO3a, and FOXO4 all interact with 14-3-3 proteins [69]–[71], and TDP-43 has been shown to interact with 14-3-3 proteins in an RNA-dependent manner [43]. To determine whether 14-3-3 relays signals from TDP-43 to FOXO, we performed competitive co-immunoprecipitation assays to address whether 14-3-3 partners with TPD-43 and FOXO in mutually exclusive protein complexes (Figure 5F–H). For this purpose, we transfected HEK293T cells with different combinations of Myc-tagged TDP-43, Flag-tagged FOXO3a, and HA-tagged 14-3-3σ. With 1 mM H2O2 treatment, 14-3-3σ was able to pull down either TDP-43 or FOXO3a when 14-3-3σ was co-transfected with either of the two proteins (Figure 5F). However, when all three proteins were expressed, the level of FOXO3a was greatly reduced in the 14-3-3σ co-immunoprecipitates (Figure 5F). Similar competitive binding to 14-3-3σ was also observed between WT TDP-43 and FOXO1 (Figure 5G). These results demonstrate that TDP-43 and FOXOs compete for binding to 14-3-3.
To further examine the relationships among these three proteins, we studied the effect of mutant TDP-43 proteins lacking the nuclear localization signal (ΔNLS) or the nuclear export signal (ΔNES) on the interaction between14-3-3 and FOXO proteins using the same competitive co-immunoprecipitation assays described above. The ΔNLS and ΔNES mutations enhanced the relative enrichment of cytoplasmic and nuclear TDP-43, respectively (Figure S8). The compartmentalization of the mutants was not exclusive, since residuals of the ΔNLS and ΔNES mutants could be found in the nuclear and cytoplasmic fractions. Nevertheless, the ΔNLS and ΔNES mutants allowed us to address how localization of TDP-43 affects the competitive binding of FOXOs to 14-3-3 proteins. We transfected HEK293T cells with tagged versions of FOXO3a, 14-3-3σ, and ΔNLS or ΔNES TDP-43, and then performed co-immunoprecipitation experiments using 14-3-3 as the bait. The ΔNLS TDP-43 was more effective than ΔNES in interfering with the interaction between 14-3-3σ and FOXO3a, indicating that the cytoplasmic fraction of TDP-43 is capable of dissociating FOXO3a from its binding to 14-3-3. This result suggests that the competitive binding of TDP-43 to 14-3-3 occurs at least in part in the cytoplasm (Figure 5H).
FOXOs have been reported to positively regulate protein quality control systems, including proteasomes and autophagy [72]–[74]. The regulation of FOXOs by TDP-43 suggests that TDP-43 may be a regulator of protein quality control. To explore this possibility, we studied the effects of loss or gain of TDP-43 on protein aggregation using a previously established SOD1 solubility assay [75]. The assay relies on differential detergent extraction to separate insoluble protein aggregates from soluble low-molecular weight monomers and oligomers. The G85R mutant, but not WT SOD1, was found in the insoluble pellet, providing a sensitive reporter for protein aggregation. When TDP-43 was ectopically expressed in HEK293T cells, there was a significant increase in the level of insoluble G85R SOD1 aggregates, with no difference in the soluble level (Figure 6A). However, when TDP-43 was knocked down in HEK293T cells, there was a marked reduction in the insoluble aggregates of G85R SOD1, together with a decrease in the level of soluble mutant SOD1 (Figure 6B). The WT SOD1 protein level was not changed when TDP-43 was knocked down or overexpressed, suggesting that TDP-43 negatively regulates the turnover of misfolded proteins.
In the present study, we have described a conserved signaling pathway in which TDP-43 senses stress and regulates protein quality control. This pathway is mediated, at least in part, by the ability of TDP-43 to regulate FOXO transcription factors. Therefore, TDP-43, a known RNA binding protein, represents a link between protein quality control and RNA metabolism, indicating a novel layer of regulation of protein homeostasis imparted by RNA-processing proteins. We propose that TDP-43 mediates stress responses designed to maintain protein homeostasis by coordinating the attenuation of protein synthesis and the selective enhancement of protein quality control systems. First, the loss of TDP-43 from the nucleus and its localization to stress granules in response to cellular stress constitute an adaptive response to keep target mRNAs from active translation [76]. Second, the cytoplasmic translocation of TDP-43 promotes protein quality control, increasing the removal of misfolded proteins through the regulation of FOXOs described here. Thus, the connection between the two metabolic processes, RNA processing and protein quality control, represents a high-level regulation of protein homeostasis, accomplished through TDP-43 coordination.
The regulation of FOXO transcription factor activity by TDP-43 has not been previously described. In this regulation, TDP-43 acts as a stress response switch to control FOXO activities. When the cell is in the resting state, TDP-43 is predominantly nuclear and exerts negative control over the FOXO transcription factors (Figure 7A). Evidence for this model includes TDP-1's negative regulation of DAF-16 transcriptional activity in C. elegans (Figure 1) and TDP-43's negative regulation of FOXO transcription activity in mammalian cells (Figure 3). When the cells are exposed to stress, TDP-43 temporarily leaves the nucleus and relaxes its negative control of the FOXO transcription factors. TDP-43 does not appear to influence the levels of FOXO proteins, which are governed by other complex regulations. Instead the increased fraction of cytoplasmic TDP-43 competes with FOXOs for binding to 14-3-3 proteins and drives the nuclear translocation of the FOXOs, further enhancing their transcription activity (Figure 7B).
The regulation of FOXOs by TDP-43 is consistent with the changes we observed in the cell's protein quality control systems. DAF-16, the only ortholog of the FOXO transcription factors in C. elegans, promotes longevity by activating the transcription of stress-resistant genes, including molecular chaperones. Accordingly, the loss-of-function TDP-1 mutant has an increased lifespan that is DAF-16-dependent, and it also has reduced levels of misfolded proteins (Figure 1) [55], [56]. In mammalian cells, the activation of FOXO transcription factors, including FOXO1 and FOXO3a, induces autophagy [72]–[74], [77]; the activation of FOXO3a also promotes proteasome activity [73]. In accordance with the negative regulation of FOXOs by nuclear TDP-43, acute reduction of TDP-43 decreases the levels of misfolded and aggregated proteins (Figure 6), suggesting enhanced protein quality control.
The coupling of RNA regulation to protein quality control by TDP-43 may represent a general coordination feature among stress-adaptive programs. There are other RNA-processing proteins that exhibit similar switching behaviors during stress responses. For example, a number of hnRNP proteins translocate to the cytoplasm from the nucleus in response to stress and are sequestered in punctate structures such as stress granules. We propose that, like TDP-43, these RNA-processing proteins function as stress response switchers to maintain cellular homeostasis. The nucleocytoplasmic translocation and sequestration of these RNA-processing proteins into stress granules represent an adaptive loss of function that serves to temporarily curtail protein synthesis.
TDP-43's role in controlling stress response and protein homeostasis may have important implications for neurodegenerative diseases. Rather than a simple gain- or loss-of-function scenario, we propose that a mechanism involving the compromised function of TDP-43 acting as a stress response switch underlies the etiology and pathology of TDP-43-related degenerative diseases. The TDP-43 proteinopathy is characterized by the cytoplasmic accumulation and concomitant nuclear clearance of the non-mutated form of the TDP-43 protein [6]. This common pathology in related neurodegenerative diseases is likely a result of the response to chronic stress. Moreover, the present study suggests that there is a duality in the cellular effects of TDP-43 cytoplasmic translocation and nuclear clearance. As a stress response, this acute reduction in TDP-43 function is protective, defending the cell against proteotoxicity through the pathway delineated above. However, if chronic stress persists, the resulting long-term reduction in TDP-43 function is deleterious. Without a resetting of the switch, the capacity of TDP-43 to buffer further stress would be lost. Also, over-activation of FOXOs promotes senescence or cell death [78], suggesting that the activation of FOXOs by the TDP-43 switch may initiate a built-in program to eliminate over-stressed cells. This duality is analogous to that in ER stress, in which the UPR program has both a protective and a deleterious effect [5].
In conclusion, TDP-43 acts as a stress response switch to regulate RNA and protein metabolism in order to maintain protein homeostasis. The role of TDP-43 in the feedback regulation of the proteotoxic stress response and quality control provides a new perspective for TDP-43-related pathogenesis. Compromise of the stress response by proteins like TDP-43 may be a general mechanism underlying neurodegenerative diseases.
All C. elegans strains are on the N2 Bristol background and cultured under standard conditions at 20°C unless otherwise indicated. To generate the Psnb-1::TDP-1-YFP(iwIs53) strain, a transgene DNA construct was generated by subcloning TDP-1 cDNA with a C-terminal YFP tag into a modified plasmid, pPD30_38 (Fire Lab Vector, Addgene), with the promoter replaced with that of the snb-1 gene, as previous described [60]. The transgene DNA solution containing 20 ng/µl of the expression construct was injected into hermaphrodite gonads [79], and multiple extrachromosomal lines were established based upon the fluorescent markers. These lines were further treated with 30 µg/ml trimethylpsoralen (Sigma-aldrich) and 300 µJ of 365 nm UV light to screen for integrated lines that stably expressed the transgenes. Each integrated line was backcrossed with the N2 strain at least four times. The Psnb-1::TDP-C25-YFP(iwIs22) strain was reported previously [60]. Some strains were provided by the Caenorhabditis Genetics Center, which is funded by the NIH Office of Research Infrastructure Programs (P40 OD010440).
The WT N2 C. elegans and mutant strains RB929 [tdp-1(ok803)], CF1038 [daf-16(mu86)], CB1370 [daf-2(e1370)], IW417 [tdp-1(ok803); daf-16(mu86)], and IW177 [tdp-1(ok803);daf-2(e1370)] were cultured under standard conditions at 20°C. Newly laid embryos were synchronized within 2 h and transferred to fresh NGM plates, with 50 embryos per plate. These animals were transferred to new plates every day until they reached the post-reproductive stage and were allowed to age under normal culture conditions. Animals were checked daily and considered dead if they showed no response to probing with a platinum pick. The animals that crawled out of the plate, had vulval burst, or died with internally hatched larvae or “bags of worms” were censored. The day when embryos were synchronized was defined as the first day for lifespan analysis. The lifespan data were analyzed using Prism 5 software.
C. elegans were cultured at 20°C until they grow to L4 larval stage. L4 larvae were individually transferred to new plates and cultured at 25°C. Every 24 hours, these animals were transferred to new plates at 25°C until they stopped producing eggs. After transferring, the eggs laid on the plates were counted. For locomotion measurement, L4 larvae grown at 20°C were subject to a thrashing assay. The animals were transferred into M9 buffer (3 mg/ml KH2PO4, 6 mg/ml Na2HPO4, 5 mg/ml NaCl, and 1 mM MgSO4) and allowed to adapt to the buffer for 1 min. Then the rate of body bending or thrashes for the animals was measured, with a thrash being counted when both the head and the tail bend over 45 degrees.
To quantify the expression of specific genes in C. elegans, animals were harvested and total RNA was isolated using a phenol-chloroform extraction with TRIzol reagent (Life Technologies), followed by purification with an RNeasy mini kit (Qiagen). A two-step RT-PCR was employed to assess relative changes in transgenic transcripts using an iScript cDNA Synthesis Kit and an IQ SYBR Green kit (Bio-Rad). Fluorescence was measured on a real-time PCR cycler (Bio-Rad), and CT values were analyzed based on standard curves. The worm gdh-1 gene was used as a control. The sequences of all the primers used are listed in the Table S1.
HEK293T cells were grown in Dulbecco's modified Eagle's medium supplemented with antibiotics and 10% fetal bovine serum. Transient transfection of HEK293T was carried out with Lipofectamine 2000 according to the manufacturer's instructions (Life Technologies). The primers for subcloning TDP-43 cDNA into pRK5-myc vector at Sal site and Gateway vector (pDonor-221) are listed in the Table S1.
To knock down the expression of TDP-43, we generated shRNA constructs targeting different regions of the TDP-43 transcript and containing a puromycin resistance gene. The TDP-43 shRNA constructs were cloned by inserting small hairpin oligonucleotides matching TDP-43 mRNA sequences into the pRFP-C-RS plasmid (Origene) using BamHI/HindIII restriction sites. The sequence of shRNA oligonucleotides (i), (ii), and (iii) as well as a control shRNA-RFP-C-RS are listed in the Table S1. HEK293T cells were plated in 60-mm dishes at a density of 3.5×105 per well, then transfected with the shRNA constructs after 24 h. After 24 h, puromycin was applied at 3 µg/ml to select for positively transfected cells, and cells were harvested at 72 h post-transfection.
A reporter construct, pGL3-FHRE-Luc was originally from M. Greenberg (Addgene Plasmid 1789), which expresses the firefly luciferase driven by a promoter containing three copies of forkhead responsive elements, was employed to measure FOXO transcription activity [80]. A control reporter construct, pSV40-Renilla (Promega), which provides constitutive expression of Renilla luciferase, was used as an internal control. Cells were plated in 24-well plates at a density of 1×105 cells per well, and after 24 h, cells were transfected with expression plasmids of FOXOs, pSV40-Renilla, pGL3-FHRE-luciferase, or TDP-43. At 48 h, luciferase activities were measured by using the Dual-Luciferase Reporter Assay System (Promega) on a Synergy H1 luminometer (Bio-Tek). The experimental firefly luciferase activity was normalized to the control Renilla luciferase activity to reflect the FOXO activities.
In addition to HEK293T cells, two U2OS stable cells lines expressing FOXO1 or FOXO3a were used (Thermofisher). Cells were plated on coverslips pre-treated with polyethylenimine (Sigma-aldrich), in 6-well plates. Cells were stressed with 0.4 M sorbitol (Sigma-aldrich) for 1 h or 0.2 M (or 0.3 M) NaCl for 3 h. Coverslips were washed twice with 1× PBS and then fixed with 4% paraformaldehyde (PFA) for 10 min at RT. After the PFA was washed, the cells were permeabilized with 0.1% Triton X-100 in 1× PBS for 10 min. After treatment with blocking buffer containing 2% BSA in PBS with 0.1% Triton X-100 for 30 min, the coverslips were incubated with primary antibody at 4°C overnight. The primary antibodies were diluted in blocking buffer; they were: polyclonal anti-TDP-43, 1∶200 (10782-2-AP, ProteinTech); monoclonal anti-G3BP, 1∶200 (611126, BD Transduction Laboratories); and monoclonal anti-phospho TDP-43 (Ser409/410), 1∶200 (MABN14, Millipore). The coverslips were then incubated with secondary antibody for 1 h at RT: goat anti-rabbit Alexa Fluor 488, 1∶1000 (A11008, Life Technologies); goat anti-mouse Alexa Fluor 555, 1∶1000 (A21422, Life Technologies); or goat anti-mouse Alexa Fluor 555, 1∶1000 (A11006, Life Technologies). The coverslips were mounted in buffer with 2.5% DABCO, 100 mM Tris-HCl (pH 8.8), 50% glycerol, and 0.2 µg/ml DAPI. Images were collected with a Zeiss AxioObserver Z1 with an Apotome imaging system.
Cells were lysed in buffer containing 50 mM Tris-HCl (pH 8.0), 150 mM NaCl, 0.4 mM EDTA (pH 8.0), 1% NP-40, 0.05% sodium deoxycholate, and complete protease inhibitor cocktail (11836153001). Cell lysates were incubated with anti-HA antibody (H6908, Sigma-aldrich) overnight at 4°C before being centrifuged at 10,000× g for 10 min. Supernatant was transferred to clean tubes and incubated with protein G-Sepharose beads (17061801,GE Healthcare life sciences) for 1 h. After five washes with lysis buffer, the beads were resuspended in SDS sample buffer and boiled for 5 min before the eluted materials were subjected to standard western blot analysis: Protein samples were separated by SDS-PAGE and transferred to nitrocellulose membranes (Bio-Rad). The membranes were blocked with 5% milk in 1× phosphate-buffered saline with 0.1% Tween 20 (PBST) and incubated with the following primary antibodies: anti-c-myc- horseradish peroxidase, 1∶5000 (11814150001, Roche); monoclonal anti-c-myc, clone 9E10, 1∶3000 (M5546, Sigma-aldrich); monoclonal anti-Flag M2, clone M2, 1∶5000 (F3165, Sigma-aldrich); polyclonal anti-GAPDH,1∶30,000 (PA1-27448, Thermofisher); monoclonal anti-V5, 1∶3000 (460705, Life Technologies); polyclonal anti-TARDBP, 1∶3000 (10782-2-AP, ProteinTech), polyclonal anti-PARP, 1∶1000 (9542, Cell Signaling), polyclonal anti-Caspase 3, 1∶1000 (9662, Cell Signaling), and polyclonal anti-Cu/Zn SOD, 1∶3000 (ADI-SOD-100, Enzo life sciences). The following secondary antibodies were used: goat anti-rabbit IgG (H+L)-HRP conjugate, 1∶3000 (170–6515, Bio-Rad); goat anti-mouse IgG (H+L)-HRP conjugate, 1∶3000 (170–6516, Bio-Rad); goat anti-rabbit IgG IRDye, 1∶40,000 (680 LT, 926–68021 or 800 CW, 926–32211, LI-COR); and donkey anti-mouse IgG, 1∶40,000 (680 LT, 926–68022; or 800 CW, 926–32212, LI-COR). After incubating with secondary antibodies, the membranes were developed on films or the Odyssey image system (Li-COR).
To isolate cytoplasmic and nuclear fractions, cultured mammalian cells were harvested in 1× PBS, centrifuged at 200× g for 1 min, and washed twice with 1× PBS. The cell pellets were resuspended in a cytoplasmic extraction buffer containing 10 mM HEPES (pH 7.9), 10 mM KCl, 1.5 mM MgCl2, 0.1 mM EDTA, 0.5 mM DTT, 0.4% NP-40, 0.5 mM PMSF, and complete protease inhibitor cocktail, then incubated on ice for 5 min. Cell lysates were then centrifuged at 600× g at 4°C for 3 min. Supernatants were transferred to clean tubes and saved as the “cytoplasmic fraction.” The pellets were washed twice with cytoplasmic extraction buffer and centrifuged at 600× g at 4°C for 3 min. After washing, the pellets were resuspended in a nuclear extraction buffer containing 20 mM HEPES (pH 7.9), 420 mM NaCl, 1.5 mM MgCl2, 25% glycerol, 0.5 mM PMSF, 0.2 mM EDTA, 0.5 mM DTT, and complete protease inhibitor cocktail and vortexed at room temperature for 1 min. The resuspended samples were incubated on ice for 10 min, then vortexed for 1 min and centrifuged at 16,000× g for 10 min. The supernatants were transferred to clean tubes and saved as the “nuclear fraction.”
A biochemical assay was used to detect insoluble aggregated proteins according to a previously described protocol, with some modifications [60]. Mammalian cells or C. elegans were extracted in 200 µl of buffer containing 10 mM Tris-HCl (pH 8.0), 100 mM NaCl, 1 mM EDTA (pH 8.0), 0.5% NP-40, 50 µM iodoacetamine, and protease inhibitor (P8340, Sigma-aldrich) by using a Bioruptor ultrasonicator at 4°C for 5 min. The lysates were then transferred to an Airfuge ultracentrifuge (Beckman Coulter) and centrifuged at 25 psi (∼130,000 g) for 5 min.
The supernatant was transferred to clean tubes and saved as the “soluble” fraction. The remaining pellet was again resuspended in extraction buffer, then sonicated for 5 min. This resuspended pellet was applied to the Airfuge and centrifuged at 25 psi for 5 min. The remaining pellet was transferred to 100 µl of resuspension buffer containing 10 mM Tris-HCl (pH 8.0), 100 mM NaCl, 1 mM EDTA (pH 8.0), 0.5% NP-40, 0.5% deoxycholic acid, and 2% SDS, and sonicated for 5 min. This fraction was considered the “insoluble” protein aggregate fraction.
Image J software was used to analyze the colocalization between TDP-43/pTDP-43 and G3BP, and the colocalization was measured by Pearson's correlative coefficient. p values for comparing Kaplan-Merier survival curves between groups were calculated by the Log-rank test. p values of qPCR, egg-laying, locomotion and Luciferase assay data were calculated with the Student's t-test.
|
10.1371/journal.ppat.1005622 | EGFR Interacts with the Fusion Protein of Respiratory Syncytial Virus Strain 2-20 and Mediates Infection and Mucin Expression | Respiratory syncytial virus (RSV) is the major cause of viral lower respiratory tract illness in children. In contrast to the RSV prototypic strain A2, clinical isolate RSV 2–20 induces airway mucin expression in mice, a clinically relevant phenotype dependent on the fusion (F) protein of the RSV strain. Epidermal growth factor receptor (EGFR) plays a role in airway mucin expression in other systems; therefore, we hypothesized that the RSV 2–20 F protein stimulates EGFR signaling. Infection of cells with chimeric strains RSV A2-2-20F and A2-2-20GF or over-expression of 2–20 F protein resulted in greater phosphorylation of EGFR than infection with RSV A2 or over-expression of A2 F, respectively. Chemical inhibition of EGFR signaling or knockdown of EGFR resulted in diminished infectivity of RSV A2-2-20F but not RSV A2. Over-expression of EGFR enhanced the fusion activity of 2–20 F protein in trans. EGFR co-immunoprecipitated most efficiently with RSV F proteins derived from “mucogenic” strains. RSV 2–20 F and EGFR co-localized in H292 cells, and A2-2-20GF-induced MUC5AC expression was ablated by EGFR inhibitors in these cells. Treatment of BALB/c mice with the EGFR inhibitor erlotinib significantly reduced the amount of RSV A2-2-20F-induced airway mucin expression. Our results demonstrate that RSV F interacts with EGFR in a strain-specific manner, EGFR is a co-factor for infection, and EGFR plays a role in RSV-induced mucin expression, suggesting EGFR is a potential target for RSV disease.
| Respiratory syncytial virus (RSV) is responsible for severe lower respiratory disease in infants and young children. Overabundant airway mucus contributes to airway obstruction in RSV bronchiolitis, and a better understanding of RSV pathogenesis may contribute to needed therapies and vaccines. We reported previously that RSV clinical isolate strain 2–20 induces more airway mucin expression in mice than prototypic RSV strains and that the 2–20 fusion (F) protein mediates mucin induction. Epidermal growth factor receptor (EGFR) has been shown to play a role in lung mucin expression. We identified a functional interaction between 2–20 F and EGFR, in that 2–20 F expression activated EGFR and, reciprocally, EGFR expression increased 2–20 F fusion activity. RSV F and EGFR co-localized in infected cells. EGFR co-immunoprecipitated with RSV F protein from various RSV strains, and the strength of this in vitro interaction correlated with strain-specific airway pathogenicity in mice. EGFR inhibition abrogated 2–20 F-mediated infection in vitro and mucin expression induction in vivo. These data identify EGFR as a novel strain-specific co-factor of RSV infection and suggest EGFR may be a target for ameliorating RSV disease.
| Respiratory syncytial virus (RSV) is a human pathogen of the Pneumovirus genus within the Paramyxoviridae family. Worldwide, the virus causes over 30 million lower respiratory tract illnesses per year in children and is a leading cause of infant pneumonia mortality [1, 2]. Despite a substantial clinical burden of disease, there are no available vaccines or RSV-specific therapeutics. A challenge to RSV vaccine and therapy strategies remains elucidation of the unclear relationship between RSV infection and pathogenesis.
RSV is an enveloped, non-segmented, negative-strand RNA virus whose genome is approximately 15.2 kb in length and encodes 10 genes which are translated into 11 proteins. RSV attachment is mediated through host glycosaminoglycans (GAGs), cellular protein nucleolin, association with cholesterol-rich microdomains, and CX3CR1 [3–9]. Mechanisms surrounding RSV entry remain unclear and other host receptors, co-receptors, and co-factors contributing to infection are likely to be identified. Two envelope proteins mediate RSV infection, the attachment glycoprotein (G) and the fusion (F) protein. Prior to infection, RSV F exists in a metastable pre-fusion conformation [10, 11]. RSV F undergoes a series of conformational changes yielding a thermodynamically stable six-helix post-fusion bundle, which drives viral and host membrane fusion [11–13]. RSV G is mucin-like, having extensive N- and O-linked glycosylation, and G is responsible for facilitating RSV attachment through interactions with GAGs and CX3CR1 [4, 6, 7, 9, 14, 15]. However, G is not absolutely required for viral entry into immortalized monolayer cells [16–18]. Mechanisms by which F and G mediate host cell entry and their interactions with other host cell targets remain uncertain.
Epidermal growth factor receptor (EGFR) is a host glycoprotein comprised of an extracellular ligand receptor and intracellular kinase domain. The latter is activated through both Src-dependent phosphorylation and autophosphorylation [19, 20]. In addition to a wide variety of host ligands including epidermal growth factor (EGF) and transforming growth factor alpha (TGFα), several viruses have been identified that employ EGFR binding and activation for viral entry and replication. These pathogens include hepatitis B virus, human cytomegalovirus (hCMV), and Epstein-Barr virus (EBV) [21–23]. Previous studies by others evaluating the role of EGFR in RSV infection have shown that RSV activates EGFR in lung epithelial cells [24, 25]. EGFR activation in these cells promotes a pro-inflammatory response including increased survival of RSV-infected cells and suppression of interferon regulatory factor (IRF) 1-dependent CXCL10 production, an important event for recruitment of lymphocytes to infected airway epithelial cells [24, 25]. Another study using a recombinant virus based on the RSV subgroup A prototypic strain A2 demonstrated that RSV cell entry is largely mediated through endocytotic macropinocytosis promoted by EGFR phosphorylation [26].
Respiratory failure is the critical consequence of RSV disease in children, and overabundant mucus obstruction of the airways contributes to this outcome. Our laboratory previously reported that clinical isolate RSV A2001/2-20 (2–20) causes more airway necrosis, inflammation, and mucin expression during infection in BALB/cJ mice than the A2 reference strain [27, 28]. Transfer of the RSV 2–20 F protein into strain A2 recapitulated higher levels of airway mucin expression in mice [28]. These studies demonstrated that the RSV F protein plays a key role in airway epithelium infection and pathogenesis in vivo and suggests that RSV F plays a role in RSV strain-specific phenotypes. EGFR phosphorylation is known to play a role in mucin expression in airway epithelial cells during influenza and rhinovirus infections [29, 30]. We hypothesized that mucin induction by RSV 2–20 F is mediated by a specific interaction with EGFR. To test this hypothesis, we evaluated the ability of A2 and 2–20 viruses and transiently expressed F proteins to activate EGFR, and we assessed the impact of disrupting these interactions on virus infectivity in vitro and mucin expression in vivo. We demonstrate that RSV 2–20 F protein specifically binds to and activates EGFR, EGFR contributes to RSV-2-20 F infectivity, and EGFR signaling mediates 2–20 F induction of airway mucin expression in mice.
As EGFR phosphorylation/activation mediates mucin expression in response to other respiratory viruses, RSV A2 is known to activate EGFR, and RSV 2–20 F is more “mucogenic” than A2 F in the context of infection, we hypothesized that the 2–20 strain F protein potently activates EGFR. Western blotting was performed to determine levels of total EGFR and phospho-EGFR (p-EGFR) after infection of HEp-2 cells with RSV strains A2 and A2-2-20F. We used serum-starved cells because serum is an activator of EGFR signaling in vitro [31, 32]. A2-2-20F-infected cells had a higher ratio of p-EGFR to EGFR than A2-infected cells at 24 h post-infection (Fig 1A). We performed similar experiments in serum-starved NCI-H292 (H292) cells, and in an earlier time course we found that A2-2-20F infection resulted in a higher p-EGFR/EGFR ratio than A2 infection at 1, 12, and 24 hr post-infection (Fig 1B). To define the role of RSV 2–20 F expression alone in EGFR activation, A2 F and 2–20 F expression constructs were transfected into serum-starved HEp-2 cells and blotted for EGFR and p-EGFR levels. There was a significantly higher 4.45-fold ratio of p-EGFR to EGFR after 2-20F expression than after A2 F expression (Fig 1C). These data demonstrate that the 2–20 RSV F protein activated EGFR in cells to a greater extent than the A2 strain F protein.
We tested whether EGFR can enhance RSV F protein fusion activity. We used a previously established cell-cell fusion assay [28, 33, 34]. The dual-split protein (DSP) assay is based on co-transfecting an “effector”/cis population of 293T cells with a construct expressing N-terminal domains of a Renilla luciferase and GFP fusion protein (DSP1-7) and an F expression construct in the presence of specific fusion inhibitor. Another population of 293T cells (“target”/trans) is transfected with a construct expressing the C-terminal domains of the luciferase-GFP fusion protein (DSP8-11) and, in this case, either equal molar amounts of empty vector (pcDNA) or an EGFR expression vector. The fusion inhibitor is washed out, the effector and target cells are mixed, and fusion is quantified by luciferase activity reconstituted by cell content mixing. A2 F had significantly more cell-cell fusion activity than 2-20F in this assay (Fig 1D). Expression of EGFR in trans enhanced 2–20 F activity but not A2 F activity (Fig 1D). Similar to published A2 F and 2–20 F fusion assay experiments [28], we found no difference, as measured by flow cytometry, between A2 F and 2–20 F surface expression in 293T cells (S1 Fig). To determine whether the boost in RSV 2-20F fusion is specific to a trans F-EGFR interaction, 2-20F was either co-expressed (in cis) with EGFR or in trans in the target cells. There was a significant boost to 2-20F fusion when EGFR was expressed in trans but not in cis, suggesting EGFR enhancement of 2–20 F activity does not occur when the proteins are overexpressed in the same cells or membrane (Fig 1E). Co-expression of EGFR with RSV F did not alter F surface expression (S1 Fig). There was no effect on 2–20 F activity by overexpression of signaling lymphocyte activation molecules (SLAM), a receptor for measles virus (MeV) serving as an irrelevant transmembrane protein control. Collectively, the data show EGFR specifically enhanced the fusion activity of RSV strain 2–20 F protein but not A2 F protein.
EGFR can function as receptor, co-receptor, or entry co-factor for other viruses, and EGFR depletion in HeLa cells was reported to reduce RSV A2 strain infectivity [26]. We explored the role of EGFR in RSV-2-20F infection. Cells were pre-treated with EGFR tyrosine kinase inhibitors (AG1478 and PD153035) then infected with recombinant RSV A2, RSV A2-2-20F, or RSV A2-2-20GF. Virus infectivity was measured by flow cytometry of virally expressed mKate2. NCI-H292 (H292) cells are a human tracheal epithelial cell line known to express mucin genes through activation of the EGFR pathway and known to support RSV replication [35–37]. Treatment of H292 cells with increasing concentrations of EGFR inhibitor AG1478 resulted in a dose-dependent reduction in infectivity of all three virus strains at an MOI of 1, and the reduction in infection was greater against A2-2-20F and A2-2-20GF than against A2 (Fig 2A). Treatment of H292 cells prior to RSV A2 infection at a higher MOI = 3 with either AG1478 or PD153035 resulted in no change in infection (Fig 2B). In contrast, EGFR inhibition prior to A2-2-20F or A2-2-20GF infection at MOI = 3 reduced infection efficiency (Fig 2B). Treatment in BEAS-2B cells, a human bronchial epithelial cell line, with AG1478 or PD153035 resulted in similar decreases in infectivity of RSV A2-2-20F and A2-2-20GF while having no effect on A2 (Fig 2B). In order to test the effect of EGFR inhibition on RSV infectivity in a model more relevant to RSV biology, we utilized normal human bronchial epithelial cells differentiated at air-liquid interface (NHBE-ALI) [38]. We found that pre-treatment of NHBE-ALI cultures with 5 μM AG1478 was not toxic, as previously described, whereas PD153035 was toxic to NHBE-ALI [39]. Pre-treatment of NHBE-ALI cultures with AG1478 resulted in significant reduction in infectivity of RSV A2-2-20F and A2-2-20GF, and no significant effect on A2 compared to vehicle (Fig 2C). Taken together, EGFR signaling mediated infection by RSV expressing the 2–20 F protein.
To determine the role of EGFR expression in RSV infectivity, BEAS-2B cells were first transduced with lentivirus expressing either EGFR-specific shRNA or scrambled (control) shRNA. EGFR shRNA reduced expression of EGFR in BEAS-2B cells by 56% compared to the scrambled shRNA control (Fig 3A and 3B). shRNA knockdown of EGFR in BEAS-2B cells resulted in lower infectivity of A2-2-20F and A2-2-20GF but not A2 (Fig 3C). MeV infection is known to be EGFR-independent; therefore MeV was used as a control for RSV specificity [40]. The knockdown data show EGFR acted as a co-factor contributing to infectivity of RSV expressing the F protein of the clinical isolate 2–20 and suggest RSV F may be interacting with EGFR in a strain-dependent manner.
We assessed whether a physical (direct or indirect) interaction can occur between overexpressed RSV F and EGFR. 293T cells were transfected with either EGFR or SLAM and RSV F. All proteins were detected at high levels in whole cell lysates (Fig 4A). RSV F was immunoprecipitated (IP) well with motavizumab mAb and probed for the presence of co-precipitated EGFR. Motavizumab binds an epitope in RSV F that is conserved among RSV strains and between the prefusion and postfusion conformations of the protein [41, 42]. EGFR was detected after precipitation of A2F and 2-20F, and there was a 2.8-fold higher ratio of EGFR bound to 2-20F than to A2F (Fig 4B). We previously deposited the 2–20 genome sequence to GenBank and have modeled the residue differences between A2 F and 2–20 F [27]. To identify putative EGFR interaction sites for RSV 2–20 F, three sets of mutations were introduced into 2–20 F that changed amino acids to the corresponding A2 residues. Mutation sets were introduced into the F head region (T63N, E66K, and G76I triple mutant), stalk region (G519V and K524N double mutant), and in the cleaved 27 residue peptide (pep27) between the furin cleavage sites (N124K) of 2–20 F. Introduction of either the head or the stalk mutation sets resulted in reductions in the efficiency of EGFR-bound to 2–20 F, as detected by IP (Fig 4B). The N124K pep27 mutation did not affect the co-IP efficiency significantly (Fig 4B). Taken together, residues 63/66/76 and residues 519/524 contributed to the co-IP interaction between 2–20 F and EGFR.
We evaluated the Co-IP interaction of F proteins from other RSV strains with EGFR. Strain Line 19 is a laboratory strain that that is more pathogenic in mice, including high viral loads, lung IL-13 levels, and airway mucin expression in BALB/c mice, compared to prototypic strains A2 and Long [43, 44]. RSV strain A2001/3-12 (3–12) is a clinical strain we previously reported to induce lower airway mucin expression in mice than 2–20 [27]. Similar to 2–20, overexpression of line19 F resulted in higher efficiency of EGFR co-IP than did overexpression of A2 F, Long F, or 3–12 F (Fig 4B) The data correlate strength of co-IP F-EGFR interaction in vitro with in vivo mucogenicity in mice of the strain from which F was derived.
We quantified molecular co-localization between EGFR and RSV F at the cell surface. H292 cells were inoculated with either mock, A2, A2-2-20F, or A2-2-20GF at 4°C, which facilitates attachment but not viral fusion. After one hour and washes, the cells were fixed and stained for RSV F and EGFR then examined by superresolution microscopy at the plasma membrane. There was no significant overlap between EGFR and RSV F puncta in the A2 strain group (Fig 5A and 5B). In contrast, there was significant overlap between RSV F and EGFR puncta on the surface of A2-2-20F- and A2-2-20GF-inoculated cells (Fig 5A and 5B). There were more RSV F puncta per cell in the A2-2-20GF group than the A2-2-20F and A2 groups, consistent with published data that 2–20 G has greater attachment activity than A2 G (Fig 5A) [18]. A2-2-20GF attachment also resulted in greater co-localization of F and EGFR than A2-2-20F (Fig 5A and 5B), suggesting the 2–20 G protein also directs RSV localization at the cell surface. Taken together, there was significant co-localization between 2–20 F and EGFR at the cell surface following virus attachment, which was enhanced by 2–20 G.
To evaluate whether RSV co-localizes with EGFR in a highly relevant primary airway epithelium model, well-differentiated primary pediatric bronchial epithelial cells (WD-PBECs) [45] were analyzed for surface expression of p-EGFR. RSV BT2a induces mucus secretion and goblet cell hyperplasia/metaplasia in WD-PBECs [45]. p-EGFR expression in WD-PBEC cultures was largely focused in the cell membrane on ciliated structures as evidenced by co-localization with beta-tubulin (Fig 5C). WD-PBEC cultures were infected with clinical isolate, BT2a, which has similar kinetics to A2 during infection in HEp-2 cells, but exhibits more cytopathogenesis in WD-PBEC cultures [45]. Many cells infected with clinical isolate BT2a qualitatively co-localized with p-EGFR (Fig 5D). Interestingly, infection of WD-PBECs with BT2a resulted in an apparent increase in surface expression of p-EGFR in RSV-infected cells, but not neighboring non-infected cells (Fig 5E). Collectively, the co-localization data suggest RSV F is closely associated with EGFR at the plasma membrane of H292 and apically in human primary airway epithelial cells.
Infection of BALB/cJ mice with RSV A2-2-20F or the parental isolate 2–20 results in airway mucin expression that peaks approximately day 8 post-infection, a time point when infectious virus is not detectable by plaque assay [27, 28]. We previously demonstrated that 2–20 infects the mouse airway epithelium, detectable by immunofluorescence day one post-infection [27]. We examined whether A2-2-20F also infects the mouse airway epithelium and whether virally expressed proteins can be detected in PAS-positive airways. Mice were mock-infected or infected with A2-2-20F that does expresses mKate2 or A2-2-20F lacking the mKate2-encoding gene. The mKate2 far-red fluorophore has extreme pH stability [46], therefore we hypothesized it will remain functional after histology processing. When adjacent, serial lung sections were compared, mKate2 signal was evident day 8 post-infection in bronchial epithelium that was also producing mucin (Fig 6). To control for autofluorescence, lung sections from recombinant A2-2-2F that does not express mKate2 infected and mock-infected mice were also examined. The mKate2 signal was clearly distinguishable from background (Fig 6), marking cells that were either previously infected and harboring mKate2 or cells actively expressing RSV-encoded gene products. These results correlate A2-2-20F infection with airway mucin expression.
Muc5ac is a major inducible and secreted mucin protein in the lung that is up-regulated by EGFR activation and during RSV infection [43, 47–49]. We first tested whether RSV infection of these cells results in MUC5AC gene expression. In DMSO-treated serum-starved H292 cells, A2-2-20GF infection resulted in higher MUC5AC mRNA levels than mock, A2, and A2-2-20F infection (Fig 7A). Therefore H292 cells provide an in vitro model of RSV strain-specific induction of mucin expression. Serum-starved H292 cells were treated with EGFR pathway inhibitors PD153035 and AG1478, mock-infected or infected with RSV strains A2, A2-2-20F, and A2-2-20GF, and MUC5AC expression was quantified. In the presence of either inhibitor, MUC5AC mRNA fold-change in A2-2-20GF infected cells was ablated. The combination of RSV 2–20 F and 2–20 G was important for MUC5AC induction in this in vitro model, consistent with the efficient attachment function of 2–20 G (Fig 5A and ref [18]), and the EGFR pathway was critical for MUC5AC induction.
We tested the role of EGFR signaling in A2-2-20F-induced airway mucin expression. BALB/cJ mice were pre-treated with erlotinib, a specific quinazoline derivative that binds to and inhibits EGFR tyrosine kinase activity, or vehicle suspension, and tested for suppression of EGFR activation in mouse lung homogenates [50]. Erlotinib caused a reduction in the signal of p-EGFR in lungs from both mock-infected and A2-2-20F-infected mice (Fig 7B). Mice pre-treated then treated daily with vehicle or erlotinib were infected with mock or A2-2-20F. Lungs were harvested day 8 post-infection and processed for periodic acid-Schiff (PAS) stains of goblet cell hyperplasia/metaplasia (Supplemental Fig 2), a surrogate of airway for mucin expression. Our lab uses a digital pathology system to morophometrically quantify PAS positivity in all airways involved in lung sections [27, 28]. Mice treated with erlotinib and infected with A2-2-20F had significantly less airway goblet cell hyperplasia/metaplasia than vehicle-treated, A2-2-20F-infected mice (Fig 7C and 7D). Taken together, these data identify a novel role for EGFR signaling in mediating RSV-induced mucin expression and airway pathology.
RSV disease in infants is associated with airway obstruction, lung inflammation, epithelial cell sloughing, and mucus production [51, 52]. The relative contributions of sloughed epithelium and mucus production to airway obstruction remain unknown [53, 54]. Distal airways are thought to have fewer mucin-secreting cells than larger airways, but overabundant mucus production is associated with infant bronchiolitis clinically [53–55]. The clinical isolate strain 2–20 to a degree recapitulates human RSV disease in the BALB/c mouse model of RSV infection [27, 28, 56]. The prototypical A2 strain does not cause airway mucin expression in mice, and chimeric RSV strains A2-2-20F is mucogenic like parental 2–20, implicating the F protein in mucin induction in vivo [28, 44]. RSV 2–20 has somewhat altered tropism in the mouse because it infects the airway epithelium and alveolar epithelial cells, whereas A2 infects predominantly alveolar epithelial cells [27]. 2–20 and A2-2-20F cause more airway necrosis in mice than A2 [28]. Therefore, infection of the mouse airways correlates with mucin induction in this model. Here, we investigated mechanisms of 2–20 F-induced airway mucin induction. In cultured cells, 2–20 F exhibited greater functional and physical interaction with EGFR than A2 F, and EGFR was able to enhance the fusion activity of the 2–20 F protein. Chemical inhibition of EGFR signaling reduced infectivity of 2–20 F-expressing RSV and ablated mucin induction in vitro and in vivo. We report that RSV F interacts with and activates EGFR and that EGFR contributes to infection in vitro and plays a critical role in RSV-induced mucin expression.
Our study shows F and EGFR interact functionally and physically. Expression of the 2–20 F protein potently activated EGFR, as measured by p-EGFR levels in cells. The efficiency of co-IP of EGFR with F depended on strain specificity of the expressed F protein. EGFR had the highest co-IP efficiency with 2–20 and line 19 F proteins, strains previously shown to be mucogenic in mice [27, 28, 44]. We mapped the enhanced co-IP efficiency with 2–20 F to two domain differing between 2–20 F and A2 F, residues 63/66/76 and residues 519/524. Amino acids 63, 66, and 76 cluster at the top of the prefusion F trimer. This region overlaps prefusion-specific antigenic site ∅, and the adjacent residue 67 is important for RSV F prefusion stability [11, 57]. As we discussed in Stokes et al, residues 519 and 524 in RSV F are membrane-proximal in the stalk, a region implicated in regulating Hendra virus F protein triggering via stabilization of the pre-fusion form [28, 58]. In our co-IP experiments, we expressed functional F, which for RSV does not require triggering by the attachment protein, so prefusion and postfusion forms were present. We speculate that 63/66/76 and 519/524 regions regulate prefusion stability, which may relate to EGFR co-IP efficiency between F species and mutants. In the DSP fusion assay, EGFR boosted 2–20 activity, so we predict the functional interaction occurs prior to postfusion F formation. Additional studies and reagents will be required to further elucidate F-EGFR molecular interactions.
RSV infection was previously shown to activate EGFR. The A2 strain activates EGFR in cells, resulting in delayed apoptosis by ERK activation and production of the pro-inflammatory cytokine IL-8 production [24]. RSV A2 strain was shown to activate EGFR following virus attachment, leading to macropinocytotic endocytosis [26]. In that study, EGFR depletion in HeLa cells by siRNA delivery resulted in reduced infectivity of the A2 strain, whereas we found that EGFR depletion in BEAS-2B cells reduced infectivity of RSV expressing 2–20 F but not of the A2 strain. The discordant findings may be related to the cell line and/or the efficiency of EGFR knockdown. Recently, RSV activation of EGFR, in addition to influenza A (H1N1) and rhinovirus infections, led to suppression of IRF1-dependent CXCL10 production [25]. CXCL10 is expressed in airway epithelial cells, is a ligand of CXCR3 (a key regulator of leukocyte trafficking), and when elevated is associated with obstructive airway diseases [25, 45, 59, 60]. Using RSV 2–20 F-expressing viruses in future studies may shed additional light on entry mechanisms, such as macropinocytosis, and downstream immunopathogenesis such as IL-8 expression, neutrophil recruitment, and CXCL10 expression.
The capacity of RSV F to engage EGFR during infection may depend in part on a function of the RSV G protein. We observed an appreciable increase in MUC5AC expression in H292 cells when using strain A2-2-20GF, not A2 or A2-2-20F. In these cells, 2–20 G conferred greater attachment, as measured by superresolution microscopy, consistent with the sequence-based prediction that 2–20 G has more glycosylation sites and our recent published data that 2–20 G has a higher apparent molecular weight than A2 G and confers enhanced cell attachment in vitro [18]. The co-localization quantification in H292 cells revealed that, irrespective of the abundance of RSV F puncta, 2–20 G greatly enhanced signal overlap between F and EGFR, suggesting 2–20 G alters location of RSV on the plasma membrane. Our current working model is that initial 2–20 G interactions with CX3CR1, GAGs, and/or other factors mediates attachment that likely precedes F-EGFR interaction, EGFR activation, and infection (Fig 8). Further studies will need to evaluate the role of G in F-EGFR interactions during infection in cells and animal models.
In summary, we for the first time identified a host protein that both interacts with the RSV F protein and promotes fusion. EGFR was expressed at the apical surface of differentiated pediatric bronchial epithelial cells, and RSV F and EGFR co-localized in infected cells. Using clinically relevant RSV strains and infection models, we found that EGFR is critical for RSV-induced airway mucin expression and laid the groundwork for defining the molecular interaction between F and EGFR.
All animal procedures were conducted according to the guidelines of the Emory University Institutional Animal Care and Use Committee, under approved protocol number 2001533. The study was carried out in accordance with recommendations in the Guide for Care and Use of Laboratory Animals of the National Institute of Health, as well as local, state, and federal laws.
The media components and origins of HEp-2, BEAS-2B, 293T, and BSR-T7/5 cells we use are previously described [27, 28]. NCI-H292 cells were purchased from ATCC (CRL-1848) and propagated in RMPI-1640 supplemented with 10% FBS (Hyclone, Thermo Scientific), 0.01 M HEPES, and 25 mM D-glucose. Normal human bronchial epithelial (NHBE) cells were obtained from Lonza (Allendale, NJ) and differentiated on collagen-coated 24-well transwell supports at air-liquid interface as we described [38]. The human codon bias-optimized RSV A2 F and 2–20 F expression plasmids are described, and the RSV 2-20GF-expression plasmid was generated using the same strategy [28]. EGFR cDNA in a pBABE retroviral vector was transferred into pTriEx-3 using standard restriction enzyme cloning to yield pTriEx-3-EGFR, and the sequence of EGFR was confirmed. A full-length SLAM cDNA was extracted from Vero-SLAM cells and cloned into the pCG expression plasmid. DSP1-7 and DSP8-11 plasmids were provided by Naoyuki Kondo and Zene Matsuda [61]. Lipofectamine 2000 (Life Technologies) was used according to the provided protocol for all cell transfections, with the exception that DNA/liposome mixture remained on cells overnight for RSV rescue. The anti-RSV F mAb motavizumab was generously provided by Nancy Ulbrandt (MedImmune AZ). Mouse anti-SLAM was purchased from Abcam (clone ab2604) and mouse anti-vinculin was purchased from Fisher Scientific (clone VLN01). HRP-conjugated secondary antibodies for immunoblotting were purchased from Jackson Immunoresearch. Compound AG1478 (LC Laboratories) was diluted in DMSO. Recombinant human EGR was obtained from Lonza (Cat #00556827).
We previously reported generation of recombinant RSV encoding the far-red fluorescent protein monomeric Katushka-2 (mKate2) in the first gene position [62]. This virus was recovered on BSR-T7/5 cells transfected with pSynkRSV-line19F BAC together with helper plasmids encoding codon-optimized N, P, L, and M2-1 [62]. The F gene of BAC pSynkRSV-line19F, flanked by SacII and SalI sites, was replaced with a synthetic cDNA (GeneArt, Life Technologies) encoding either the A2 strain F open reading frame, the 2–20 strain F open reading frame, or the 2–20 strain G and F open reading frames, flanked by non-coding regions identical to those in pSynkRSV-line19F BAC [18]. The recombinant A2-2-20F that does not encode mKate2 (A2-2-20F –mKate2) was previously described [28]. The kRSV-A2, kRSVA2-2-20F, and kRSVA2-2-20GF viruses were plaque-purified and amplified in HEp-2 cells. Virus stocks used were sequence confirmed for the G and F genes and determined to be Mycoplasma free using the Venor GeM Mycoplasma detection kit (Sigma-Aldrich). RSV clinical isolate BT2a was used for infection of well-differentiated primary pediatric bronchial epithelial cells (described below) [45]. Measles virus expressing GFP (MeV-GFP) used in this study was previously described [63].
shRNA constructs targeting EGFR (TRCN0000039635, TRCN0000039634, TRCN0000039633, TRCN0000010329, TRCN0000121067) were purchased from Sigma. The control plasmid pLKO.1, which has a scrambled shRNA, was also purchased from Sigma. Two lentivirus helper plasmids, psPAX2, and pMD2.VSVG, were kindly provided by Gregory Melikian (Emory University). Lentiviruses for the puromycin pLKO constructs were produced in 293T cells with two helper constructs. 293T cells (6.5 × 105) were transfected with pLKO containing the EGFR shRNA, psPAX2, and pMD2.VSVG. As a control, 293T cells were transfected with pLKO.1 containing a scramble shRNA. Supernatants were harvested at 24 h following transfection. For infection, 30 to 40% confluent BEAS-2B cells were spinoculated with virus-containing supernatant at 4°C, 2900 x g. Following overnight incubation, media containing 1 μg/mL puromycin was added for 48 hours post-infection. A kill curve of puromycin on BEAS-2B cells determined that 1μg/mL puromycin killed 100% of cells. Surviving BEAS-2B cells following puromycin treatment were used for RSV infection and subsequently used for determination of EGFR knockdown efficiency.
Flow cytometry analysis was performed to quantify RSV infectivity and levels of EGFR surface expression. Phycoerythrin-EGFR antibody (sc-101 PE, Santa Cruz) was used for the detection of EGFR on non-permeabilized 293T or BEAS-2B cells using an LSRII flow cytometer (BD Biosciences). Motavizumab and Alexa 488-anti-human secondary Ab (H17101, Life Technologies) were used for measuring RSV F cell surface expression on non-permeabilized 293T cells. BEAS-2B cells were harvested, fixed, and acquired using a 561nm laser. Isotype control antibody (PE-IgG) was used as a negative control. For measuring of RSV and MeV infectivity, cells were harvested 24 hr post-infection and acquired using an LSRII, by detecting the mKate2 and GFP signals, respectively. Data were analyzed using FlowJo software (TreeStar, Ashland, OR).
293T cells transfected with equal molar amounts of DNA were used for each group, and plasmid quality was checked via agarose gel and spectrophotometer quantitation. 24 h after treatment, cells were harvested. Cell pellets were thawed on ice and lysed with RIPA buffer (Sigma) plus protease inhibitors (Thermo Scientific). Protein concentrations were determined using Bradford Reagent (Sigma). 50% (v/v) protein-A magnetic beads (Cell Signal) were conjugated with motavizumab by mixing 50 μL of bead slurry with 500 μL ice cold PBS and 0.75 μg of motavizumab (1.75 mg/mL) in a microcentrifuge tube rotating end-over-end for 4 h at 4°C. Excess antibody was washed from the beads by pelleting, aspirating the antibody suspension, and washing 3 times with 1 mL of RIPA buffer. 120 μL of cell lysate was pre-cleared with 30 μL of non-conjugated bead slurry. 100 μL of cleared lysate was then added to the conjugated bead pellet with 10 μL of 10% BSA (Sigma). The lysate/bead slurry was allowed to mix overnight, rotating end over end at 4°C. Beads were removed using a magnetic tube rack on ice at 4°C, were washed 4x in ice cold RIPA, and washed 1x in ice cold PBS. Beads were then pelleted by and re-suspended in 3x SDS loading buffer. For Westerns, lysates or beads mixed with 3x SDS loading buffer were heated 95°C 5 min then fractionated on 10% SDS-Page gels (Bio-Rad). Proteins were transferred to PVDF membranes (Bio-Rad). Membranes were blocked with 2% non-fat milk, 1% FBS (Gemini) in TBST. Membranes were incubated in primary antibody overnight (p-EGFR, EGFR, or SLAM) or 2 h (motavizumab or vinculin).
The dual-split protein (DSP) reporter cell-cell fusion assay was previously adapted to measuring RSV F protein activity [28, 33, 34]. 293T “effector” (cis) cells were transfected with RSV A2 F or 2–20 F and DSP1-7 in the presence of fusion inhibitor BMS-433771 (a gift from Jin Hong, Alios Biopharma, San Francisco, CA). 293T “target cells” (trans) were transfected with DSP8-11 and pCG-SLAM or pcDNA3.1 empty vector. Effector cis or target trans cells were transfected with pTriex3-EGFR. Effector and target cells were washed with PBS 24 h post-transfection and harvested by pipetting in media containing EnduRen live cell luciferase substrate (Promega). Equal volumes of effector and target cells were mixed and placed into an opaque 96-well plate in quadruplicate. Plates were incubated at 37°C and luciferase activity as a measure of cell-cell fusion was assayed on a TopCount Luminescence counter (Perkin Elmer) 4, 6, and 8 h after cell mixing. A positive control of DSP1-7 and DSP8–11 transfected into the same cell population was used to validate replicates.
H292 cell monolayers were serum-starved for 24 h then mock-infected or infected with RSV. The cells were washed with PBS and lysed with TRIzol reagent (Life Technologies) at 20 h post-infection. Total RNA was isolated according to the TRIzol protocol. Quantitative real-time PCR was performed using the AgPath-ID OneStep RT-PCR kit (Applied Biosystems) and an ABI 7500 sequence detector system (Applied Biosystems). The primers and probes for MUC5AC gene (forward, 5′CGTGTTGTCACCGAGAACGT3′; reverse, 5′ ATCTTGATGGCCTTGGAGCA 3′, probe, 5′ Fam- CTGCGGCACCACAGGGACCA-BHQ-1 3′) were obtained from Integrated DNA technologies (IDT) [64]. The primers and probes for GAPDH, the control, were forward, 5’ GAAGGTGAAGGTCGGAGT 3’, reverse, 5’ GAAGATGGTGATGGGATTTC 3’, and probe, 5’ Fam CAAGCTTCCCGTTCTCAGCC 3’. Threshold cycles (Ct) and ΔCt for each sample was calculated. Assays were performed in duplicate in 3 independent experiments.
H292 cells were cultured on 35-mm glass-bottom dishes (MatTek Corp). Cells were inoculated at MOI = 3 at 4°C rocking for 1 h., conditions at which attachment but not fusion can occur, washed twice in chilled PBS, and fixed in 10% buffered formalin (Thermo-Fisher) for 10 min. Cells were then washed 3X at room temperature with PBS. The fixed cells were blocked overnight in serum-free protein block (Dako). Antibodies were diluted in an antibody diluent with background reducing components (Dako). For the detection of RSV-F, a 1:2,000 dilution of 1.75 mg/mL motiavizmab was utilized. EGFR was detected with a rabbit polyclonal (Millipore) used at a 1:500 dilution. Primary antibodies were allowed to bind to cells for 4 h rocking at 4°C. Secondary antibodies anti-human IgG, IgA FITC conjugated (LifeTech) and anti-rabbit Alexa-568 (LifeTech) were used at dilutions 1:5,000 and 1:2,000 respectively. Secondary antibodies were incubated for 1 h rocking at room temp, followed additional 3 wash steps. Cells were kept at 4°C under PBS until imaging. Super-resolution images were acquired using a DeltaVision OMX Blaze (GE Healthcare Life Sciences) for three-dimensional structured illumination microscopy (3D-SIM). 3D-SIM reconstructions were generated by softWoRx (v6.1.3). The reconstructed files were further analyzed in Imaris (v8.1.2) where appropriate channel thresholds were manually set. An overlap mask channel was created using Imaris where the thresholded Mander's coefficient was calculated to quantify 3D overlap.
For WD-PBECs, pediatric bronchial epithelial cells (PBEC) were obtained, via written parental consent, from bronchial brushings of children undergoing elective surgery at the Royal Belfast Hospital for Sick Children, and the procedures were approved by the Office of Research Ethics Committees Northern Ireland [45]. PBEC were expanded in collagen-coated flasks using airway epithelial cell media and supplements (Lonza), then seeded onto transwell inserts (Corning), and then air-liquid interface (ALI) cultures were initiated and maintained 21 days in order to establish well-differentiated (WD)-PBECs, as described in further detail [45]. Paraformaldehyde-fixed and permeabilized WD-PBEC were stained for anti-β-tubulin, MUC5AC, or RSV F protein expression as described [45] and were stained with anti-phospho-(p)-EGFR (Abcam, ab40815). WD-PBEC cultures were infected with RSV subgroup A clinical isolate BT2a as described [45]. Fluorescent images were obtained with a SP5 confocal DMI 6000 inverted microscope (Leica).
7-week old female BALB/cJ mice (The Jackson Laboratory) were orally gavaged with 100 mg/kg of erlotinib (Selleck Chemicals LLS, Catalog S1023) or vehicle (0.5% carboxymethylcellulose /0.1% Tween 80) in a total volume of 100 μL daily, beginning two days prior to infection, and continuing for the duration of the experiment. Mice were infected intranasally with 1 x 106 PFU of A2-2-20F or mock virus preparation. On day 8 post-infection, the lungs were harvested and placed in 10% neutral buffered formalin for histopathology sectioning and periodic acid Schiff (PAS) staining for goblet cell hyperplasia as a measure of airway mucin expression. PAS positivity was quantified for greater than 285 individual airways total from 5 separate mice per group by digital morphometric analysis as described previously [27]. In a separate group of mice, the left lung was harvested, snap frozen, and later homogenized in RIPA buffer containing a protease inhibitor cocktail. The amount of total protein the lung homogenates was determined by Bradford assay and equivalent proteins from each group were loaded on a 4–17% SDS-PAGE gel and separated by electrophoresis before Western blotting for total EGFR (Abcam, AB15669) or p-EGFR Monoclonal (Abcam, AB24928).
7-week old female BALB/cJ mice (The Jackson Laboratory) were infected intranasally with 1 x 106 PFU of A2-2-20F (-mKate2), 1 x 105 FFU of A2-2-2-20F, or mock virus preparation. On day 8 post infection, lungs were harvested and prepared as above for PAS staining or, after sectioning, were deparaffinized with Clear-rite (Thermo Scientific), rehydrated through graded alcohols to water, and then stained with Prolong DAPI Gold (Life Technologies). PAS slides were imaged using a Mirax Imaging System as descried previously [27]. DAPI-stained slides analyzed by fluorescence microscopy using the Mirax Image System. Equal exposure time was used for each channel (DAPI 120μs, mKate2 900μs) across all groups. Excitation was provided by HXP-120 light source (LEj) through Zeiss filter set 2 (DAPI) or 45 (mKate2). Images were analyzed in Panoramic Viewer v1.15.2 (3DHISTECH), where pseudocoloring levels were kept constant for each channel across all groups. Images of representative airways from the central portion of each lung were exported as TIF files.
|
10.1371/journal.pgen.1004224 | Modeling 3D Facial Shape from DNA | Human facial diversity is substantial, complex, and largely scientifically unexplained. We used spatially dense quasi-landmarks to measure face shape in population samples with mixed West African and European ancestry from three locations (United States, Brazil, and Cape Verde). Using bootstrapped response-based imputation modeling (BRIM), we uncover the relationships between facial variation and the effects of sex, genomic ancestry, and a subset of craniofacial candidate genes. The facial effects of these variables are summarized as response-based imputed predictor (RIP) variables, which are validated using self-reported sex, genomic ancestry, and observer-based facial ratings (femininity and proportional ancestry) and judgments (sex and population group). By jointly modeling sex, genomic ancestry, and genotype, the independent effects of particular alleles on facial features can be uncovered. Results on a set of 20 genes showing significant effects on facial features provide support for this approach as a novel means to identify genes affecting normal-range facial features and for approximating the appearance of a face from genetic markers.
| The face is perhaps the most inherently fascinating and aesthetic feature of the human body. It is a principle subject of art throughout human history and across cultures and populations. It provides the most significant means by which we communicate our emotions and intentions in addition to health, sex, and age. And yet features such as the strength of the brow ridge, the spacing between the eyes, the width of the nose, and the shape of the philtrum are largely scientifically unexplained. Here, we use a novel method to measure face shape in population samples with mixed West African and European ancestry from three locations (United States, Brazil, and Cape Verde). We show that facial variation with regard to sex, ancestry, and genes can be systematically studied with our methods, allowing us to lay the foundation for predictive modeling of faces. Such predictive modeling could be forensically useful; for example, DNA left at crime scenes could be tested and faces predicted in order to help to narrow the pool of potential suspects. Further, our methods could be used to predict the facial features of descendants, deceased ancestors, and even extinct human species. In addition, these methods could prove to be useful diagnostic tools.
| The craniofacial complex is initially modulated by precisely-timed embryonic gene expression and molecular interactions mediated through complex pathways [1]. As humans grow, hormones and biomechanical factors also affect many parts of the face [2], [3]. The inability to systematically summarize facial variation has impeded the discovery of the determinants and correlates of face shape. In contrast to genomic technologies, systematic and comprehensive phenotyping has lagged. This is especially so in the context of multipartite traits such as the human face. In typical genome-wide association studies (GWAS) today phenotypes are summarized as univariate variables, which is inherently limiting for multivariate traits, which, by definition cannot be expressed with single variables. Current state-of-the-art genetic association studies for facial traits are limited in their description of facial morphology [4]–[7]. These analyses start from a sparse set of anatomical landmarks (these being defined as “a point of correspondence on an object that matches between and within populations”), which overlooks salient features of facial shape. Subsequently, either a set of conventional morphometric measurements such as distances and angles are extracted, which drastically oversimplify facial shape, or a set of principal components (PCs) are extracted using principal components analysis (PCA) on the shape-space obtained with superimposition techniques, where each PC is assumed to represent a distinct morphological trait. Here we describe a novel method that facilitates the compounding of all PCs into a single scalar variable customized to relevant independent variables including, sex, genomic ancestry, and genes. Our approach combines placing spatially dense quasi-landmarks on 3D images [8], [9], principal component analysis (PCA), and a new partial least squares regression (PLSR, [10]) derived method we call “bootstrapped response-based imputation modeling” (BRIM) to measure and model facial shape variation (Text S1, Figures S1, S2, S3).
Given the multivariate nature of the face and the large number of genes likely affecting variation in the face, we chose to focus attention on the between-population variation with a genetic admixture approach using research participants from three West African/European admixed populations. Ancestry informative markers (AIMs) can be used to estimate individual genomic ancestry from DNA [11], which can be used to investigate population differences and map genes for genetically determined traits that vary between populations. Non-random mating and continuous gene flow in admixed populations results in admixture stratification or variation in individual ancestry [12], [13]. The process of admixture also results in admixture linkage disequilibrium or the non-random association among both AIMs and traits that vary between the parental populations. These characteristics make admixed populations uniquely suited to investigations into the genetics of such traits [14]–[16]. By simultaneously modeling facial shape variation as a function of sex and genomic ancestry along with genetic markers in craniofacial candidate genes, the effects of sex and ancestry can be removed from the model thereby providing the ability to extract the effects of individual genes.
A spatially dense mesh of 7,150 quasi-landmarks was used to map 3D images of participants' faces onto a common coordinate system (Figure 1). Quasi-landmarks are defined here as largely homologous vertices in this mapped mesh. The mesh is applied automatically, eliminating the difficult and error-prone procedure of manually indicating facial landmarks [8], [9], [17]. Deviations from bilateral symmetry were removed by averaging each face with its mirror image [18], [19]. PCA on the symmetrized 21,450 quasi-landmark 3D coordinates (X, Y, and Z for each of the 7,150 quasi-landmarks) using all 592 participants produces 44 principal components (PCs) that together summarize 98% of the variation in face shape and define a multidimensional face space. The effects of the first 10 PCs are illustrated in Figure 2. Some of these PCs (e.g., PC4, PC5) capture the effects of changes in only particular parts of the face. However, many PCs (e.g., PC1, PC2, PC3) capture effects in multiple parts of the face. Moreover, although the PCs are statistically independent, any particular part of the face is affected by several PCs. As such, it is likely incorrect to assume that each PC represents a distinct morphological trait resulting from the action of specific genes. Our use of BRIM to combine the independent effects of PCs is agnostic about their biological meaning, if any, and provides for the compounding of the information from any or all of the PCs together into a single variable that is customized to the predictor variable being modeled. In this way, BRIM also overcomes the problem of multiple testing inherent to other methods for summarizing facial variation. In other words, the hypothesis, does this gene have significant effects on facial shape, can be addressed with a single statistical test (Text S1).
BRIM is an extension of existing relationship modeling techniques that uses response variables to refine and, in some cases, to transform one or more initial predictor variables. In other words and in contrast to alternate techniques, BRIM uses a multivariate matrix of response variables in a leave-one-out forced imputation setup to update the initial predictor variable values, creating a new type of variable – the response-based imputed predictor (RIP) variable (Figure S2). The BRIM process is bootstrapped, and estimator improvement over successive iterations can be monitored (Figures S5, S6, S7, S8, S9). BRIM also functions to correct observation error, misspecification of predictor values, and other sources of statistical confounding (Text S1). Within the iterative bootstrapping scheme, a nested leave-one-out approach is used to avoid model over-fitting and to allow hypothesis testing using standard statistical techniques, such as correlation analysis, ANOVA, and receiver operating characteristic (ROC) curve analysis [20], to test the significance of the association between the predictors and RIP variables. Likewise, the relationships between the RIP variables and the response variables, e.g., the 21,450 facial parameters, allows for the visualization and quantitation of their effects on face shape.
RIP variables modeling sex (RIP-S) and genomic ancestry (RIP-A), as well as those modeling the effects of particular genetic markers (RIP-Gs), can be visualized using two primary methods – shape transformations and heat maps. We used three summary statistics (area ratio, normal displacement, and curvature difference), which can be illustrated using heat maps, to quantify the particular changes to the face that result. These measures of facial change, along with particular inter-landmark distances, angles, and spatial relationships, can together be termed face shape change parameters (FSCPs). FSCPs provide a means of translating face shape changes from the abstract face space into both visual representations into words. Such terms are used in clinical and anthropological descriptions of faces and by doing so we can compare these to the BRIM results (e.g., Figures S28, S29, S30, S31, S32, S33, S36, S37, S38, and Table S1). The statistical significance of these and related FSCPs can be tested using permutation.
As expected, many parts of the face are affected by both ancestry and sex. Figure 3 illustrates the partial effects of RIP-A and RIP-S on facial shape using transformations and heat maps for effect size (R2) and the three primary FSCPs. Facial regions that are statistically significant (p<0.001) for effect size and the FSCPs are shown in Figure 3 as the yellow (not green regions in the bottom panels). The RIP-A and RIP-S shape transformations shown are set to the points three standard deviations plus and minus the mean RIP-A and RIP-S levels in these samples. As seen in the effect-size (R2) panels in Figure 3, the proportion of the total variance in particular facial features explained by RIP-A and RIP-S can be substantial. In general, up to a third of the variance in several parts of the face is explained by these two variables. RIP-A primarily affects the nose and lips and, to lesser extents, the roundness of the face, the mandible, and supraorbital ridges. Sex has a much larger effect than ancestry on the supraorbital ridges and cheeks, and smaller effects on the nose and under the eyes. The FSCPs help to illustrate the specific ways in which particular RIP variables affect the face. For example, the area ratio shows increased surface area for the medial canthus, sides of the nose, and front of the chin on the European end of RIP-A and a greater surface area for the nostrils and lips on the West African end of RIP-A. The curvature difference highlights the top of the philtrum as a facial feature that is highly convex on the European end and highly concave on the West African end of RIP-A. Regions showing curvature differences for RIP-A are also seen in the nasal bridge, supraorbital ridges, and chin. RIP-S shows greatest effects on the supraorbital ridges, nasal bridge, nasal ridge, zygomatics, and cheeks. The nose, lips, medial canthus, and mandible are also affected by RIP-S. The largest differences in facial curvature related to changes in RIP-S are on the supraorbital ridges and the nasal bridge.
Despite the complex ways in which faces are affected by RIP-A and RIP-S, these variables are useful summaries of the degree to which particular faces are more or less ancestry-typical and sex-typical, respectively. This is evident in the strong relationship observed between RIP-A and genomic ancestry as measured with a panel of 68 AIMs (r = 0.81, p<0.001; Figure 4A). Approximately two thirds of the variation in RIP-A across these three West African/European admixed populations is explained by genomic ancestry. Likewise, as seen in Figure 4B, RIP-S is very distinctive between the sexes. ROC analyses (Figure S32) show that the AUC for RIP-S on sex is 0.994 (p<0.001). Genomic ancestry, independently from sex, explains 9.6% of the total facial variation, while sex independently from ancestry explains 12.9% of the total facial variation (Table S3). Most facial variation, like human genetic variation in general, is shared among different human populations and by members of both sexes.
We used alternate subsets of AIMs and alternate population samples to test the robustness of the facial ancestry (RIP-A) estimation. RIP-A values were derived using different initial predictor variables and compared. The pairwise correlations of RIP-A estimates are high (R2>0.99), showing that very similar estimates of facial ancestry result from different panels of AIMs (Figure S9) and alternate population samples (Figures S10, S11). The robustness of RIP-A estimates to both marker panel and population sample substantiates the generality and, thus, practical usefulness of these models.
We also see that RIP-A estimates generated using AIMs panels with lower ancestry-information content show stronger correlations with more accurate genomic ancestry estimates than with the genomic ancestry estimates that were used to generate them (Figure S9). To further evaluate the performance of BRIM when less information is available, we performed noise injection experiments by adding or subtracting randomly defined quantities from the estimates of genomic ancestry and misclassifying the sex of persons in the sample (Figures S4, S5, S6, S7, S8 and Figures S12, S13, S14, respectively). These experiments demonstrate the same patterns noted above using alternate panels of AIMs: Accurate RIP variables for these two traits are possible with incorrect coding of sex and imprecise estimates of genomic ancestry. The initial predictor variable values of both sex and ancestry can be reduced in precision by as much as 30% (i.e., r2 = 0.7 between the original predictor variable and the noise predictor injected variable) and still show correlation coefficients of about r = 0.95 between the RIP measures generated with these noisy estimates and RIP measures generated with the original estimates (Figure S8 and Figure S14). BRIM is efficient in using the latent covariance structure of the facial PCs to discover the paths through face space that reflect sex and ancestry and can accurately summarize the relative positions of individual faces on these paths as RIP-S and RIP-A, respectively.
Humans are also very adept at observing faces and can infer many aspects of the variability among faces [21], [22]. Given this, we attempted to test whether the human observer might provide a means of validating the RIP-A and RIP-S variables. Observers were shown false-colored 3D animated GIF images of research participants' faces and asked to rate the proportion of West African ancestry (from 0% to 100%) and the femininity (using a Likert scale from 1 to 7). Observers were also asked to judge the sex and the population group. As shown in Figures 5A and 5B, the correlations between RIP-A and observer ratings of proportional facial ancestry and judgments of facial population are strong (all r>0.85 and p<0.0001). Similarly, RIP-S and observer ratings of facial femininity and judgments of facial sex are also highly correlated (r>0.85 and p<0.0001; Figures 5C and 5D). These findings provide additional validation that RIP-A and RIP-S are informative summary statistics representing the relative levels of facial ancestry and facial femininity.
Like sex and genomic ancestry, SNP genotypes can be used as initial predictor variables in BRIM resulting in one RIP-G variable per SNP. We performed a partial BRIM analysis modeling genotype effects independent of sex and ancestry for each of 76 West African/European ancestry-informative SNPs located in 46 craniofacial candidate genes. These 46 genes were selected primarily from a set of 50 craniofacial genes that also showed genomic signatures of accelerated evolution in a survey of 199 genes (Table S2). Since properly conditioned tests of genetic association in admixed populations are an efficient approach to discover genes affecting traits that differ between populations and since RIP-A is an efficient means of summarizing overall facial ancestry, it is perhaps somewhat counterintuitive that RIP-A conditioning is superior to genomic ancestry conditioning in our partial BRIM modeling (Figures S15, S16, S17, S18, S19, S20 and S27). Likewise, RIP-S proved to be a better conditioning variable than sex in the partial BRIM analyses to estimate RIP-G (Figures S21, S22, S23, S24, S25, S26). We performed ANOVAs to test for average differences in RIP-G by genotype category (e.g., CC, CT, and TT coded as −1, 0, and 1 assuming additive allelic effects). Given the substantial a priori evidence, viz., that these genes show evidence of accelerated evolution in one or both of the parental populations and that mutations in these genes can cause overt murine or human craniofacial dysmorphology, we consider our analysis of each gene to be a separate statistical test and, as such, do not require adjustments for multiple testing. Twenty-four of 76 RIP-G variables (in 20 different genes) show p<0.1 (Table S2). The relatively low threshold was motivated by the strong a priori evidence for each gene noted above, the single trait summary provided by RIP-G, and an expected small effect of single genes on normal-range variation across the whole face. Additionally, given the general finding that clinically relevant genes can also affect subclinical and normal-range variation (e.g., [23]), we performed detailed post hoc descriptions of the effects of these RIP-Gs using FSCPs (Figures S34, S35, Figures S39, S40, S41, S42, S43, S44 and Table S4).
Summaries of the effects of three of these 24 RIP-G variables (rs1074265 in SLC35D1, rs13267109 in FGFR1 and rs2724626 in LRP6) presented in Figures 6A, 6B, and 6C illustrate these results. A detailed analysis and description of each of the 24 SNP effects using FSCPs is given in the supporting material (Text S1). The gene solute carrier family 35 member D1 gene (SLC35D1; OMIM#610804) is located on human chromosome 1p31.3 [24]. Mutations in SLC35D1 have been shown to result in Schneckenbecken dysplasia (OMIM#269250), which affects the face causing the characteristic feature of “superiorly oriented orbits.” The normal-range results of the SNP in rs1074265 in SLC35D1 (Figure 6A) indicate strong effects at the eyes and periorbital regions, including notable differences at the supraorbital region, as well as at the midface and the chin. Mutations in the human fibroblast growth factor receptor 1 (FGFR1;OMIM#136350) gene located on chromosome 8p21.23-p21.22 can result in four autosomal dominant craniofacial disorders: Jackson-Weiss syndrome (OMIM#123150), which is characterized by craniosynostosis and midfacial hypoplasia; trigonocephaly (OMIM#190440), which is characterized by a keel-shaped forehead resulting in a triangle-shaped cranium when viewed from above; osteoglophonic dysplasia (OMIM#166250), which is characterized by craniosynostosis prominent supraorbital ridge and depressed nasal root; and Pfeiffer syndrome (OMIM#101600), which is characterized by midface hypoplasia and, depending on the subtype, ocular proptosis, short cranial base, and cloverleaf skull. The normal-range results of the SNP rs13267109 in FGFR1 depicted in Figure 6B indicate the strongest effects in the supraorbital ridges, the eyes, the midface, the nose, and the corners of the mouth. The strongest differences in the shape transformations are indeed the forehead, supraorbital ridges and nasal bridge. The mouse homologue of the human low-density lipoprotein receptor-related protein 6 (LRP6; OMIM#603507) gene is known to be critical for the development of lips in the mouse resulting in bilateral cleft lips in the knockout LRP6 mouse model [25]. As yet, no human craniofacial diseases have been linked to the LRP6 gene or to the gene region on human chromosome 12p13.2 although the gene product is known to interact on a molecular level with WNT signaling. Observing the shape transformation in Figure 6C, a change from a prominent lip region, including the appearance of a thick and convex vermilion, to a less prominent lip region, including an apparently thinner and less convex (more concave) vermilion, is noted. This is confirmed by inspecting the normal displacement results and the significance maps, in which the lips are clearly delineated (Figure S43).
In general, some RIP-G variables show localized effects (e.g., rs1074265 in SLC35D1), changing only certain aspects in facial shape, while others display changes in several facial regions (e.g., rs13267109 in FGFR1). Summary statistics for the underlying distributions of effect sizes across the quasi-landmarks are presented in Table S3. In the case where multiple SNPs in the same gene are modeled, overlapping and similar effects are seen across the different SNPs for the same gene (e.g., DNMT3B and SATB2) and different SNPs from genes within the same biological pathway (e.g., WNT3, FGFR1, and FGFR2). We present a graphical user interface (GUI) so that effects of changes in these 24 RIP-G variables, RIP-A, RIP-S, or any of the top 44 PC variables can be visualized in more detail. These transformations can be visualized with the texture map as well as shape only, and the GUI (http://tinyurl.com/DNA2FACEIN3D) allows for the illustration of the comparison of transformed faces to the consensus face using the three primary FSCPs.
Since both categorical and continuous variables can be modeled using BRIM, this approach might be used to test for relationships between facial features and other factors, e.g., age, adiposity, and temperament. The methods illustrated here also provide for the development of diagnostic tools by modeling validated cases of overt craniofacial dysmorphology. Most directly, our methods provide the means of identifying the genes that affect facial shape and for modeling the effects of these genes to generate a predicted face. Although much more work is needed before we can know how many genes will be required to estimate the shape of a face in some useful way and many more populations need to be studied before we can know how generalizable the results are, these results provide both the impetus and analytical framework for these studies.
Population samples were collected in the United States (State College, PA, Williamsport, PA, and The Bronx, NY); Brasilia, Brazil; and Cape Verde (São Vicente, and Santiago), all under a Penn State University Internal Review Board (IRB) approved research protocol titled, “Genetics of Human Pigmentation, Ancestry and Facial Features.” Skin pigmentation was measured using narrow-band reflectometry with the DermaSpectrometer (Cortrex Technology, Hadsund, Denmark) in the United States and Brazil and the DSMII (Cortrex Technology, Hadsund, Denmark) in Cape Verde. DermaSpectrometer readings were rescaled to the DSMII scale by multiplying by 1.19, the slope derived from a comparison of readings with both instruments on the same set of participants (data not shown). Height, weight, age, self-reported ancestry, and sex were collected by survey. DNA was collected both with buccal cell brushes and using finger-stick blood on four-circle Whatman FTA cards (Whatman, Florham Park, NJ).
To minimize age-related variation in facial morphology, we only recruited participants between the ages of 18 and 40. From these recruits, we selected individuals with >10% West African ancestry and <15% combined Native American and East-Asian ancestry as measured with the 176 ancestry informative marker (AIM) panel. We assigned these cutoff points to reduce admixture from parental populations other than West African and European. Ancestry-based exclusion criteria were not applied to Cape Verdeans given the largely dihybrid nature of this population. Finally, we excluded participants whose 3D images were obstructed by facial or head hair. After excluding participants by these criteria, we were left with 592 participants (154 from the US, 191 from Brazil, and 247 from Cape Verde).
Genotyping of 176 AIMs for the US and Brazilian samples was performed on the 25 K SNPstream ultra-high-throughput genotyping system (Beckman Coulter, Fullerton, CA) as previously described [11]. Ancestry was estimated using the various panels of AIMs by one of two methods. Ancestry using full set of 176 AIMs was estimated in the US and Brazilian subsample using maximum likelihood on a four-population model; European, West African, Native American, and East Asian [11].The 68-AIM ancestry estimates were generated using the full sample (U.S., Brazilian, and Cape Verdean) using ADMIXMAP as these markers were available on all 592 participants. One marker (rs917502) from the original 176 had a call rate of less than 30% and was omitted from the ADMIXMAP analyses.
The Cape Verdean sample was assayed for the Illumina Infinium HD Human1M-Duo Beadarray (Illumina, San Diego, CA) following the manufacturer's recommendations. A total of 537,895 autosomal SNPs that passed quality controls were used to estimate ancestry using the program FRAPPE [26], assuming two ancestral populations (West African and European). HapMap genotype data, including 60 unrelated European-Americans (CEU) and 60 unrelated West Africans (YRI), were incorporated in the analysis as reference panels (phase 2, release 22, The HapMap Project; [27]).
We identified a list of selection-nominated candidate genes for testing against normal-range facial variation in admixed individuals of European and West African descent. Ancestry information and tests for accelerated evolution [28] were used to prioritize among a larger set of craniofacial genes. Since most genomic regions show low levels of allele frequency change across human populations, genes affecting traits that vary across populations are usually distinctive in showing large differences in frequency and other features of local variation and allele frequency spectra consistent with rapid local evolution. A preliminary set of craniofacial candidate genes was developed by searching the Online Mendelian Inheritance in Man (OMIM) database [24]. The keywords “craniofacial” and “facial” were searched to determine a set of genes known to affect craniofacial development. The OMIM entries for each gene included in the search output were then scanned manually to remove genes where the term appeared as a result of phrases such as “no craniofacial associations found” and other similar negative results. OMIM searching resulted in a list of 199 unique craniofacial candidate genes. Because this work focused on admixed populations of West African and European descent, the statistical power to detect linkage with craniofacial variation is greatest for SNPs that show large allele frequency differences between West African and European parental populations. Therefore, allele frequency differences among parental groups were further used to prioritize among the candidate genes. SNP frequency data in putative parental population (CEPH Europeans (CEU) and Yoruban (YRI) West Africans) for all SNPs within the 199 OMIM candidate genes were pulled from the HapMap database. This reduced subset of genes was then tested for signatures of non-neutral evolution in a 200 kb window surrounding each gene using a combination of three statistical tests: Locus-Specific Branch Length (LSBL) [29], the log of the ratio of the heterozygosities (lnRH) [30], and Tajima's D [31]. Because these tests are inferring different concepts regarding population history, we considered as significant any gene with statistical evidence of selection for all three measures or strong evidence of non-neutral evolution for two measures in either West African and/or European parental populations as a Selection-nominated candidate gene. It is notable that these steps were taken to increase the likelihood that a functional SNP would be available to test the ability of methods like BRIM to model individual gene effects on the human face. We are making no strong claims in this analysis that craniofacial genes generally or this subset in particular have been subject to greater than average levels of non-neutral evolution or that these genes do in fact have genetic variation that is affecting normal range facial variation in this sample. A total of 50 autosomal genes were thus selected (SKI, LMNA, SIL1, EDN1, RSPO2, TRPS1, POLR1D, MAP2K1, ADAMTS10, TBX1, PEX14, HSPG2, CAV3, CTNND2, TFAP2A, PEX6, PEX3, MEOX2, RELN, ROR2, NEBL, CHUK, FGFR2, WT1, PEX16, BMP4, FANCA, RAI1, FOXA2, ECE1, DPYD, ZEB2, SATB2, FGFR3, NIPBL, NSD1, ENPP1, GLI3, COL1A2, BRAF, ASPH, FREM2, SNRPN, FBN1, MAP2K2, RPS19, DNMT3B, GDF5, and UFD1L) and a set of SNPs with high allele frequency differences (delta >0.4) in these 50 craniofacial Selection-nominated candidate genes to test for associations with facial shape variation.
3D images composed of surface and texture maps were taken using the 3dMDface system (3dMD, Atlanta, GA). Participants were asked to close their mouths and hold their faces with a neutral expression for the picture. Images were then exported from the 3dMD Patient software in OBJ file format and imported into a scan cleaning program for cropping and trimming, removing hair, ears, and any dissociated polygons. The complete work flow involved in processing face scans is depicted in Figure 1. Five positioning landmarks were placed on the face to establish a rough facial orientation. Subsequently, an anthropometric mask (7,150 quasi-landmarks) was non-rigidly mapped onto the original 3D images and their reflections [8], [9], [17], which were constructed by changing the sign of the x-coordinate [18], [32]. This established homologous spatially-dense quasi-landmark (Q-L) configurations for all original and reflected 3D images (8). Note that, by homologous, we mean that each quasi-landmark occupies the same position on each face relative to all other quasi-landmarks. Subsequently, a generalized Procrustes superimposition [18], [33] is used to eliminate differences in position, orientation, and scale of both original and reflected configurations combined was performed. This constructed a tangent space of the Kendall shape-space centered on the overall consensus configuration [25]. Procrustes shape coordinates, representing the shape of an object [34], were obtained for all 3D faces and their reflections. After Procrustes superimposition, the overall consensus configuration is perfectly symmetrical and a single shape can be decomposed into its asymmetric and its bilaterally symmetric part [18]. The average of an original and its reflected configuration constitutes the symmetric component while the difference between the two configurations constitutes the asymmetric component [19], [35]. The analyses in this report were all based on facial shape as represented using the component of symmetry only. Although deviations from bilateral symmetry are thought to be the effects of developmental noise and/or environmental factors [36], it is likely there are genetic effects on asymmetry, which would compel independent investigation.
Principal components analysis (PCA) [9] on the superimposed and symmetrized quasi-landmark configurations of the panel of 592 participants resulted in 44 PCs that together summarize 98% of the total variation in face space. To examine the effect of excluding lower PCs, we first reconstructed actual quasi-landmark configuration from the 44 PCs only and compared these to the original remapped face. We found that the average root mean squared error (RMSE) is as small as 0.2 mm per quasi-landmark. The localized differences between the original faces and the faces as represented by the first 44 PCs are largest around the iris, eyelids, under the nose, and the corners and opening of the mouth and are at most about 0.45 mm. How a PC or any other independent variable affects the face can be shown with heat maps and shape transformations: heat maps use contrasting colors to highlight the specific parts of the face that are affected, while shape transformations illustrate the changes in overall face shape with two or more images of the face at set intervals. Shape transformations are obtained from the average face in the direction of each PC at −3 and +3 times the accompanying standard deviation (square-root of the eigenvalue). Figure 2 shows how the first 10 PCs affect the face. Some of these PCs (e.g., PC1, PC2, PC3) summarize effects on many parts of the face, while other PCs (e.g., PC4, PC5) summarize the effects of changes in only particular parts of the face. The effects of each of the 44 PCs as well as the RIP variables can be visualized with a GUI software tool that we have written called DNA2FACEIN3D.EXE. The program and instruction manual can be downloaded here: http://tinyurl.com/DNA2FACEIN3D.
We have used three methods to visualize and quantify facial difference so that we can systematically express the effects of particular response-based imputed predictor (RIP) variables on the face into anatomically interpretable results. These are based on comparing faces pairwise, such as comparing the most feminine RIP-S to the most masculine RIP-S transformed consensus faces using three fundamental measures: area ratio, normal displacement, and curvature ratio. These two ratios and one displacement along with particular inter-landmark distances and angles can together be termed “face shape change parameters” (FSCPs) and are a means of translating face shape changes from the abstract face space into language of facial characteristics such that comparisons between clinical or anthropological descriptions of faces can be compared to bootstrapped response-based imputation modeling (BRIM) results. The statistical significance of these FSCPs can be estimated using permutation. A more detailed description on how this is done is given in the supplementary online material.
|
10.1371/journal.pgen.1003486 | A Statistical Framework for Joint eQTL Analysis in Multiple Tissues | Mapping expression Quantitative Trait Loci (eQTLs) represents a powerful and widely adopted approach to identifying putative regulatory variants and linking them to specific genes. Up to now eQTL studies have been conducted in a relatively narrow range of tissues or cell types. However, understanding the biology of organismal phenotypes will involve understanding regulation in multiple tissues, and ongoing studies are collecting eQTL data in dozens of cell types. Here we present a statistical framework for powerfully detecting eQTLs in multiple tissues or cell types (or, more generally, multiple subgroups). The framework explicitly models the potential for each eQTL to be active in some tissues and inactive in others. By modeling the sharing of active eQTLs among tissues, this framework increases power to detect eQTLs that are present in more than one tissue compared with “tissue-by-tissue” analyses that examine each tissue separately. Conversely, by modeling the inactivity of eQTLs in some tissues, the framework allows the proportion of eQTLs shared across different tissues to be formally estimated as parameters of a model, addressing the difficulties of accounting for incomplete power when comparing overlaps of eQTLs identified by tissue-by-tissue analyses. Applying our framework to re-analyze data from transformed B cells, T cells, and fibroblasts, we find that it substantially increases power compared with tissue-by-tissue analysis, identifying 63% more genes with eQTLs (at FDR = 0.05). Further, the results suggest that, in contrast to previous analyses of the same data, the majority of eQTLs detectable in these data are shared among all three tissues.
| Genetic variants that are associated with gene expression are known as expression Quantitative Trait Loci, or eQTLs. Many studies have been conducted to identify eQTLs, and they have proven an effective tool for identifying putative regulatory variants and linking them to specific genes. Up to now most studies have been conducted in a single tissue or cell type, but moving forward this is changing, and ongoing studies are collecting data aimed at mapping eQTLs in dozens of tissues. Current statistical methods are not able to fully exploit the richness of these kinds of data, taking account of both the sharing and differences in eQTLs among tissues. In this paper we develop a statistical framework to address this problem, to improve power to detect eQTLs when they are shared among multiple tissues, and to allow for differences among tissues to be estimated. Applying these methods to data from three tissues suggests that sharing of eQTLs among tissues may be substantially more common than it appeared in previous analyses of the same data.
| Regulatory variation plays an essential role in the genetics of disease and other phenotypes as well as in evolutionary change [1]–[3]. However, in sharp contrast to nonsynonymous variants in the human genome, which can now be identified with great accuracy, it remains extremely difficult to know which variants in the genome may impact gene regulation in any given tissue or cell type. [Henceforth we use “tissue” for brevity, but everything applies equally to cell types.] Expression QTL mapping (e.g. [4]–[6] represents a powerful approach for bridging this gap, by allowing regulatory variants to be identified, and linked to specific genes. Indeed, numerous studies (e.g. [7], [8]) have shown highly significant overlaps between eQTLs and SNPs associated with organismal-level phenotypes in genome-wide association studies (GWAS), suggesting that a large fraction of GWAS associations may be due to variants that affect gene expression.
Ultimately, understanding the biology of organismal phenotypes, such as diseases, is likely to require understanding regulatory variation in many different tissues ([9], [10]). For example, if regulatory variants differ across tissues, then, in understanding GWAS hits, and using them to understand the biology of disease, we would like to know which variants are affecting which tissues. At a more fundamental level, identifying differential genetic regulation in different tissues could yield insights into the basic biological processes that influence tissue differentiation. To date, eQTL studies have been performed in a relatively narrow range of tissue types. However, this is changing quickly: for example, the NIH “Genotype-Tissue Expression” (GTEx) project aims to collect expression and genotype data in 30 tissues across 900 individuals. Motivated by this, here we describe and illustrate a statistical framework for mapping eQTLs in expression data on multiple tissues.
While statistical methods for identifying eQTLs in a single tissue or cell type are now relatively mature (e.g. [11]) current analytic tools are limited in their ability to fully exploit the richness of data across multiple tissues. In particular, available methods fall short in their ability to jointly analyze data on all tissues to maximize power, while simultaneously allowing for differences among eQTLs present in each tissue. Indeed relatively few papers have considered the problem. The simplest approach (e.g. [12], [13]) is to analyze data on each tissue separately (“tissue-by-tissue” analysis), and then to examine overlap of results among tissues. However, this fails to leverage commonalities among tissues to improve power to detect shared eQTLs. Furthermore, although examining overlap of eQTLs among tissues may appear a natural approach to examining heterogeneity, in practice interpretation of results is complicated by the difficulty of accounting for incomplete power. Both [13] and [14] provide approaches to address this, but only for pairwise comparisons of tissues.
Compared with tissue-by-tissue analysis, joint analysis of multiple tissues has the potential to increase power to identify eQTLs that have similar effects across tissues. Both [15] and [16] conduct such joint analyses – the first using ANOVA, and the second using a weighted -score meta-analysis – and [16] confirm that their joint analysis has greater power than tissue-by-tissue analysis. The ANOVA and -score methods each have different advantages. The ANOVA framework has the advantage that, by including interaction terms, it can be used to investigate heterogeneity in eQTL effects among tissues. Gerrits et al. ([15]) use this to identify eQTLs that show significant heterogeneity, and then classify these eQTLs, post-hoc, into different types based on estimated effect sizes. The weighted -score method has the advantage that, unlike ANOVA, it allows for different variances of expression levels in different tissues (which are likely to occur in practice). However, it does not so easily allow for investigation of heterogeneity; Fu et al. ([16]) hence assess heterogeneity for pairs of tissues by using a resampling-based procedure to assess the significance of observed differences in scores. Other papers, including [17] and [18], also show that joint analyses provide more power. Our work goes beyond these papers in its modeling of heterogeneity, and in its use of a hierarchical model to borrow information across genes to estimate weights associated with different types of heterogeneity.
Here we introduce a statistical framework for the joint analysis of eQTLs among multiple tissue types, that combines advantages of some of the methods above, as well as introducing some new ones. In brief, our framework integrates recently-developed GWAS meta-analysis methods that allow for heterogeneity of effects among groups [19]–[23], into a hierarchical model (e.g. [24], [25]) that combines information across genes to estimate the relative frequency of patterns of eQTL sharing among tissues. Like ANOVA, our approach allows investigation of heterogeneity among several tissues, not just pairs of tissues. However, in contrast to ANOVA, our framework allows for different variances in different tissues. Moreover, unlike any of the methods described above, our framework explicitly models the fact that some tissues may share eQTLs more than others, and estimates these patterns of sharing from the data (a similar idea was applied to ChIP-Seq data by [26]). Our methods also allow for intra-individual correlations when samples are obtained from a common set of individuals. While we focus here on comparing and combining information across different tissue types, our framework could be applied equally to comparing and combining across other units, e.g. different experimental platforms, multiple datasets on the same tissue types, or data on individuals from different populations.
The remainder of the paper is as follows. After providing a brief overview of our framework, we use simulations to illustrate its power compared to other methods, and then apply it to map eQTLs, and assess heterogeneity among tissues, using data from Fibroblasts, LCLs and T-cells ([12]). Consistent with results from [16], we show that our joint analysis framework provides a large gain in power compared with a tissue-by-tissue analysis. Furthermore, compared with previous analyses of these data, we find a much higher rate of tissue-consistent eQTLs.
Consider mapping eQTLs in tissues. In our applications here the expression data are from micro-arrays, and so we assume a normal model for the expression levels, suitably-transformed. (These methods can also be applied to RNA-seq data after suitable transformation; see Discussion). That is, in each tissue, , we model the potential association between a candidate SNP and a target gene by a linear regression:(1)where denotes the observed expression level of the target gene in tissue for the individual, the mean expression level of this gene in tissue , the effect of a candidate SNP on this gene expression in tissue , the genotype of the individual at the SNP (coded as 0,1 or 2 copies of a reference allele) and the residual error for tissue and individual . Note that the subscript on residual variance indicates that we allow the residual variance to be different in each tissue. In addition, when tissues are sampled from the same set of individuals, we allow that the residual errors may be correlated (with the correlation matrix to be estimated from the data).
The primary questions of interest are whether the SNP is an eQTL in any tissue, and, if so, in which tissues. To address these questions we use the idea of a “configuration” from [21], [23]. A configuration is a binary vector where indicates whether the SNP is an eQTL in tissue . If then we say the eQTL is “active” in tissue . The “global null hypothesis”, , that the SNP is not an eQTL in any tissue, is therefore . Every other possible value of can be thought of as representing a particular alternative hypothesis. For example, represents the alternative hypothesis that the SNP is an eQTL in all tissues, and represents the alternative hypothesis that the SNP is an eQTL in just the first tissue.
Our aim is to perform inference for . A natural approach is to specify a probability model, , being the probability of obtaining the observed data if the true configuration were , and then perform likelihood-based inference for . The support in the data for each possible value of , relative to the null , is quantified by the likelihood ratio, or Bayes Factor (BF, [27]):(2)Specifying these likelihoods requires assumptions about , the distribution of the effect sizes for each possible configuration (as well as less crucial assumptions about nuisance parameters such as and ). Of course, if then by definition, but for the tissues where various assumptions are possible – for example, one could assume that the effect is the same in all these tissues, or allow it to vary among tissues. Here we use a flexible family of distributions, (see Methods), where the hyper-parameters can be varied to control both the typical effect size, and the heterogeneity of effects across tissues (see below).
The value of measures the support in the data for one specific alternative configuration , compared against the null hypothesis . To account for the fact that there are many possible alternatives, the overall strength of evidence against at the candidate SNP is obtained by “Bayesian Model Averaging” (BMA), which involves averaging over the possible alternative configurations , weighting each by its prior probability, :(3)Further, under an assumption of at most one eQTL per gene, the overall evidence against for the entire gene (i.e. that the gene contains no eQTL in any tissue) is given by averaging across all candidate SNPs [28]. In either case, at either the SNP or gene level, large values of constitute strong evidence against . has a direct Bayesian interpretation as the strength of the evidence against , but here we also use it as a frequentist test statistic ([28], [29]), assessing significance by permutation or simulation. The latter has the advantage that and obtained in this way are “valid” even if not all the prior assumptions are exactly correct.
Note that depends on the choice of , and the power of as a test statistic is expected to depend on how well this choice of these hyper-parameters captures the range of alternative scenarios present in the data. Here we make use of three different choices:
Each of these choices has something to recommend it. The first, being data driven, is the most attractive in principle, but also the most complex to implement. The default choice is simpler to implement, and is included partly to demonstrate that one does not have to get the hyper-parameter values exactly “right” for to be a powerful test statistic. Finally, has the advantage that it is easily applied to large numbers of tissues; neither of the other methods scales well, either computationally or statistically, with the number of tissues, because the number of terms in the sum in equation (3) is .
When there is strong evidence against , the Bayes Factors can also be used to assess which alternative configurations are consistent with the data. Specifically the posterior probability on each configuration is:(4)and the posterior probability that the SNP is an eQTL in tissue is obtained by summing the probabilities over configurations in which :(5)The second of these is particularly helpful when the data are informative for an eQTL in tissue , but ambiguous in other tissues: in such a case the probability will be close to 1, even though the “true” configuration will be uncertain (so none of the probabilities (4) will be close to 1). Because both (4) and (5) are sensitive to choice of hyper-parameters, we compute them using (where the hyper-parameters are estimated from the data).
Further details of methods used are provided in the Methods section.
We now analyze data from [12], consisting of gene expression levels measured in fibroblasts, LCLs and T-cells from 75 unrelated individuals genotyped at approximately 400,000 SNPs. The data were pre-processed similarly to the original publication, as described in the Methods section. Throughout we focus on testing SNPs that lie within 1 Mb of the transcription start site of each gene (the “cis candidate region”), and on a subset of 5012 genes robustly expressed in all three cell-types.
In this work, we have presented a statistical framework for analyzing and identifying eQTLs, combining data from multiple tissues. Our approach considers a range of alternative models, one for each possible configuration of eQTL sharing among tissues. We compute Bayes Factors that quantify the support in the data for each possible configuration, and these are used both to develop powerful test statistics for detecting genes that have an eQTL in at least one tissue (by Bayesian model averaging across configurations), and to identify the tissue(s) in which these eQTLs are active (by comparing the Bayes factors for different configurations against one another). Our framework allows for heterogeneity of eQTL effects among tissues in which the eQTL is active, for different variances of gene expression measurements in each tissue, and for intra-individual correlations that may exist due to samples being obtained from the same individuals. For eQTL detection, our framework provides consistent, and sometimes substantial, gains in power compared to a tissue-by-tissue analysis and ANOVA or simple linear regression. Concerning the tissue specificity of eQTLs, our framework efficiently borrows information across genes to estimate configuration proportions, and then uses these estimates to assess the evidence for each possible configuration. When re-analyzing the gene expression levels in three cell types from 75 individuals ([12]), we found that there appears to be a substantial amount of sharing of eQTLs among tissues, substantially more than suggested by the original analysis.
In the next few years, we expect that expression data will be available on large numbers of diverse tissue types in sufficient sample sizes to allow eQTLs to be mapped effectively (for example, the NIH GTEx project aims to collect such data). The methods presented here represent a substantive step towards improved analyses that fully exploit the richness of these kinds of data. However, we also see several directions for potential extensions and improvements. First, our current framework can only partially deal with the challenges of large numbers of tissues. Specifically, because with tissues, there are possible configurations of eQTL sharing among tissues, some of our current methods, which consider all possible configurations, will become impractical for moderate (speculatively, above about 10, perhaps). Our test statistic partially addresses this problem, by allowing for heterogeneity while averaging over only configurations, which is practical for very large . Our simulation results suggest that is a powerful test statistic for identifying SNPs that are an eQTL in at least one tissue. However our preferred approach for identifying which tissues such SNPs are active in involves a hierarchical model that estimates the frequency of different patterns of sharing from the data, and this hierarchical model scales poorly with . In particular, having a separate parameter for each possible configuration is unattractive (both statistically and computationally) for large , and alternative approaches will likely be required. There are several possible ways forward here: for example, one would be to reduce the number of distinct configurations by clustering “similar” configurations together; another would be to focus less on the discrete configurations, and instead to focus on modeling heterogeneity in effect sizes in a continuous way - perhaps using a mixtures of multivariate normal distributions with more complex covariance structures than we allow here. We expect this to remain an area of active research in the coming years, especially since these types of issues will likely arise in many genomics applications involving multiple cell types, and not only in eQTL mapping.
Another important issue to address is that most future expression data sets will likely be collected by RNA-seq, which provides count data that are not normally distributed. Previous eQTL analyses of RNA-seq (e.g. [32]) have nonetheless performed eQTL mapping using a normal model, by first transforming (normalized) count data at each gene to the quantiles of a standard normal distribution. Although this approach would not be attractive in experiments with small sample sizes, with the moderate to large sample sizes typically used in eQTL mapping experiments this approach works well. As a first step, this approach could also be used to apply our methods to count data. However, ultimately it would seem preferable to replace the normal model with a model that is better adapted to count-based data, perhaps a quasi-Poisson generalized linear model ([33]); Bayes Factors under these models could be approximated using Laplace approximations, similar to the approximations used here for the normal model [21]. The quasi-Poisson model has the advantage over the normal transformation approach that it preserves the fact that there is more information about eQTL effects in tissues where a gene is high expressed than in tissues where it is low expressed. This information is lost by normal transformation. In our primary analyses here we addressed this by analyzing only genes that were robustly expressed in all tissues, but this is sub-optimal, and will become increasingly unattractive as the number of tissues grows.
Our analyses here assess (cis) eQTL sharing among tissues by performing association testing at the level of individual SNPs. A different approach to investigating eQTL sharing among tissues is to study the “cross-heritability” of expression levels among tissues (e.g. [34], [35]). These methods are based on polygenic models, and attempt to estimate the combined influence of all shared eQTLs; this contrasts with our analysis, where the focus is on sharing of individually-identifiable eQTLs of moderate-to-large effect. Both [34] and [35] estimate cross-tissue heritability to be low. [34], studying expression in Blood and Adipose tissues from Icelanders, estimated cross-tissue heritability as 3%; [35] obtained an estimate of mean genetic correlation close to zero for Blood and LCLs in monozygotic twins (). These results may appear to conflict with our results (both from our model-based approach, and the less-model-based pairwise analysis approach from [13]), which suggest that most large-effect cis eQTLs are shared among fibroblasts, LCLs and T cells. However, these low estimates of cross-tissue heritability reflect not only the extent of sharing of eQTLs, but also the absolute size of the eQTL effects. If eQTL effects are small, explaining only a small proportion of the total variance in gene expression, then cross-tissue heritability will be also small, even if all eQTLs have exactly the same effect in all tissues. Thus, to assess eQTL sharing in the heritability-based approaches, it is helpful to contrast cross-tissue heritability, , with within-tissue heritability, , (which is also affected by eQTL effect size, but not by sharing). Specifically, within the polygenic model it can be shown that the correlation coefficient of the eQTL effects in two tissues and is: . Applying this to the cis estimates of and from [34], for adipose and blood, yields . Although this estimate of effect correlation within a polygenic model, is not directly comparable with our estimate of sharing of eQTLs in a decidedly non-polygenic model (and for different cell types!), this result suggests that the two analyses may be less in conflict than they initially appear.
Software implementing our methods are available on the website http://stephenslab.uchicago.edu/software.html.
For we use a hierarchical model, similar to [24], [25], which combines information across genes, to estimate the grid weights 's and configuration weights 's. Following both [25], [26] we make the simplifying assumption that each gene has at most one eQTL (which may be active in multiple tissues), and that each SNP is equally likely to be the eQTL. Let be the number of SNPs in the cis-region for gene . Then, if denotes the Bayes Factor (12) computed for SNP in gene , the “overall Bayes Factor” measuring the evidence for an eQTL in gene , , is obtained by averaging over the possible eQTL SNPs, the possible configurations , and the grid of values for , weighting by their probabilities:(13)Furthermore, if we let denote the probability that each gene follows the null (i.e. contains no eQTL) then the likelihood for gene , as a function of , is given by(14)(15)The overall likelihood for our hierarchical model is obtained by multiplying these likelihoods across genes:(16)Note that although the expression levels for different genes are not independent, because the SNPs being tested in different genes are mostly independent this independence assumption for the likelihoods across genes is a reasonable starting point. We have developed an EM algorithm to estimate the parameters by maximum likelihood (see Supplementary information).
For our simulations, when simulating SNP-gene pairs, the genotypes at each SNP in each individual were simulated as Binomial(2,0.3): that is, with minor allele frequency 30% and assuming Hardy-Weinberg equilibrium. Phenotypes with eQTLs were simulated, with effect size based on an expected proportion of variance explained (PVE) of 20%; (see Text S1). For Figure 1A and 1B the error variances (one per tissue) were all equal to 1. For Figure 1C the error variances were randomly drawn from , all equally likely.
The ANOVA/LR method uses the same linear model as our Bayesian methods (1), except that the residual errors are assumed to be equal across tissues . Within this model we tested the global null hypothesis ( for all ) using an test comparing the null model with the unconstrained alternative ( unconstrained). See Text S1.
The phenotypes from Dimas et al. ([12]) were retrieved from the Gene Expression Omnibus (GSE17080). We mapped the 22,651 non-redundant probes to the hg19 human genome reference sequence (only the autosomes) using BWA ([37]), kept 19,965 probes mapping uniquely with at most one mismatch, and removed the probes overlapping several genes from Ensembl. This gave us 12,046 genes overlapped by 16,453 probes. For genes overlapped by multiple probes, we chose a single probe at random. In our analyses we considered only genes that were robustly expressed in all tissues. A gene was considered robustly expressed in a given tissue if its mean expression level across individuals in this tissue was larger than or equal to the median expression level of all genes across all individuals in this tissue. As a result, we focused on 5012 genes.
Genotypes were obtained from the European Genome-phenome Archive (EGAD00000000027). We extracted the genotypes corresponding to the 85 individuals for which we had phenotypes and converted the SNP coordinates to the hg19 reference using liftOver ([38]). To detect outliers, we performed a PCA of these genotypes using individuals from the CEU, CHB, JPT and YRI populations of the HapMap project using EIGENSOFT ([39]). As in the original study, we identified 10 outliers and removed them from all further analyses, which were therefore performed on 75 individuals.
Gene expression measurements suffer from various confounders, many of which may be unmeasured ([40]), but which can be corrected for using methods such as principal components analysis (PCA). Following [32], we applied PCA in each tissue separately on the 501275 matrix of expression levels of each gene in each individual. We sorted principal components (PCs) according to the proportion of variation in the original matrix they explain, and selected PCs so that adding another PC would explain less than 0.0025% of the variation. As a result, this procedure identified 16 PCs in Fibroblasts, 7 in LCLs and 15 in T-cells. We then regressed out these PCs from the original matrix of gene expression levels, and used the residuals as phenotypes for all analyses.
All methods we compared assume that the errors are distributed according to a Normal distribution. Before analysis we therefore rank-transformed the expression levels at each gene to the quantiles of a standard Normal distribution ([28]).
On the data set from Dimas et al., we assessed the performance of two methods, the tissue-by-tissue analysis and the BMA joint analysis, by comparing the number of genes identified as having at least one eQTL in any tissue, at a given FDR. For each method, we defined a test statistic, which was computed for each gene. For the tissue-by-tissue analysis, the test statistic is the minimum of the linear regressions between the given gene and each cis SNP in each tissue (so the minimum is taken across all SNPs and all tissues). For the BMA joint analysis, the test statistic is the average of the Bayes Factors for the given gene and each cis SNP. (When applying the tissue-by-tissue analysis to test for eQTLs in a single tissue, the test statistic is the minimum of the linear regressions between the given gene and each cis SNP in that tissue.)
In each case we converted the test statistic to a for each gene, testing the null hypothesis that the gene contains no eQTL in any tissue, by comparing the observed test statistic with the value of the test statistic obtained on permuted data obtained by permuting the individuals labels (using the same permutations in each tissue to preserve any intra-individual correlations between gene expression in different tissues). Specifically, let denote the total number of permutations (we used ), the value of the test statistic for gene on the non-permuted data, and the value of the test statistic on the -permuted data. The for gene from the tissue-by-tissue analysis is: . For the BMA joint analysis, the is: . Note that permutations were performed for each gene, since the null distribution of the test statistic will vary across genes (not least because the genes have different numbers of SNPs in their cis candidate region; see Figure S3).
From the calculated for each gene we estimate using the qvalue package ([31]), and determine the number of genes having at least one eQTL in any tissue at an FDR of by computing the number of genes with .
When performing the tissue-by-tissue analysis on a single tissue, we performed the permutations in each tissue separately.
|
10.1371/journal.pbio.1001041 | A Genetically Encoded Tag for Correlated Light and Electron Microscopy of Intact Cells, Tissues, and Organisms | Electron microscopy (EM) achieves the highest spatial resolution in protein localization, but specific protein EM labeling has lacked generally applicable genetically encoded tags for in situ visualization in cells and tissues. Here we introduce “miniSOG” (for mini Singlet Oxygen Generator), a fluorescent flavoprotein engineered from Arabidopsis phototropin 2. MiniSOG contains 106 amino acids, less than half the size of Green Fluorescent Protein. Illumination of miniSOG generates sufficient singlet oxygen to locally catalyze the polymerization of diaminobenzidine into an osmiophilic reaction product resolvable by EM. MiniSOG fusions to many well-characterized proteins localize correctly in mammalian cells, intact nematodes, and rodents, enabling correlated fluorescence and EM from large volumes of tissue after strong aldehyde fixation, without the need for exogenous ligands, probes, or destructive permeabilizing detergents. MiniSOG permits high quality ultrastructural preservation and 3-dimensional protein localization via electron tomography or serial section block face scanning electron microscopy. EM shows that miniSOG-tagged SynCAM1 is presynaptic in cultured cortical neurons, whereas miniSOG-tagged SynCAM2 is postsynaptic in culture and in intact mice. Thus SynCAM1 and SynCAM2 could be heterophilic partners. MiniSOG may do for EM what Green Fluorescent Protein did for fluorescence microscopy.
| Electron microscopy (EM) once revolutionized cell biology by revealing subcellular anatomy at resolutions of tens of nanometers, well below the diffraction limit of light microscopy. Over the past two decades, light microscopy has been revitalized by the development of spontaneously fluorescent proteins, which allow nearly any protein of interest to be specifically tagged by genetic fusion. EM has lacked comparable genetic tags that are generally applicable. Here, we introduce “miniSOG”, a small (106-residue) fluorescent flavoprotein that efficiently generates singlet oxygen when illuminated by blue light. In fixed tissue, photogenerated singlet oxygen locally polymerizes diaminobenzidine into a precipitate that is stainable with osmium and therefore can be readily imaged at high resolution by EM. Thus miniSOG is a versatile label for correlated light and electron microscopy of genetically tagged proteins in cells, tissues, and organisms including intact nematodes and mice. As a demonstration of miniSOG's capabilities, controversies about the localization of synaptic cell adhesion molecules are resolved by EM of miniSOG fusions in neuronal culture and intact mouse brain.
| The most general techniques for imaging specific proteins within cells and organisms rely either on antibodies or genetic tags. EM is the standard technique for ultrastructural localization, but conventional EM immunolabeling remains challenging because of the need to develop high-affinity, high-selectivity antibodies that recognize cross-linked antigens, and because optimal preservation of ultrastructure and visibility of cellular landmarks requires strong fixation that hinders diffusibility of antibodies and gold particles. Thus the target proteins most easily labeled are those exposed at cut tissue surfaces. Replacement of bulky gold particles by eosin enables catalytic amplification via photooxidation of diaminobenzidine (DAB), but eosin-conjugated macromolecules still have limited diffusibility and need detergent permeabilization to enter cells [1]. Genetic labeling methods should overcome many of these shortcomings, just as fluorescent proteins have revolutionized light microscopic imaging in molecular and cell biology [2]. However, no analogous genetically encoded tag for EM contrast has yet proven widely applicable. Metallothionein has been proposed as a genetic tag that can noncatalytically incorporate cadmium or gold [3], but its main applications to intact cells have been to Escherichia coli conditioned to tolerate 0.2 mM CdCl2 for 18 h [4] or 10 mM AuCl for 3 h [4],[5]. Such high concentrations of heavy metal salts would not seem readily transferable to most multicellular organisms or their cells. Also many higher organisms express endogenous metallothionein, which would contribute background signals unless genetically deleted or knocked down [5]. Horseradish peroxidase can be a genetic label in the secretory pathway but is greatly limited by its requirements for tetramerization, glycosylation, and high Ca2+, so that it is not functional when expressed in the cytosol [6]. Furthermore, its DAB reaction product tends to diffuse from sites of enzymatic generation, resulting in poorer resolution than immunogold or the reaction product of photogenerated singlet oxygen (1O2, the metastable excited state of O2) with DAB [1],[7],[8]. The best previous genetically targetable generator of 1O2 was the biarsenical dye ReAsH, which binds to genetically appended or inserted tetracysteine motifs [9]. However, ReAsH has modest 1O2 quantum yield (0.024) (Figure S1), requires antidotes to prevent cell toxicity, needs careful precautions to reduce nonspecific background signal, and has been difficult to apply to multicellular tissues and organisms [10]. Although fluorescence photooxidation using GFP has been reported [11],[12], the 1O2 quantum yield of the naked GFP chromophore is extremely low (0.004), and the 1O2 quantum yield of the intact protein was yet lower and unquantifiable [13], presumably because the beta-barrel of the protein shields the chromophore from oxygen. The phototoxic fluorescent protein “Killer Red” [14] is now acknowledged not to work through 1O2 [15], and we have confirmed that its 1O2 quantum yield is negligible (Figure S1).
Here, we introduce miniSOG, a small, genetically encodable protein module that needs no exogenous cofactors to fluoresce and photogenerate 1O2 with a substantial quantum yield. MiniSOG provides major improvement in correlated light and electron microscopy in cells and multicellular organisms via photooxidation techniques.
The LOV (light, oxygen, and voltage) domain of phototropin (a blue light photoreceptor) binds flavin mononucleotide (FMN) [16],[17], which by itself is an efficient singlet oxygen photosensitizer [18]. FMN is ubiquitous in cells and performs indispensable biological functions such as mitochondrial electron transport, fatty acid oxidation, and vitamin metabolism [19]. In phototropin, the excited state energy of FMN is consumed to form a covalent bond with a cysteine [20]. To divert this energy into 1O2 generation, we carried out saturation mutagenesis of the relevant cysteine (Cys426) of the LOV2 domain of Arabidopsis thaliana phototropin 2 (AtPhot2). To screen for optimal 1O2 production, these site-specific mutants were fused to an infrared fluorescent protein, IFP1.4, which is readily bleached by 1O2 (Figure S2) [21]. Colonies of E. coli expressing the fusion proteins were imaged in the IFP channel (ex 684/em708 nm) before and after blue light (488 nm) illumination (Figure 1A). Several colonies showed a decrease of IFP fluorescence from wild-type colonies and two with the largest decrease (∼70%) had the single site substitution of Cys426 to Gly. The small side chain of the glycine residue may provide space around the cofactor that would allow O2 close apposition to FMN for efficient energy transfer. To increase the brightness of the C426G mutant, we also performed saturation mutagenesis of other residues surrounding the chromophore binding site. DNA shuffling of the improved mutants plus random mutagenesis led to a new protein, miniSOG (106-residue) (Figure 1B and C, Figure S3), which absorbs maximally at 448 nm with a shoulder at 473 nm with extinction coefficients (16.7±0.7)×103 and (13.6±0.5)×103 M−1cm−1, respectively (Figure 1D). Excitation of miniSOG leads to green emission with two peaks at 500 and 528 nm (Figure 1D). The 1O2 quantum yield of miniSOG (0.47±0.05) was measured using anthracene-9,10-dipropionic acid (ADPA) as 1O2 sensor (Figure 1E) [22]. Free FMN was used as the standard for the measurement of 1O2 generation (quantum yield 0.51) [10].
MiniSOG was determined by light scattering to be monomeric in solution, with a molecular weight of 13.9±0.4 kDa, close to the theoretical value of 15.3 kDa. Absence of oligomerization was further supported by the good separation by gel filtration of miniSOG from its tandem dimer (td-miniSOG) (Figure S4). Mass spectrometry confirmed that the flavin cofactor is FMN (Figure S5). Equilibrium dialysis reported a dissociation constant of 170±8 pM (Table S2), similar to values for some flavoproteins (e.g. 260±60 pM for a flavodoxin [23]) and consistent with the crystal structures of LOV domains, which show FMN deeply buried inside the protein core [24]. Furthermore, overexpression of miniSOG in HEK293 cells caused the FMN content to increase ∼3-fold, presumably to keep miniSOG nearly saturated with FMN (Figures S6–S8), but caused no obvious toxicity in the absence of light (Table S1). Feedback pathways involving enzymes such as riboflavin kinase (EC 2.7.1.26) and FAD (flavin adenine dinucleotide) diphosphatase (EC 3.6.1.18) probably regulate intracellular FMN to titrate endogenous flavoproteins and miniSOG [25]. Riboflavin kinase phosphorylates riboflavin into FMN, while FAD diphosphatase catalyzes the production of FMN from FAD.
We used the fluorescence from miniSOG fusion proteins to successfully localize a wide variety of proteins and organelles in cultured mammalian cells (Figure 2). Its green fluorescence, while modest compared to GFP (quantum yield of 0.37 versus 0.6), revealed that labeled components appeared to have correct localizations (Figure 2A–H). Figure 2A shows ER-targeted miniSOG, indicating that miniSOG can work within the secretory pathway. Figure 2B–F show Rab5a, zyxin, tubulin, β-actin, and α-actinin as examples of proteins tagged in cytosolic compartments. Mitochondrial targeting and nuclear histone 2B-fusions (Figure 2G,H) show that miniSOG expresses within those organelles. Using the fluorescence and photo-generated 1O2 from miniSOG for fluorescence photooxidation of DAB (Figure 3A), correlated confocal and EM imaging could be performed with several miniSOG fusion proteins (Figure 3B–E), producing excellent EM contrast, efficient labeling, and good preservation of ultrastructure.
The successful localization of a variety of proteins by light and EM in cultured cells as well as mitochondria in C. elegans and SynCAM2 in intact mouse brain demonstrates the value of miniSOG for correlated light and EM localization of specific proteins in cells and multicellular organisms. MiniSOG is advantageous over conventional immuno-gold staining because the protein of interest is genetically tagged before fixation and all subsequent components (O2, DAB, and OsO4) are small molecules that easily permeate tissues. Tissues or cells can be fixed using established methods for good preservation of ultrastructure without concern for retention of antigenicity. Thus, permeabilizing detergents such as Triton X-100 that degrade membranes to facilitate the diffusion of bulky antibodies and secondary labels are unnecessary. This is demonstrated by the well-preserved ultrastructure in SynCAM-miniSOG labeled mice where unlabeled synapses (arrowhead), nonsynaptic plasma membrane, and synaptic vesicles are clearly observed (Figure 5). Such landmarks were essential to assign the precise location of the SynCAMs. While super-resolution fluorescence techniques [38]–[40] could provide improved localizations, each landmark of interest would need to be labeled with fluorophores emitting at different color.
MiniSOG probes have several advantages over other correlated LM/EM probes. MiniSOG needs no exogenous cofactors and produces 1O2 with about 20 times higher quantum efficiency than ReAsH on a tetracysteine motif. Therefore, miniSOG photooxidation has considerably better sensitivity and lower background than ReAsH labeling. MiniSOG is much smaller than GFP, and unlike GFP can mature and become fluorescent in the absence of O2. GFP-based photooxidation is very difficult due to its extremely low 1O2 quantum yield [13]. Genetically encoded horseradish peroxidase is tetrameric and far larger than GFP, only becomes functional inside the secretory pathway [6], and produces relatively diffuse precipitates [1],[7],[8]. Metallothionein fusions would seem most appropriate for purified macromolecules [3], because imaging of intact cells requires them to survive prolonged incubation in high concentrations of Cd2+ or Au+ [4],[5] and not to express endogenous metallothionein.
Our results with miniSOG fusions demonstrate that SynCAM1 and SynCAM2 are localized to pre- and post-synaptic membranes, respectively, and these observations are consistent with the reported strong heterophilic interaction between SynCAM1 and SynCAM2 in the formation of trans-synaptic structures [41]. The presynaptic membrane localization of SynCAM1 is also consistent with the recent report that SynCAM1 is expressed in growth cones in the early developmental stages of mouse brain and is involved in shaping the growth cones and the assembly of axo-dendritic contact [41]. Analogous trans-synaptic pairs include neurexin/neuroligin [42], EphrinB/EphB, and netrinG/netrin-G ligand (NGL). New synaptic proteins continue to be reported, such as leucine rich repeat transmembrane proteins (LRRTMs), NGL-3, and leukocyte common antigen-related (LAR) [43],[44]. The large variety of these molecules may be necessary to establish and support the great diversity of neuronal synapses; dissecting their locations within synapses will be a complex task.
As demonstrated here, our miniSOG-based photooxidation technique provides a method to determine the detailed distribution of these and other important macromolecules. In combination with SBFSEM, miniSOG fusion proteins should find wide applications in the ultrastructural localization of proteins, including 3-d reconstruction of neuronal circuits by large scale automated SBFSEM to mark cells of interest and trace them across large numbers of sections (Figure S13) [37]. Additionally, a logical next step will be to further enhance the preservation of cellular ultrastructure in these types of specimens by combining chemical fixation and high pressure freezing [45] with photooxidation using miniSOG.
Spatiotemporally controlled local photogeneration of 1O2 should also be useful for rapidly inactivating proteins of interest [46], reporting protein proximities over tens of nanometers [47] by 1O2 transfer from a SOG to a 1O2 sensitive fluorescent protein (e.g. IFP1.4) and ablating cells by photodynamic damage. Thus, further development and application of miniSOG using 1O2 generation should greatly expand its utility in imaging and functional studies.
A gene encoding LOV2 domain of Phototropin 2 with codons optimized for E. coli was synthesized by overlap extension PCR [48]. Genetic libraries were constructed by saturation and random mutagenesis and DNA shuffling [21]. Mutants were fused to IFP1.4 by overlap extension PCR and cloned into a modified pBAD vector containing the heme oxygenase-1 gene from cyanobacteria [21]. Libraries were expressed in E. coli strain TOP10 and screened by imaging the agar plates with colonies in the IFP channel before and after blue light illumination [21]. Protein purification and spectroscopic characterization experiments were done as described [49].
DNA encoding miniSOG with codons optimized for mammals was synthesized by overlap extension PCR [48]. MiniSOG fusions were cloned into pcDNA3.1 vector. HEK293 and HeLa cells were transfected with miniSOG or chimera cDNAs using Fugene, then imaged 24–48 h later. Cultured cortical neurons were transfected by Amaxa electroporation (Lonza AG, Germany) and imaged 1–2 wk later.
Transfected cells cultured on glass bottom culture dishes (P35G-0-14-C, MatTek Corp., Ashland, MA) were fixed with 2% glutaraldehyde (Electron Microscopy Sciences, Hatfield, PA) in pH 7.4 0.1 M sodium cacodylate buffer (Ted Pella Inc., Redding, CA) for 30–60 min, rinsed several times in chilled buffer, and treated for 30 min in blocking buffer (50 mM glycine, 10 mM KCN, and 5 mM aminotriazole) to reduce nonspecific background reaction of diaminobenzidine (DAB). Confocal images were taken with minimum exposure using a BioRad MRC-1024 inverted confocal microscope or similar inverted fluorescence microscope to identify transfected cells and for correlative light microscopic imaging. Detailed protocols for performing fluorescence photooxidation of DAB have been published [2],[6]. It is important to use an inverted microscope to ensure direct open access to the DAB solution. An objective of numerical aperture ≥0.7 is desirable to maximize illumination intensity. For photooxidation, diaminobenzidine tetrahydrochloride (Sigma-Aldrich, St. Louis, MO) was freshly diluted to 1 mg/ml in 0.1 M sodium cacodylate buffer, pH 7.4, filtered through a 0.22 micron syringe filter (Millipore), and placed on ice and added to the cells. The region of interest was identified by the fluorescence and an image recorded with care not to bleach the area. A small tube attached to an oxygen tank was placed near the top of the dish and a stream of pure oxygen was gently blown continuously over the top of the solution. Alternately, the DAB solution on ice was bubbled with oxygen and the solution in the dish refreshed every few minutes. The samples were then illuminated using a standard FITC filter set (EX470/40, DM510, BA520) with intense light from a 150W xenon lamp. Illumination was stopped as soon as a very light brown reaction product began to appear in place of the green fluorescence as monitored by transmitted light (typically 2–10 min, depending on the initial fluorescence intensity, the brightness of the illumination, and the optics used). Care was taken to avoid overreacting the samples, as this can lead to overstaining and the degradation of ultrastructure in the region of photooxidation. Multiple areas on a single dish could be reacted if the solution was refreshed every few minutes. The cells were then removed from the microscope and washed in chilled buffer (5×2 min) and post-fixed in 1% osmium tetroxide (Electron Microscopy Sciences) in 0.1 M sodium cacodylate buffer for 30 min on ice. Cells were washed in chilled buffer twice and rinsed in distilled water, then en bloc stained with 2% aqueous uranyl acetate (Ted Pella Inc.) for 1 h to overnight at 4°C. The samples were then dehydrated in a cold graded ethanol series (20%, 50%, 70%, 90%, 100%, 100%) 2 min each, rinsed once in room temperature anhydrous ethanol, and infiltrated in Durcupan ACM resin (Electron Microscopy Sciences) using 1∶1 anhydrous ethanol and resin for 30 min, then 100% resin 2×1 h, then into fresh resin and polymerized in a vacuum oven at 60°C for 48 h.
Transgenic worms were made by injection of cDNAs of mitochondrially targeted miniSOG driven by myo-3 promoter at 50 ng/µl. The worms were chemically fixed with 2% glutaraldehyde, washed, and blocked as described above. The cuticle was sharply cut to allow diffusion of DAB into the inner body for photooxidation. After confocal imaging and fluorescence photooxidation, the worms were processed for EM imaging as described above.
Endotoxin-free DNA (∼3 µg) of the SynCAM2-miniSOG fusion construct was delivered into the lateral ventricle of embryos by in utero electroporation [50]. The offspring at p7 or p21 were anesthetized and fixed by vascular perfusion as previously described [51] with Ringer's solution followed by 4% formaldehyde made fresh from paraformaldehyde (Electron Microscopy Sciences) in 0.15 M cacodylate buffer. Brains were removed and placed in the same fixative at 4°C for 1 h for p21 and overnight for p7. In this case we avoided glutaraldehyde in combination with paraformaldehyde due to the increased autofluorescence that occurs with glutaraldehyde. The autofluorescence obscured miniSOG fluorescence and made it impossible to locate transfected neurons in the brain slices for photooxidation. Brains were then sliced to 100 µm sections using a vibratome (Leica). Areas of interest were identified by confocal microscopy. The sections were then postfixed with 2% glutaraldehyde for 30 min, rinsed in cold buffer, blocked, and then photooxidized as described above. Subsequent procedures for EM processing were similar to those described above except the vibratome sections were resin embedded between two liquid release agent coated glass slides (Electron Microscopy Sciences).
Photooxidized areas of embedded cultured cells were identified by transmitted light and the areas of interest were sawed out using a jeweler's saw and mounted on dummy acrylic blocks with cyanoacrylic adhesive. The coverslip was carefully removed, ultrathin sections were cut using an ultramicrotome, and electron micrographs recorded using a 1200 TEM (JEOL) operating at 80 keV. For tissue sections, one of the glass coverslips was removed using a razorblade and the area of interest identified by transmitted light microscopy. The tissue was removed from the slide, mounted, sectioned, and imaged as above. For electron tomography, 0.5 micron thick sections of cells expressing photooxidized H2B-miniSOG were cut and imaged using a 4000 IVEM (JEOL) operated at 400 keV. Images were tilted and recorded every 2° from ±60° to −60°. The image stack was aligned and reconstructions were obtained using R-weighed back projection methods with the IMOD tomography package. For serial block face scanning electron microscopy, a 3View system (Gatan Inc., Pleasanton, CA) mounted in a Quanta FEG scanning electron microscope (FEI Company, Eindhoven, The Netherlands) was employed. Imaging was performed as previously described [37]. Individual image planes were hand segmented to outline the plasma membrane of the target neuron and denote labeled post-synaptic densities, then thresholded and projected using Amira (Visage Imaging, Germany).
|
10.1371/journal.pntd.0003825 | Human Immunodeficiency Virus, Antiretroviral Therapy and Markers of Lymphatic Filariasis Infection: A Cross-sectional Study in Rural Northern Malawi | Lymphatic filariasis (LF) and human immunodeficiency virus (HIV) are major public health problems. Individuals may be co-infected, raising the possibility of important interactions between these two pathogens with consequences for LF elimination through annual mass drug administration (MDA).
We analysed circulating filarial antigenaemia (CFA) by HIV infection status among adults in two sites in northern Malawi, a region endemic for both LF and HIV. Stored blood samples and data from two geographically separate studies were used: one a recruitment phase of a clinical trial of anti-filarial agent dosing regimens, and the other a whole population annual HIV sero-survey. In study one, 1,851 consecutive adult volunteers were screened for HIV and LF infection. CFA prevalence was 25.4% (43/169) in HIV-positive and 23.6% (351/1487) in HIV-negative participants (p=0.57). Geometric mean CFA concentrations were 859 and 1660 antigen units per ml of blood (Ag/ml) respectively, geometric mean ratio (GMR) 0.85, 95%CI 0.49-1.50. In 7,863 adults in study two, CFA prevalence was 20.9% (86/411) in HIV-positive and 24.0% (1789/7452) in HIV–negative participants (p=0.15). Geometric mean CFA concentrations were 630 and 839 Ag/ml respectively (GMR 0.75, 95%CI 0.60-0.94). In the HIV-positive group, antiretroviral therapy (ART) use was associated with a lower CFA prevalence, 12.7% (18/142) vs. 25.3% (67/265), (OR 0.43, 95%CI 0.24-0.76). Prevalence of CFA decreased with duration of ART use, 15.2% 0-1 year (n=59), 13.6% >1-2 years (n=44), 10.0% >2-3 years (n=30) and 0% >3-4 years treatment (n=9), p<0.01 χ2 for linear trend.
In this large cross-sectional study of two distinct LF-exposed populations, there is no evidence that HIV infection has an impact on LF epidemiology that will interfere with LF control measures. A significant association of ART use with lower CFA prevalence merits further investigation to understand this apparent beneficial impact of ART.
| Lymphatic filariasis (LF) and HIV are both major public health problems worldwide and where they co-exist have the potential to interact. The main strategy for LF elimination is annual mass drug administration (MDA). A particular concern is whether HIV, through its impact on the immune system, will interfere with the effectiveness of this approach to control and eliminate LF. We report findings from cross-sectional studies in two separate populations in northern Malawi where both HIV and LF are common. One group (1,851 individuals) were studied at enrolment into a trial of anti-LF treatments, whilst the other study used samples stored from adult participants in a whole population HIV survey (7,863 individuals). Between 5–10% of the study participants were HIV-positive and 24% were LF-infected. We found no evidence that LF infection was more or less common in HIV-positive adults in either population. However, we identified robust evidence that antiretroviral therapy use was associated with lower LF prevalence rates. We have no evidence to suggest HIV will have a detrimental effect on LF control. On the contrary, the evidence suggests that antiretroviral therapy may have beneficial effects and merits further careful evaluation of the anti-filarial properties of these compounds.
| Human immunodeficiency virus (HIV) and parasitic infections affect widely overlapping populations in sub-Saharan Africa. Of the estimated 35 million people infected with HIV worldwide at the end of 2013, about 70% were from sub-Saharan Africa [1]. Parasitic infections, including lymphatic filariasis (LF) are also widespread in sub-Saharan Africa, raising the possibility of clinically significant interactions between the two pathogens. It has been suggested that HIV and parasitic co-infections may have bidirectional deleterious interactions by affecting susceptibility to HIV, impacting on HIV progression and potentially worsening clinical outcomes of filarial infection [2]. Previous in-vitro studies have shown helminth infections to increase susceptibility of peripheral blood mononuclear cells to HIV infection [3]. In addition deworming can result in increases in CD4+ cells and reduction in plasma HIV-1 RNA concentrations [4]. Derangements in the immune response associated with HIV-infection might also be expected to alter susceptibility to, or complications from, filarial infection or other helminths such as Strongyloides [5]. To date, there are few studies that have investigated LF and HIV co-infection and to our knowledge, none have been on a large population scale. A cross-sectional study of 907 adults undertaken in Tanga region of Tanzania reported increased circulating filarial antigen (CFA) concentration in HIV-positive persons [6], although a further evaluation of this group of individuals did not support any association between HIV and Wuchereria bancrofti infection [7]. Similarly, in urban southern India, no quantitative difference in W. bancrofti CFA levels by HIV status was found in a study of 432 HIV-positive and 99 HIV-negative patients [8].
Malawi embarked on a programme of mass drug administration (MDA) for LF control and elimination in 2009 [9]. Concerns that the programme may be less effective in areas of high HIV and LF prevalence prompted this study in Karonga, a district in the northern region of Malawi which was known to be highly endemic for LF infection [10]. Karonga is bordered by Lake Malawi to the east, the Songwe river to the north (which also forms the boundary with Tanzania), and by the Nyika plateau and escarpment to the west and south. The population is rural and dependent on subsistence agriculture including rice growing and fishing from the lake.
Two previous studies had been undertaken by co-authors in the district and both had serum samples stored with approval for later testing. The first was the recruitment phase of a randomised controlled clinical trial that investigated alternative schedules and dosing regimens of ivermectin and albendazole use in MDA programmes (study 1). Findings from this clinical trial are reported elsewhere [11]. The second study was nested within a comprehensive population-based HIV survey which enabled a longitudinal assessment of CFA in a whole adult population (study 2). In this paper we report the prevalence and relationship of LF and HIV infections from these two studies.
Study 1 used samples and data collected as part of the screening phase of a clinical trial of the effectiveness of increasing the dose and frequency of albendazole and ivermectin as antifilarial agents for clearing LF microfilaraemia (clinical trials registration number NCT01213576) [11]. It was undertaken between January 2009 to March 2012 in the northern portion of Karonga district along the Tanzanian border and Songwe river delta. Villages in this area had previously been shown to have a high prevalence of filarial antigenaemia and chronic manifestations of LF [10]. No mass treatment interventions had been undertaken in the area at the time the study was started.
Enrolment to the clinical trial required individuals to have a microfilarial count of >80 microfilariae per ml of blood. Consequently a population-based screening process was undertaken to identify suitable participants for enrolment in the therapeutic trial. The estimated total adult population of the target villages was 36,643 [12]. Sensitization meetings were held with community members, the Traditional Authority (TA) and all the village headmen and their aides who are administratively responsible for the study area. At these meetings the aims and procedures of the study were explained. Following verbal approval by the community leaders, a team of field workers went house-to-house seeking written consent and recruiting individual participants. All households in selected villages were visited in sequence before moving on to the next village. This screening phase was planned to continue until 120 eligible individuals with appropriate microfilarial levels were recruited into the trial. However, screening and recruitment into this study was discontinued following the rollout of the albendazole/ivermectin national MDA programme in the study area as further recruitment into the clinical trial became impractical. At the end of the recruitment phase individuals from 16 villages were included. For analysis purposes smaller villages were combined in a geographically appropriate way to produce 10 village location categories with suitable numbers of participants.
Individuals were eligible if they provided written informed consent, were residents of the area and aged between 18 and 55. At each home visit, the study was introduced and explained to all members and individual participants were asked to provide written informed consent by signing (or thumb printing if they were illiterate) the informed consent forms that were translated in the local language (Tumbuka). A questionnaire was administered to capture personal details. Eligible individuals were screened for CFA by the immunochromatographic (ICT) card test and for HIV by trained counsellors following the national HIV rapid testing algorithm. CFA positive individuals were asked to provide a night blood sample between 22:00 and 02:00 hours when a 5ml sodium citrate sample was collected for microfilaria counting and later stored in the project laboratory archive at -20°C. All individuals who were CFA positive but declined to participate in the clinical trial or did not meet the eligibility criteria for the trial were offered standard dose antifilarial therapy with albendazole and ivermectin. Individuals who were HIV positive were referred for HIV treatment and care. Assessment of HIV clinical stage, CD4 count and viral load were not performed on these individuals as a part of the study protocol.
Study 2 used samples and data from an annual whole adult population survey. Repeated rounds of data and sample collection spanned the time periods before and after the introduction of MDA. It was undertaken in the Karonga Health and Demographic Surveillance Site (KHDSS) and nested within a comprehensive population-based HIV sero-survey [13]. The KHDSS area is mapped and divided into geographically defined clusters of 20–30 households which are further aggregated into 21 geographically distinct reporting groups. The KHDSS was established between August 2002 and August 2004 to serve as a sampling frame for on-going epidemiological and clinical studies. Unlike study 1, data from study 2 included detailed socio-demographic information allowing for inclusion of these factors in statistical analysis. Since its establishment, the initial population of the KHDSS has been under continuous demographic surveillance. In addition, between September 2007 and October 2011 four annual HIV serological surveys have been conducted in all individuals aged 15 or more years using rapid point-of-care HIV tests on finger-prick whole blood samples [14]. Community sensitisation meetings to explain the aims and procedures of the study were held in each village and were followed by house-to-house visits by counsellors to recruit participants. The counsellors were trained and certified by Malawi Ministry of Health staff to perform HIV counselling and testing and referral using standard procedures [15]. Written informed consent was obtained as in study 1. In addition, all consenting adults were asked to provide a 5ml blood sample for quality control and storage for further laboratory analysis including for other diseases of importance in Karonga district. Plasma samples from consenting participants were stored in the project laboratory archive at -20°C. Samples from the first surveillance round, which took place between September 2007 and October 2008, prior to MDA introduction, were included in our study. Viral load, CD4 counts and clinical staging were not measured as a part of the survey.
The ICT card test (Binax, Portland, ME) [16] was only used as the screening test for LF infection in study 1. This is a portable rapid point-of-care test suitable for screening in non-laboratory settings and was used in accordance with manufacturer’s instructions. Microfilaria counting was done on ICT card test positive individuals in study 1 by the nucleopore membrane filtration technique [17]. This involves filtering a measured volume of venous blood through a 5μM pore size nucleopore filter. After filtration, the filter is removed, placed on a glass slide and mounted on a light microscope for examination and counting of microfilariae. Filarial antigenaemia was quantified in all plasma samples in both study 1 and study 2 by means of the Og4C3 antigen-capture ELISA (TropBio, Australia) [18]. Samples were analysed in accordance with the manufacturer’s instructions and expressed as antigen units per ml of blood (Ag/ml) based on control samples supplied with the ELISA kits. Samples with CFA concentration more than 128Ag/ml were considered clear positives for CFA. Samples with CFA concentration of 32Ag/ml and below were considered negatives. Samples with a titre between 32 and 128Ag/ml were considered indeterminate.
In both studies HIV testing used whole blood rapid diagnostic tests according to Malawian national guidelines [15] with demonstrated high accuracy in community settings in Karonga [14]. The initial screening test was with Determine TM HIV-1/2 (Abbott Japan Co Ltd, Japan) and confirmatory testing was done with UniGold TM HIV-1/2 (Trinity Biotech PLC, Ireland). Samples with a non-reactive screening test were considered negative and those with a reactive screening and confirmatory test were considered positive. Where the screening and confirmatory tests were discordant, a tie breaker using a third rapid test, (SD Bioline, Korea) was used.
Study forms were checked, coded, double entered and verified using Microsoft Access software. Statistical analysis was done in Stata 12 software (StataCorp, Texas, USA). Continuous variables were log transformed prior to analysis to achieve an approximate normal distribution. Linear regression was used for crude and adjusted analyses with results expressed as geometric mean ratios (GMR) and their 95% confidence intervals. The association of age, gender and CFA status with HIV positivity was estimated using χ2 tests for crude analyses and a logistic regression model for adjusted Odds Ratios (OR). In a risk factor analysis for CFA positivity, logistic regression was used to estimate crude and adjusted ORs. Variables were retained in the model if significant associations were identified in the unadjusted estimates. Geographic identifiers for groups of survey villages were also incorporated in the models to adjust for geographic confounding as previous surveys had indicated heterogeneity of CFA prevalence across the region [19]. Rather than a binary variable, HIV was treated as a categorical variable in the model with the HIV—negative group and two HIV-positive groups based on ART use at the time of blood sampling. A sub-group analysis to investigate the effect of cotrimoxazole and duration of ART use on CFA prevalence was performed on the HIV-positive group only. Logistic regression models were used to estimate odds ratios. Difference in CFA prevalence with increased use of ART was investigated with a χ2 test for linear trend with odds ratios derived from a 2xn table. Adjusted odds ratios were estimated with logistic regression.
The National Health Sciences Research Committee of the Malawi Ministry of Health (protocol numbers 495 and 419) and the Ethical Committee of the London School of Hygiene and Tropical Medicine (protocol numbers 5344 and 5081) gave ethical clearance for both studies 1 and 2. Study participants in both studies were consented for storage and later testing of samples at the time of enrolment. This covered testing for HIV and other diseases of local significance. The National Health Sciences Research Committee (protocol number 908) and the Ethical Committee of the Liverpool School of Tropical Medicine (protocol number 11.77) approved the additional analysis conducted on stored samples.
The overall population of Karonga district during the study period was 272,789 [12] with the KHDSS population accounting for 33,500. HIV prevalence in the KHDSS in the 2007–2008 survey year was measured at 7.4% but estimated to be 10.4% when adjusted for non-testing by those who already knew they were HIV-positive [20]. At baseline 54.8% of those aged 15 years or more reported previous HIV testing.
From the estimated total of 36,643 adults of the target villages, 1,851 individuals were eligible and consented to participate. Of these 1,851 individuals screened for LF antigen by the ICT card test, 447 (24.2%) were CFA positive (Fig 1). A total of 1,656 individuals accepted HIV testing and 169 (10.2%) of these were HIV-positive. HIV-positive individuals tended to be older (Table 1). CFA positivity was present in 43 (25.6%) of HIV-positive and 351 (23.6%) of HIV-negative (crude OR 1.11, 95% CI 0.77–1.60) with an LF/HIV co-infection prevalence rate of 2.6%. CFA positivity did not differ by HIV infection status (Table 1). There was heterogeneity in the prevalence of CFA by village location, median 24.4%, range 15.7–33.3% (Pearson χ2 22.7, p<0.01, 9 degrees of freedom). Data on the use of antiretroviral and cotrimoxazole was incomplete in the context of this study.
Microfilaria counting was done in the 311 (69.6%) LF antigen positive individuals who were eligible and gave consent for night blood sampling. The remainder either refused or left before follow up. HIV prevalence in those lost to follow up was broadly similar to those sampled (10.3% vs. 9.3% respectively, χ2 test p = 0.90). Microfilariae were present in 49.5% of the 311 sampled individuals. Microfilarial detection and levels did not differ by HIV infection status (Fig 1 and Table 1).
Of 311 stored baseline night blood plasma samples, 290 (93.2%) were CFA positive using the Og4C3 antigen-capture ELISA. CFA was positive in 26 (89.7%) of HIV-positive individuals, 231 (93.9%) of the HIV-negative individuals and 33 (97.1%) of HIV-unknown individuals respectively (p = 0.47). The geometric mean CFA concentration levels by HIV status were 859 and 1660 for HIV-positive and HIV-negative respectively (GMR 0.85, 95% CI 0.49–1.50). CFA and MF counts showed reasonable positive correlation (Pearson correlation coefficient r = 0.56, p<0.01).
The total eligible population of 15 year olds and older in the KDHSS at the baseline survey was 11,756. From this group, 7,863 (66.9%) underwent HIV testing and consented to storage of their blood sample. Of these, 1,875 (23.9%) individuals were CFA positive by the Og4C3 ELISA. HIV infection was identified in 411 (5.2%) participants. HIV-positive adults tended to be older and more likely to be female (Table 2). CFA positivity was present in 86 (20.9%) of HIV-positive and 1789 (24.0%) of HIV-negative (crude OR 0.84, 95% CI 0.66–1.07) with an HIV/LF co-infection prevalence rate of 4.6%. In the female participants, CFA positivity was present in 17.8% of the HIV-positive and 19.8% of the HIV-negative (OR 0.88, 95% CI 0.64–1.21) and in the male participants, 26.8% of the HIV-positive and 29.4% of the HIV-negative (OR 0.88, 95% CI 0.60–1.28) respectively. Geometric mean CFA concentration was lower in the HIV-positive individuals by 25% although this association was weakened when adjusted for age and sex (Table 2).
Several risk factors were associated with an increased prevalence of CFA (Table 3). These included male gender, age between 30 and 39 and lower quality housing, whilst decreased CFA prevalence was associated with higher levels of education (p<0.01, χ2 for linear trend) the availability of piped tap water or the use of lake water and the use of antiretrovirals. Bed net ownership was high, however ownership or the number of nets owned in the household was not associated with CFA prevalence. Individuals were found in all 21 reporting groups, with a median of 314 participants (range 129–820). There was considerable heterogeneity in the prevalence of CFA by reporting group, median 23.2% range 5.7–37.2% (Pearson χ2 354.7, p<0.01, 20 degrees of freedom).
Of the 411 HIV-positive adults, 142 (34.5%) were taking antiretroviral therapy (ART) and 117 (28.5%) were using cotrimoxazole prophylaxis (CTX) with only 4 of the 117 taking CTX without ART at the time of sampling. In 6 of the 411 individuals, information on ART and/or CTX use at the time of sampling was unavailable. ART consisted of Lamivudine, Stavudine and Nevirapine (Triomune-30) in 94% of cases with Zidovudine or Efavirenz substitutions in the remainder. No protease inhibitors were in use. In the HIV-positive group, ART use was associated with a lower prevalence of CFA when compared to those not on ART [12.7% vs. 25.3% (OR 0.43, 95% CI 0.24–0.76)]. Similarly, CTX use was associated with lower CFA prevalence [12.8% vs. 24.1% (OR 0.46, 95% CI 0.25–0.85)]. In a multivariable model incorporating ART and CTX use along with age, sex and geographical location, the adjusted odds ratio for ART use was 0.47 (95% CI 0.17–1.31) and for CTX use 0.92 (95% CI 0.31–2.71). When the ART treated group were further sub-divided by year since treatment started, there was a significant trend to decreased prevalence of CFA with increasing time on treatment; 25.3% no treatment (n = 265), 15.2% year 1 treatment (n = 59), 13.6% year 2 treatment (n = 44), 10.0% year 3 treatment (n = 30) and 0% year 4 treatment (n = 9), (p<0.01 χ2 for linear trend). This relationship persisted after adjustment for age, gender and reporting group. In the HIV-positive individuals with detectable CFA, the geometric mean concentration of CFA was not significantly different between those off and on ART, 647 vs. 512 Ag/ml respectively, GMR 1.27, 95% CI 0.76–2.08 (Fig 2), nor did the GMC differ by ART duration category 647, 392, 762, 516 & 0 Ag/ml for no treatment, year 1, 2, 3 & 4 of treatment respectively.
We present data from two separate studies undertaken in Karonga district, northern Malawi. In both studies a high LF and HIV prevalence was measured with HIV co-infection rates of 2.6% and 4.6% among those who were CFA positive and 15 years and older. We found no evidence that HIV is associated with an increased risk of LF infection. Initial findings from study 1, a clinical trial not powered to definitively test the impact of HIV on LF infection, revealed a tendency to lower CFA and microfilarial density in the HIV-positive adults. Subsequent investigation of these parameters in the much larger population sample revealed a tendency to lower CFA prevalence in the HIV-positive group, attributable to significantly lower CFA prevalence in the ART treated sub-group, a finding that persisted following adjustment for key potential confounders and showed a significant trend to lower CFA prevalence with duration of ART use. There was no significant effect of CTX therapy, when analysed in a multivariable model. CFA concentration was also persistently lower in the HIV-positive group although at a level of uncertain clinical or public health significance.
Previous studies have reported divergent findings with some showing an association between LF and HIV infections but these have tended to be small samples and in selected populations [6–8]. In contrast to these studies, our second study had a larger sample taken from a whole population survey, including a high proportion of the at-risk population in an area with high prevalence rates of LF and HIV. The findings in relation to ART use are novel and we are unaware of other studies that have investigated this association. Individuals receiving ART may represent a select group of the HIV-positive population who have better health seeking behaviour, may be more educated, live in better accommodation and/or may live in close proximity to health providers. However, as this work was undertaken in the context of a demographic survey we were able to investigate these potential confounders by adjusting for reported educational status, housing quality, access to clean water and geographic location. The finding of ART associated with a lower CFA prevalence appears robust.
An explanation for these findings remains less clear and merits further work. Residual confounding or an unrecognised selection bias remains possible, but seems unlikely given the highly significant lower CFA prevalence with duration of ART therapy. The crude association of CTX with lower CFA prevalence seems adequately explained by concomitant use of ART, and there is no evidence to support either sulphonamides or trimethoprim, the components of CTX, as effective antifilarial agents. If LF infection adversely impacts on the success of ART therapy, then over time the prevalence of CFA positivity in this group will reduce as the LF/HIV co-infected die. There is no evidence from the Malawi national HIV programme that outcomes from ART treatment are worse in regions of the country endemic for LF compared to those with low LF prevalence. Helminth infections have been linked to increased viral load in non-ART treated individuals [21] but not to evidence of faster HIV progression [22]. Similarly, LF infection had no significant effect on HIV disease progression in a study of W. bancrofti and HIV coinfections in south India [23]. Altered diagnostic accuracy of the Og4C3 ELISA in the presence of ART has not been reported. ART has been rarely linked to false negative HIV results in children and adults but this is more likely to be due to low levels of virus and/or antibody than a direct inhibitory effect. The reduction in CFA prevalence by ART treatment duration and the antigen capture nature of the Og4C3 ELISA would be difficult to explain by ART inhibition of the assay. Immune reconstitution as a result of ART does not adequately explain our finding either as there is a similar prevalence of CFA in the HIV-negative and the HIV-positive untreated. There is no precedent for immune recovery following ART leaving the immune system in a more competent state than an HIV-negative person. ART treatment is an imprecise proxy marker of duration of HIV infection. If the natural history of LF in the HIV-positive is a steady fall in antigenaemia could this explain the association? We do not have accurate seroconversion dates for the majority of this population so are not able to fully consider this possibility. However with ART use the “natural history” of HIV is dramatically altered and it might be expected that any tendency to lower antigenaemia with time would also be altered and this would be inconsistent with our findings. The most plausible explanation for this finding is a direct filaricidal activity of the major ART agents. We are unaware of any data on the effect of Lamivudine, Stavudine or Nevirapine on helminths. Further evaluation of these molecules as antihelminthics would be appropriate.
Of the other factors associated with CFA positivity all have been reported previously, providing reassurance that the epidemiology of LF disease in Karonga is not unique and results are generalizable to other similar regions. One surprise was the lack of association with bed net ownership. However most households possessed bed nets limiting the power of any comparison, and during this survey we did not specifically ask about usage, or condition of the nets, thus limiting the value of this finding. More detailed evaluation of this will be needed in future work.
Both of our studies had some degree of selection bias, but it is unlikely that this has fundamentally altered our findings. In study 1, we targeted villages known historically to have a high prevalence of LF infection. If participation by HIV-positive individuals was reduced because of perceived stigma associated with an HIV test, we may have had reduced power to identify an association between LF and HIV. However, a similar finding in the much larger study 2 provides consistency. In study 2, we know HIV-positive adults were under-represented. Adults who knew their status from earlier HIV testing studies or through routine service provision in the district, declined participation [20]. However it is difficult to see a mechanism whereby LF co-infection would disproportionately lead to non-participation by HIV-positive adults and in particular ART treated HIV-positive adults thereby obscuring the true association.
More females than males were included in both studies. This may represent the easier access to females at the time of recruitment since females are more likely to be at home. Although we know men are more likely to be infected with LF in this population, we do not think this under-representation has meaningfully affected the LF/HIV association. Sub-group analyses showed similar odds ratios for the LF/HIV association by gender in study 2 suggesting no major effect modification.
The measurement of our exposure (HIV) and outcome endpoints (LF status) were based on accurate and well described tests and we do not believe these have introduced significant bias into the study. We used different tests for assessment of circulating filarial antigen in the two studies with different sensitivities and specificities, the ICT card test with sensitivity and specificity reportedly close to 100% and the Og4C3 ELISA test with 100% sensitivity and specificity of at least 94% [24–26]. There was some disparity between these two tests identified in study 1. This is consistent with previous studies that have reported overall agreement between the ICT and Og3C4 tests but different sensitivities and specificities [24,25]. In study 2, we were not able to assess MF counts due to the use of a stored sample collection. Whilst we cannot categorically rule out an association between MF density and HIV, data from study 1 showed a positive correlation between CFA levels and MF density. Previous studies have also shown a positive correlation between CFA levels and MF density [26,27]. This implies that the CFA relationship will broadly apply to MF counts.
In summary, we did not demonstrate a significant detrimental association between LF and HIV in these studies that will have a negative impact on plans to eliminate lymphatic filariasis. However ART treated adults had significantly lower CFA prevalence, a finding that merits further careful evaluation to exclude an adverse impact of LF on HIV, or the potential of antiretrovirals as molecules with antihelminthic properties.
|
10.1371/journal.pbio.1001945 | De-Differentiation Confers Multidrug Resistance Via Noncanonical PERK-Nrf2 Signaling | Malignant carcinomas that recur following therapy are typically de-differentiated and multidrug resistant (MDR). De-differentiated cancer cells acquire MDR by up-regulating reactive oxygen species (ROS)–scavenging enzymes and drug efflux pumps, but how these genes are up-regulated in response to de-differentiation is not known. Here, we examine this question by using global transcriptional profiling to identify ROS-induced genes that are already up-regulated in de-differentiated cells, even in the absence of oxidative damage. Using this approach, we found that the Nrf2 transcription factor, which is the master regulator of cellular responses to oxidative stress, is preactivated in de-differentiated cells. In de-differentiated cells, Nrf2 is not activated by oxidation but rather through a noncanonical mechanism involving its phosphorylation by the ER membrane kinase PERK. In contrast, differentiated cells require oxidative damage to activate Nrf2. Constitutive PERK-Nrf2 signaling protects de-differentiated cells from chemotherapy by reducing ROS levels and increasing drug efflux. These findings are validated in therapy-resistant basal breast cancer cell lines and animal models, where inhibition of the PERK-Nrf2 signaling axis reversed the MDR of de-differentiated cancer cells. Additionally, analysis of patient tumor datasets showed that a PERK pathway signature correlates strongly with chemotherapy resistance, tumor grade, and overall survival. Collectively, these results indicate that de-differentiated cells up-regulate MDR genes via PERK-Nrf2 signaling and suggest that targeting this pathway could sensitize drug-resistant cells to chemotherapy.
| The development of multidrug resistance is the primary obstacle to treating cancers. High-grade tumors that are less differentiated typically respond poorly to therapy and carry a much worse prognosis than well-differentiated low-grade tumors. Therapy-resistant cancer cells often overexpress antioxidants or efflux proteins that pump drugs out of the cell, but how the differentiation state of cancer cells influences these resistance mechanisms is not well understood. Here we used genome-scale approaches and found that the PERK kinase and its downstream target, Nrf2—a master transcriptional regulator of the cellular antioxidant response—are key mediators of therapy resistance in poorly differentiated breast cancer cells. We show that Nrf2 is activated when cancer cells de-differentiate and that this activation requires PERK. We further show that blocking PERK-Nrf2 signaling with a small-molecule inhibitor sensitizes drug-resistant cancer cells to chemotherapy. Our results identify a novel role for PERK-Nrf2 signaling in multidrug resistance and suggest that targeting this pathway could improve the responsiveness of otherwise resistant tumors to chemotherapy.
| Multidrug resistance (MDR) is the primary obstacle to treating malignant tumors [1]. Cancer cells develop MDR by overexpressing antioxidant enzymes that neutralize the reactive oxygen species (ROS) required for chemotherapy toxicity or by up-regulating drug efflux pumps [2],[3]. In many cancers, these MDR mechanisms are up-regulated by mutation or amplification of genes encoding antioxidant enzymes or drug efflux pumps. Many other cancers, however, up-regulate these genes through nonmutational mechanisms that remain poorly understood.
One nonmutational mechanism by which cancer cells acquire MDR is de-differentiation. De-differentiation is a well-established marker of poor prognosis tumors and can occur when differentiated cells are induced into a more primitive stem-cell–like state [4]–[6]. One mechanism by which both cancerous and noncancerous cells can be de-differentiated is through induction of an epithelial-to-mesenchymal transition (EMT) [7]–[14]. De-differentiated cancer cells generated by EMT and cancer stem-like cells are both resistant to a wide range of chemotherapies [15]–[19]. Conversely, cells experimentally induced to differentiate are more sensitive to chemotherapies [20]–[23]. Although de-differentiation is known to up-regulate MDR mechanisms as described above, how this occurs is poorly understood.
In this article, we examine this question by employing a global transcriptional profiling approach to identify ROS-induced genes that are preactivated in de-differentiated cells. Many of these genes—which are activated in de-differentiated cells even in the absence of oxidative damage—are regulated by a single signaling pathway. We further show that this pathway is critical for de-differentiated cells to resist chemotherapies.
To study the effects of differentiation state on MDR, we used isogenic pairs of human breast epithelial cells (HMLE) that were either differentiated and expressed a control vector, or de-differentiated through induction of an EMT—achieved by expressing the Twist transcription factor [24],[25]. These de-differentiated HMLE-Twist cells were more resistant to the chemotherapy drugs Paclitaxel (Tax) and Doxorubicin (Dox) than differentiated HMLE-shGFP cells, consistent with prior reports (1.5× and 2.5×, respectively; Figure 1a) [26],[27]. To determine how Twist-induced de-differentiation caused MDR, we assessed whether known mechanisms were up-regulated in these cells. Twist overexpression significantly increased efflux pump activity (Figure 1b) and lowered ROS levels—both basal and induced by the oxidizer menadione or Dox (Figure 1c,d) [28]. Additionally, HMLE-Twist cells displayed significantly lower amounts of lipid peroxidation compared to HMLE-shGFP cells (Figure 1e). As a measure of overall reducing capacity of the cells, we also show that HMLE-Twist cells had a greater pool of reduced glutathione, which could be maintained even in the presence of menadione (Figure 1f). Finally, Twist overexpression led to a significant increase in expression of enzymes involved in ROS metabolism: superoxide dismutase 1 (SOD1) and catalase (CAT) (Figure 1g).
We suspected that these MDR mechanisms were up-regulated through a normal regulator of the cellular antioxidant response. To identify putative regulators, we transcriptionally profiled HMLE-shGFP and HMLE-Twist cells treated with vehicle or menadione (Table S1). In the absence of oxidative stress, 1,694 genes were differentially expressed between the two cell types, several of which were ROS and efflux-related genes (Tables S2 and S3). Treatment with menadione induced the expression of 181 and 170 genes in HMLE-shGFP and HMLE-Twist cells, respectively, with 44 genes being commonly induced in both cell types (Table S4; hypergeometric test, p value<1.0×10−10). Of the 181 genes induced by menadione in HMLE-shGFP cells, 54 were already up-regulated in HMLE-Twist cells in the absence of treatment (Table S5; hypergeometric test, p value<1.0×10−10, Figure 1h). Of these 54 genes, 38 were uniquely induced in HMLE-shGFP but not HMLE-Twist cells treated with menadione. This suggests that some oxidative stress response genes are “preactivated” in de-differentiated HMLE-Twist cells.
The most significantly preactivated gene in HMLE-Twist cells was heme oxygenase 1 (HMOX-1)—expressed at 8-fold higher levels in HMLE-Twist cells compared to HMLE-shGFP cells and induced 22-fold in differentiated cells treated with menadione. HMOX-1 is a well-characterized enzyme involved in the metabolism of heme, but is also a major target of master antioxidant regulator Nrf2 [29]–[31]. The Nrf2 transcription factor activates an arsenal of antioxidant genes and ABC transporters, and its up-regulation is associated with acquired MDR [32]–[35]. To test whether Nrf2 might be basally active in HMLE-Twist cells, but not HMLE-shGFP cells, we examined Nrf2 target gene expression. Of 1,013 Nrf2 direct-target genes, a significant number—142 genes—were up-regulated in HMLE-Twist cells compared to HMLE-shGFP cells in the absence of oxidative stress (Table S6; hypergeometric test, p value<1.0×10−10) [36]. Further, 7 of the 54 oxidative stress response genes “preactivated” in HMLE-Twist cells were Nrf2 direct-target genes, representing a significant enrichment over the number predicted by random chance (Table S5; hypergeometric test, p value = 4.9×10−5, Figure 1h). To confirm Nrf2 activation in HMLE-Twist cells, we assessed its subcellular localization by immunofluorescence. In HMLE-shGFP cells, Nrf2 was sequestered in the cytoplasm and translocated to the nucleus when cells were treated with menadione (Figure 1i). In HMLE-Twist cells, however, Nrf2 was constitutively in the nucleus, and treatment with menadione only modestly increased its nuclear accumulation (Figure 1i). These findings demonstrate that Nrf2 is constitutively active in de-differentiated HMLE-Twist cells—even in the absence of exogenous stress.
We next examined why Nrf2 was constitutively active in HMLE-Twist cells, even though basal ROS levels are low. Although ROS activate Nrf2 by oxidation, it can also be activated in the absence of oxidative stress by several kinases [37]–[39]. In particular, Nrf2 is directly phosphorylated and activated by the ER-membrane kinase PERK, which is canonically activated under conditions of ER stress as part of the unfolded protein response (UPR) [40]–[42]. In this context, PERK relieves ER stress by slowing protein translation through phosphorylation of eiF2α. We have recently shown that PERK is also activated upon EMT-induced de-differentiation—even in the absence of overt ER stress [43]. Consistent with this, we found that PERK is constitutively phosphorylated in HMLE-Twist cells, but not in HMLE-shGFP cells, and inhibition of PERK with a small-molecule inhibitor blocked its phosphorylation (Figure 2a) [44]. To understand if PERK controls constitutive Nrf2 activation in HMLE-Twist cells, we assessed Nrf2 localization following PERK inhibition. We found that inhibition of PERK fully reversed the nuclear localization of Nrf2 in HMLE-Twist cells, but did not prevent oxidative stress-induced nuclear accumulation of Nrf2 in either HMLE-shGFP or HMLE-Twist cells (Figure 2b). As a complementary approach to PERK inhibition and to rule out off-target effects of the small-molecule PERK inhibitor, we also generated cell lines in which PERK expression was stably inhibited by two different shRNAs (Figure 2c). Inhibition of PERK by shRNA significantly decreased Nrf2 nuclear localization in HMLE-Twist cells, mirroring the results obtained with the small-molecule PERK inhibitor (Figure 2d). Collectively, these results demonstrate that Nrf2 nuclear localization is controlled by PERK in de-differentiated HMLE-Twist cells.
To confirm that Nrf2 nuclear localization correlated with its activation, we assessed Nrf2 target gene expression following PERK inhibition. We found that PERK inhibition significantly decreased HMOX-1 expression in HMLE-Twist cells, but did not prevent induction of HMOX-1 in response to oxidative stress (Figure 2e). Moreover, using microarray gene expression analyses, we found that PERK inhibition decreased the expression of 58 of the 142 Nrf2-target genes (41%) activated in HMLE-Twist cells (Table S6). Amongst these PERK-Nrf2-target genes were ABC transporters, enzymes involved in glutathione metabolism and ROS buffering, and several proteins with known roles in drug resistance. These findings confirm that the exit of Nrf2 from the nucleus correlates with down-regulation of its target genes.
PERK has previously been shown to bind to, directly phosphorylate, and activate Nrf2, though the exact phosphorylation sites have not yet been determined [42]. To show that PERK directly regulates Nrf2 in our system, we performed PERK immunoprecipitation followed by western blot with a Nrf2-specific antibody—which confirmed that PERK and Nrf2 directly interact in HMLE-Twist cells (Figure 2f). We also immunoprecipitated Nrf2 in either the presence or absence of the PERK inhibitor, which demonstrated that Nrf2 phosphorylation was markedly reduced by PERK inhibition (Figure 2g). These data, combined with our finding that inhibiting PERK decreases nuclear accumulation of Nrf2, suggest that PERK directly interacts with Nrf2 to mediate its nuclear translocation and activation.
We next tested whether inhibition of PERK would eliminate MDR phenotypes associated with HMLE-Twist cells. PERK inhibition caused a 45% increase in mitochondrial ROS levels in HMLE-Twist cells, but did not affect HMLE-shGFP cells (Figure 3a). PERK inhibition also significantly increased lipid peroxidation in HMLE-Twist cells, but not in HMLE-shGFP cells (Figure 3b). PERK inhibition compromised ROS buffering—cells pretreated with the PERK inhibitor produced 25%–55% more ROS than vehicle-treated cells (Figure 3c). Additionally, PERK inhibition led to a significant decrease in the expression of ROS metabolizing enzymes SOD1 and CAT (Figure 3d). Lastly, inhibition of PERK signaling reduced the percentage of high-effluxing HMLE-Twist cells by 50% and did not affect efflux in HMLE-shGFP cells (Figure 3e). Together these results demonstrate that a simple change in differentiation state confers MDR phenotypes, and these are mediated by constitutive PERK signaling.
To understand how this applies in the context of cancer, we expanded our analyses to include several luminal and basal-like breast cancer cell lines, which represent epithelial-like/differentiated and mesenchymal-like/de-differentiated cells, respectively [45]. Previous work has shown that PERK is preferentially activated in basal compared to luminal cell lines [43]. Consistent with our results in the HMLE system, basal breast cancer cells had lower overall ROS than luminal cells, and addition of the PERK inhibitor caused a dramatic increase in ROS levels in basal cells but not luminal cells (Figure 3f). Likewise, inhibition of PERK caused a 25% reduction in the ratio of reduced to oxidized glutathione in only the basal cell lines, indicative of decreased ROS buffering (Figure 3g). This indicates that PERK contributes to the enhanced oxidative stress buffering ability of both noncancerous and cancerous de-differentiated cells.
In order to affect chemotherapy resistance, we rationalized that PERK inhibition would need to occur prior to chemotherapy exposure to allow time for reversal of MDR phenotypes (Figure 4a). Pretreatment with the PERK inhibitor greatly sensitized both HMLE-Twist and HMLE-shGFP cells to subsequent treatment with Tax and Dox—the number of surviving cells was reduced significantly in both cell types (Figure 4b). Treatment with a ROS-scavenging agent n-acetyl cysteine (NAC) was able to rescue this decreased survival, indicating that PERK pathway activation contributes to chemotherapy resistance in significant part via ROS buffering (Figure 4c,d) [46]. We also utilized a small molecule—oltipraz—capable of inducing Nrf2 activation (Figure 4e,f) [47]. Activation of Nrf2 significantly rescued PERK-dependent decreases in cell survival. To rule out the possibility that off-target effects of oltipraz were responsible for this effect, we performed the same rescue experiment in cells with stable Nrf2 knockdown achieved by two independent shRNAs. When Nrf2 was inhibited, oltipraz was no longer able to rescue the effects of PERK inhibition, confirming that these effects were mediated by Nrf2 (Figure 4g–i). These results indicate that PERK signaling through Nrf2 is responsible for the acquisition of MDR.
These results prompted us to test the effect of PERK inhibition in vivo, utilizing xenografted tumors derived from therapy-resistant basal breast cancer cells. We utilized a treatment plan involving cycles of pretreatment with the PERK inhibitor, followed immediately by treatment with Dox. The combined treatment resulted in significantly smaller tumors compared to single or mock treatments (Figure 5a). To test if PERK inhibition affected ROS buffering in vivo, we harvested tumors from each of the four treatment groups and measured the expression of the ROS-metabolizing enzyme SOD1. We found that the Dox, PERK inhibitor, and combined treatment groups all had significantly reduced expression of SOD1 compared to control tumors, with the dual-treated tumors having the lowest expression (Figure 5b). Additionally, the combined treatment group had the most necrotic cells compared to the other treatment groups (Figure 5c).
We next adjusted the dosage schedule to highlight the synergistic interactions between PERK inhibition and Dox treatment and found that reducing the total dosage and frequency of treatments further emphasized the sensitization effect—dual-treated tumors were 4 times smaller than Dox-treated tumors and >5 times smaller than PERK inhibitor or mock-treated groups (Figure 5d). Although prior research has shown that PERK is critical for tumor growth and angiogenesis [48]–[50], we found that low-dose inhibition only minimally impacted tumor growth in the absence of chemotherapy. To assess the in vivo effects on ROS buffering, we measured the levels of reduced glutathione (GSH) in tumors harvested from each treatment group. Dox, PERK inhibitor, and combined treatment groups all had decreased levels of GSH compared to the control group, with the dual-treated tumors having the lowest amount (Figure 5e).
As an important control to demonstrate that the observed in vivo results were not due to off-target effects of the PERK inhibitor, we utilized xenografted tumors derived from luminal breast cancer cells. Although treatment with Dox led to a reduction in tumor size, inhibition of PERK did not provide any additive benefit in the luminal tumors (Figure 5f). This confirms that the effects observed in the basal breast cancer xenografts are not due to off-target effects of the PERK inhibitor, as luminal cells—unlike basal cells—do not constitutively activate PERK and do not significantly respond to PERK inhibition. Together our results suggest that combining Dox treatment with PERK inhibition compromises the ROS-buffering capacity of basal-like breast cancer cells and sensitizes them to chemotherapy-induced cell death.
To assess the clinical relevance of our findings, we analyzed primary human breast tumor datasets. Utilizing two independent datasets (comprised of 413 patient tumors), we first tested for correlations between the expression of PERK pathway genes and genes associated with the basal subtype of breast cancer. We found that a PERK gene expression signature correlated positively with a basal breast cancer gene signature, suggesting that the PERK signaling pathway is active in basal breast tumors (Figure 5g) [51]. As a negative control, an IRE1 gene expression signature did not show a significant correlation (Figure 5g). Additionally, we found that PERK pathway activity could stratify patient response to therapy—85% of PERK-low tumors displayed complete or partial response to therapy, compared to only 38% of PERK-high tumors (Figure 5h). Finally, PERK pathway expression also correlated to differentiation state and overall survival in invasive high-grade glioma—tumors stratified into a PERK-high group were almost exclusively poorly differentiated grade 4 GBM and had significantly worse overall survival than the PERK-low group (Figure 5i,j). These results highlight the relevance of our work in primary tumors, and suggest that targeting PERK signaling may be beneficial in highly aggressive and malignant tumor types.
These findings identify PERK-Nrf2 signaling as one mechanism by which de-differentiated cells gain MDR. Because they constitutively activate Nrf2, these de-differentiated cells constitutively express antioxidant enzymes and drug efflux pumps. Remarkably, in this setting, Nrf2 is not activated by oxidation, but rather through a previously reported mechanism involving its phosphorylation by PERK [42]. This finding is of particular interest given Nrf2's known role in promoting chemotherapy survival [34] and its constitutive activation by mutation in a subset of tumors [52]–[54]. Our findings indicate that a change in cellular state, in the absence of mutation or oxidative stress, can also lead to constitutive Nrf2 activation. This enables de-differentiated cells to survive chemotherapy by preventing cellular damage before it occurs. In contrast, differentiated cells activate Nrf2 only after proteins and DNA have been oxidized. Although this defensive response may succeed in neutralizing toxins, the damage to cellular components would have already occurred.
Our findings also highlight the importance of stress signaling in cancer. Cancer cells activate stress response pathways to protect themselves from harsh environments encountered during tumor growth and metastasis—for example, hypoxia and nutrient deprivation—and also during the course of chemotherapy. We show that de-differentiated tumor cells preactivate PERK-Nrf2 signaling in the absence of stress and that inhibition of PERK sensitizes these cells to chemotherapy. These observations complement prior studies establishing a role for the UPR and its downstream targets in chemosensitization [55]–[67]. Collectively, our findings provide mechanistic insights into how cellular de-differentiation promotes MDR and suggest that inhibiting PERK-Nrf2 signaling may reverse the MDR of cancer cells that are otherwise drug resistant.
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 Animal Care and Use Committee of the Massachusetts Institute of Technology (Protocol No. 0611-071-14). All surgery was performed under isoflurane anesthesia, and every effort was made to minimize suffering.
HMLE-shGFP and HMLE-Twist cell lines were a kind gift of Dr. Robert Weinberg and cultured as described previously [8]. Basal breast cancer (MDA.MB.231, Hs578t) and luminal breast cancer (MCF7, T47D, MDA.MB.361) cell lines were purchased from ATCC and cultured in DMEM+10% FBS. The SUM159 basal breast cancer cell line was purchased from Asterand and cultured in Ham's F-12+5% FBS, Insulin and Hydrocortisone. Chemical oxidizer menadione, Nrf2 activator oltipraz, and ER-stress inducer thapsigargin were purchased from Sigma-Aldrich. Cumene hydroperoxide (CH) was purchased from Life Technologies. The PERK inhibitor (PERKi) was described previously and purchased from EMD Millipore [44]. Lentiviral short hairpin RNA (shRNA) constructs were generated as described previously [68]. Lentiviral integration was selected with 1 µg/ml puromycin or 10 µg/ml blasticidin for 7 d, and knockdown efficiency was measured by quantitative RT-PCR.
ROS production was measured by fluorescent imaging or flow cytometry analysis of MitoSOX or CellROX probes (Life Technologies) according to manufacturer instructions. Lipid peroxidation was assessed using the Click-iT Lipid Peroxidation Imaging Kit (Life Technologies) according to the manufacturer instructions. Total, reduced, and oxidized glutathione were determined using the GSH/GSSG-Glo™ Assay (Promega) according to manufacturer instructions.
MDR1 efflux ability was measured by flow cytometry quantification of DiOC2(3)-dye efflux (EMD Millipore). Efflux assays were conducted according to manufacturer instructions. Briefly, cells were loaded with DiOC2(3)-dye for 10 min and either kept on ice or placed at 37°C for 1.5 h to allow efflux of the dye. Control and efflux samples were then immediately analyzed by flow cytometry.
Cells were lysed with cold RIPA buffer plus complete protease inhibitor cocktail (Roche Applied Science). The signal was detected using the SuperSignal ECL system (Thermo Scientific). The following antibodies were used for immunoblotting: SOD1 and Nrf2 (Santa Cruz Biotechnology) and CAT, HMOX-1, pan-phospho, and β-actin (Cell Signaling Technologies).
HMLE-Twist cells grown in the presence or absence of 1 µM PERK inhibitor for 48 h were lysed in nondenaturing lysis buffer (20 mM Tris pH 7.5, 150 mM NaCl, 2 mM EDTA, 1% NP-40, supplemented with cocktails for phosphatase and protease inhibition). Equal protein amounts were used for immunoprecipitation using PERK or Nrf2 antibody as per the vendors' instructions. Samples were analyzed by immunoblotting using antibodies to Nrf2 and phospho-Ser/Thr–containing proteins.
Anti-Nrf2 (C-2) and anti–phospho-PERK (pPERK) antibodies was purchased from Santa Cruz Biotechnology. Cells were fixed on glass chamber slides in 4% PFA for 5 min, blocked with 5% BSA in PBS, and incubated with primary antibody at a 1∶50 dilution for 2 h. Slides were then washed with PBS and incubated with an Alexa Fluor 488 anti-rabbit secondary antibody. The nuclei were then stained with DAPI prior to analysis.
Immediately after harvest, tumors were fixed in 4% PFA for 24 h and paraffin-embedded. For staining, slides were deparaffinized in xylene and then rehydrated with ethanol and double distilled water. Hydrogen peroxide was used to block nonspecific sites, and Diva Decloaker (BioCare Medical) solution and microwaving were used for antigen retrieval. Sections were incubated for 2 h at room temperature with a SOD1 antibody (Santa Cruz Biotechnology). Expression was detected using HRP anti-rabbit secondary antibody (BioCare Medical) and betazoid DAB (BioCare Medical). The slides were counterstained with hematoxylin.
HMLE-shGFP and HMLE-Twist cells were treated with 1 µM of PERKi or DMSO for 48 h and then treated with 40 µM menadione or DMSO for 2 h immediately before RNA extraction. Total RNA were extracted using Qiagen RNeasy kit, and integrity and quality verified prior to analysis. Gene expression analyses were conducted using Affymetrix GeneChip Human Genome U133 Plus 2.0 Arrays according to standard Affymetrix protocols, with normalization as described previously [69]. Alteration of gene expression by PERKi and/or menadione was calculated by comparing the expression of each gene across treatment groups for each cell type. The gene-expression data have been deposited in the NCBI Gene Expression Omnibus public database (GEO; GSE59780).
Cells were seeded and treated according to the schedule described in Figure 4a. Briefly, cells were seeded on day 0, pretreated with PERKi (1 µM) or DMSO for 48 h on days 1–3, and rescued with NAC (3 mM), oltipraz (25 µM), or DMSO for 24 h on day 3. Cells were then treated with Dox (30 nM), Tax (2 nM), or DMSO on day 4. Cells were followed for an additional 5 d, with complete media and drug replacement on day 7. Cell survival was assessed on day 9 by manual cell count and normalized as described for each experiment.
Female NOD/SCID mice were purchased from Jackson Labs. The Animal Care and Use Committee of the Massachusetts Institute of Technology approved all animal procedures. For tumor regression studies, 1×106 cells were injected bilaterally into the mammary fat pad of 6–8-wk-old female NOD/SCID mice. After reaching 60–80 mm3, tumors were treated with PERKi (7.5 mg/kg/tumor) or DMSO by intratumoral injection on days 1, 2, 4, 5, 8, 9, 11, and 12, and Dox (2.5 mg/kg) or PBS by intraperitoneal (IP) injection on days 2, 5, 9, and 12 unless otherwise specified. Tumor volume over time and average tumor weight at sacrifice were measured and presented as the average ± standard error of mean for 10 tumors per treatment group.
For correlation analyses, the PERK gene expression signature was defined as the top 500 genes down-regulated in de-differentiated cells treated with 1 µM PERKi for 48 h. The PERK signature scores were calculated for each patient sample from human breast cancer (GSE3143, GSE41998) and glioma (GSE4412) datasets by summing the log-transformed normalized expression values for each probe in the signature set. High- and low-PERK signature groups were defined as the top or bottom 15% of samples within each group. Spearman's rho was used to measure correlation, and a p value was determined by Monte Carlo sampling as described previously [43].
All data are presented as mean ± standard error of mean unless otherwise specified. Student t test (two-tailed) was used to calculate p values, and p<0.05 was considered significant.
|
10.1371/journal.pntd.0003458 | Deep Sequencing of the Trypanosoma cruzi GP63 Surface Proteases Reveals Diversity and Diversifying Selection among Chronic and Congenital Chagas Disease Patients | Chagas disease results from infection with the diploid protozoan parasite Trypanosoma cruzi. T. cruzi is highly genetically diverse, and multiclonal infections in individual hosts are common, but little studied. In this study, we explore T. cruzi infection multiclonality in the context of age, sex and clinical profile among a cohort of chronic patients, as well as paired congenital cases from Cochabamba, Bolivia and Goias, Brazil using amplicon deep sequencing technology.
A 450bp fragment of the trypomastigote TcGP63I surface protease gene was amplified and sequenced across 70 chronic and 22 congenital cases on the Illumina MiSeq platform. In addition, a second, mitochondrial target—ND5—was sequenced across the same cohort of cases. Several million reads were generated, and sequencing read depths were normalized within patient cohorts (Goias chronic, n = 43, Goias congenital n = 2, Bolivia chronic, n = 27; Bolivia congenital, n = 20), Among chronic cases, analyses of variance indicated no clear correlation between intra-host sequence diversity and age, sex or symptoms, while principal coordinate analyses showed no clustering by symptoms between patients. Between congenital pairs, we found evidence for the transmission of multiple sequence types from mother to infant, as well as widespread instances of novel genotypes in infants. Finally, non-synonymous to synonymous (dn:ds) nucleotide substitution ratios among sequences of TcGP63Ia and TcGP63Ib subfamilies within each cohort provided powerful evidence of strong diversifying selection at this locus.
Our results shed light on the diversity of parasite DTUs within each patient, as well as the extent to which parasite strains pass between mother and foetus in congenital cases. Although we were unable to find any evidence that parasite diversity accumulates with age in our study cohorts, putative diversifying selection within members of the TcGP63I gene family suggests a link between genetic diversity within this gene family and survival in the mammalian host.
| Trypanosoma cruzi, the causal agent of Chagas disease in Latin America, infects several million people in some of the most economically deprived regions of Latin America. T. cruzi infection is lifelong and has a variable prognosis: some patients never exhibit symptoms while others experience debilitating and fatal complications. Available data suggest that parasite genetic diversity within and among disease foci can be exceedingly high. However, little is know about the frequency of multiple genotype infections in humans, as well as their distribution among different age classes and possible impact on disease outcome. In this study we develop a next generation amplicon deep sequencing approach to profile parasite diversity within chronic Chagas Disease patients from Bolivia and Brazil. We were also able to compare parasite genetic diversity present in eleven congenitally infants with parasite genetic diversity present in their mothers. We did not detect any specific association between the number and diversity of parasite genotypes in each patient with their age, sex or disease status. We were, however, able to detect the transmission of multiple parasite genotypes between mother and foetus. Furthermore, we also detected powerful evidence for natural selection at the antigenic locus we targeted, suggesting a possible interaction with the host immune system.
| Trypanosoma cruzi is a kinetoplastid parasite and the causative agent of Chagas disease (CD) in Latin America. T. cruzi infects approximately 8 million people throughout its distribution and causes some 13,000 deaths annually [1]. Chagas disease follows a complex course. Infection, often acquired in childhood, is generally lifelong but progression from the indetermined (asymptomatic) to symptomatic stage occurs in only 30% of cases [2]. A broad pathological spectrum is associated with clinical CD including potentially fatal cardiological and gastrointestinal abnormalities [3]. The relative contributions of parasite and host immunity in driving disease pathology are a matter of continuing debate [4]. Recently, for example, bioluminescent parasite infections in BALB/c mouse models have suggested that heart disease can progress in the absence of detectable local parasite load [5].
It is widely recognized that natural parasitic infections are often comprised of several parasite clones [6]. Malariologists use the term ‘multiplicity of infection’ (MOI) to describe when multiple Plasmodium sp. genotypes occur within the same host [7,8]. A similar phenomenon has been observed in T. cruzi in vectors (e.g. [9]), as well as mammalian reservoir hosts (e.g [10]) and humans hosts (e.g. [11]) using solid phase plating and cell sorting techniques. The occurrence of multi-genotype infections has fundamental implications for host immunity [12], as well as for accurate evaluation of pathogen drug resistance [13], transmission rate, epidemiology and population structure (e.g. [7,11]). The efficiency with which it is possible to sample pathogen clonal diversity from biological samples has soared in recent years with the advent of next generation sequencing. Deep sequencing approaches have long been applied to study the dynamics of HIV anti-viral therapy escape mutations. As a result amplicon sequencing increasingly features in a clinical diagnostic context [14]. Plasmodium falciparum MOI can be resolved at merozoite surface protein loci at far greater depths than possible by standard PCR approaches [15]. Furthermore, targeting low copy number antigens in parasite populations via amplicon sequencing can provide important clues to frequency-dependent selection pressures within hosts, between hosts and between host populations [16].
T. cruzi can persist for several decades within an individual host. Unsurprisingly perhaps, therefore, T. cruzi shows significant antigenic complexity. T. cruzi surface proteins are encoded by several large, repetitive gene families that are distributed throughout the parasite genome [17]. Among these gene families the mucins, transialidases, ‘dispersed gene families’ (DGFs), mucin-associated surface proteins (MASPs) and GP63 surface proteases comprise the vast majority of sequences—10–15% of the total genome size [17,18]. Whilst the role of some of the proteins encoded by surface gene families in host cell recognition and invasion is relatively well understood (e.g. the transialidases [19]), the role of others (e.g. the MASPs, DGFs) is not. Furthermore, the role each plays in evading an effective host response remains largely unknown.
The GP63 surface proteases are found in a wide variety of organisms, including parasitic trypanosomatids [20]. In Leishmania spp. GP63 proteases are the most common component of the parasite cell surface with crucial roles in pathogenicity, innate immune evasion, interaction with the host extracellular matrix and ensuring effective phagocytosis by macrophages [21]. In T. brucei subspp. the role of GP63 proteins is less well defined, although some protein classes are thought to be involved with variant surface glycoprotein processing between different life cycle stages [22]. In T. cruzi at least four classes of GP63 gene are recognized [20]. Like many GP63 proteases in Leishmania spp., surface expressed T. cruzi GP63 (TcGP63) genes are anchored to the cell membrane via glycosyl phosphatidylinositol moieties [23,24]. Among these are the TcGP63 Ia & Ib genes (collectively TcGP63I). TcGP63 Ia & Ib encode 78kDa 543 amino acid proteins, are expressed in all life cycle stages and are implicated in the successful invasion of mammalian cells in vitro [23,24].
In the current study we target TcGP63I genes as markers of antigenic diversity among three cohorts of Chagas disease patients: two in Cochabamba, Bolivia and one in Goias, Brazil. We also targeted a maxicircle gene for the NADH dehydrogenase subunit 5 to provide basic T. cruzi genotypic information for each case. Diversity at each of the two T. cruzi loci within each patient was characterized using a deep amplicon sequencing approach, generating several million sequence reads. Our results shed light on the diversity of parasite DTUs within each patient, as well as the extent to which parasite strains pass between mother and foetus in congenital cases. We were unable to find any evidence that parasite diversity accumulates with age in our study cohorts, or to detect a link between parasite diversity and clinical profile. However, we were able to detect evidence of putative diversifying selection within members of the TcGP63 gene family, suggesting a link between genetic diversity within this gene family and survival in the mammalian host.
Ethical permissions were in place at the two centres where human sample collections were made, as well as at the London School of Hygiene and Tropical Medicine (LSHTM). Local ethical approval for the project was given at the Plataforma de Chagas, Facultad de Medicina, UMSS, Cochabamba, Bolivia by the Comite de Bioetica, Facultad de Medcina, UMSS. Local ethical permission for the project was given at the Hospital das Clínicas da Universidade Federal de Goias (UFG), Goias, Brazil by the Comite de Etica em Pesquisa Médica Humana e Animal, protocol number 5659. Ethical approval for sample collection at the LSHTM was given for the overall study, “Comparative epidemiology of genetic lineages of Trypanosoma cruzi” protocol number 5483. Samples were collected with written informed consent from the patient and/or their legal guardian.
Parasite isolation protocols were different between centres. At the UMSS, 0.5 mL of whole venous blood was taken from chronic patients and inoculated directly into biphasic blood agar culture. T. cruzi positive samples were minimally repassaged and cryopreserved at log phase (precise repassage history unavailable). For infants, 0.5 mL of chord blood was taken at birth and inoculated into culture. Again, positive samples were cryopreserved at log phase after minimal repassage (precise repassage history unavailable). DNA extractions, using a Roche High-Pure Template Kit, were made directly from the cryopreserved stabilate. At the UFG, 17 mL of whole blood was collected into EDTA, centrifuged for 10 minutes at 1200g at 4°C and the plasma replaced with 8mL Liver Infusion Tryptone (LIT) medium. After a further 10 minutes at 1200g (4°C), the supernatant was again removed. Two mL of packed red blood cells were subsequently transferred to 3 mL of LIT medium and checked periodically for signs of epimastigote growth by light microscopy. Positive cultures were not repassaged. Instead primary cultures were stabilized by the addition of guanidine 6 M-EDTA 0.2 M (Sigma-Aldrich, UK). DNA extractions were made from the full volume using the QIAamp DNA Blood Maxi Kit (Qiagen, UK) according to the manufacturer’s instructions. Among Bolivian strains, DNA concentrations submitted to PCR were standardized after quantitation using a PicoGreen assay. In view of presence of human genetic material in Goias samples, parasite DNA concentrations were standardized to within the same order of magnitude via qPCR as previously described [25]. All samples collected for in this study are listed in Table 1.
The two areas studied have dissimilar histories in terms of Chagas disease transmission intensity. Vector-borne T. cruzi transmission in Goias and its surrounding states (where samples were collected—Table 1) was interrupted approximately 20 years before the sampling detailed in this study [26,27]. In the sub-Andean semi-arid valleys of Cochabamba and its environs, however, vector-borne domestic transmission is still a likely source of new infections, albeit at a reduced rate since intensive spraying campaigns in the mid 2000s [28]. Clinical data collected in this study were categorised simply into symptomatic and asymptomatic classes for statistical tests in view of samples sizes. Sub-categories within symptoms were defined as 1) Cardiopathy (including any electrocardiographic and/ or echocardiographic abnormalities, X-ray with cardiac enlargement. Patients with atypical cardiac abnormalities i.e. those not exclusively associated with Chagas disease, were included in the symptomatic class in the context of this study.) 2) Megaesophagous (including achalasia and barium swallow abnormalities) 3) Megacolon (constipation associated with dilation as by barium enema) and 4) Normal (no symptoms or signs on examination and a normal electrocardiogram) (Table 1)
Degenerate primers for a 450bp fragment of the maxi-circle NADH dehydrogenase 5 were designed as described in Messenger et al. 2012 [29]. Degenerate primer design for the TcGP63I family surface proteases (including Ia and Ib sublaclasses) [24] was achieved by reference to sequences retrieved from EuPathDB for Esmeraldo (TcII), CL Brener (TcVI), Silvio (TcI) and JR (TcI) (http://eupathdb.org/). Primer biding site positions in relation to TcGP63I putative functional domains are displayed in S1 Fig. Homologs were identified by BLAST similarity to a complete TcGP63I sequence (bit score (S) ≥ 1000). Alignments of resulting sequences were made in MUSCLE [30] and primers were designed manually to target a variable region within and between individual strains with a final size of 450bp. ND5b primer sequences were ND5b_F ARAGTACACAGTTTGGRYTRCAYA; ND5b_R CTTGCYAARATACAACCACAA. The final TcGP63 primers were TcGP63_F RGAACCGATGTCATGGGGCAA and TcGP63_R CCAGYTGGTGTAATRCTGCYGCC. Amplification was undertaken using the Fluidigm platform and a reduction of the manufacturer’s recommended number of cycles to total of 26 was made in an attempt to minimise PCR amplification bias. Thus, the manufacturer’s recommended conditions were adapted to the following protocol: one cycle of 50°C for 2 minutes, 70°C for 20 minutes, and 95°C for 10 minutes; six cycles of 95°C for 15 seconds, 60°C for 30 seconds, 72°C for 60 seconds; two cycles of 95°C for 15 seconds, 80°C for 30 seconds, 60°C for 30 seconds and 72°C for 60s; five cycles of 95°C for 15 seconds, 60°C for 30 seconds, 72°C for 60 seconds; two cycles of 95°C for 15 seconds, 80°C for 30 seconds, 60°C for 30 seconds and 72°C for 60 seconds; five cycles of 95°C for 15 seconds, 60°C for 30 seconds, 72°C for 60 seconds, and finally five cycles of 95°C for 15 seconds, 80°C for 30 seconds, 60°C for 30 seconds and 72°C for 60 seconds. Amplifications were performed using the FastStart High Fidelity PCR System (Roche). Three PCR reactions were pooled per sample prior to sequencing in an attempt to further reduce amplification biases [31]. Equimolar concentrations of ND5 and TcGP63I amplicons from 96 DNA samples were multiplexed on Illumina runs using dual index sequence tags (Illumina Inc). Sequencing was undertaken using a MiSeq platform using a 2 x 250 bp (Reagent Kit version 2) according to the manufacturer’s protocol. In addition to the clinical samples, we included a dilution series of control samples. The controls comprised artificially mixes of DTUs I-VI genomic DNA at equimolar concentrations. At the ND5 locus, comparison between the expected DTU abundance ratios and diversity of artificial control mixes and that defined via amplicon sequencing was made (S2 Fig.).
De-multiplexed paired-end sequences were submitted to quality control and trimming in Sickle [32] and mate pairs trimmed in FASTX Toolkit (http://hannonlab.cshl.edu/fastx_toolkit/). ND5, TcGP63 and contaminating sequences were then sorted against a reference using BOWTIE2 [33]. Individual paired reads were found to be overlapping in only a minority of cases. Thus we chose to proceed with analysis of a sequence fragment with a truncated central section for both targets. Further sequence manipulations were undertaken using FASTX Toolkit and custom awk scripts to parse files and concatenate each mate pair into a single sequence for downstream analysis. MUSCLE [30] was used for alignment of amplicon sequences in each patient sample. Next, analysis was undertaken in the Mothur software package [34] for the elimination of putative PCR chimeras and individual sequence clustering. The Shannon index of diversity was calculated at the intra-patient level based on sequence types (STs) defined at 97% and 99% identity in Mothur [34]. Comparisons of Shannon diversity were made between patients in each cohort (Bolivia chronic, Bolivia congenital, Goias chronic) via analyses of covariance and linear regression in the R package (http://CRAN.R-project.org). TcGP63I sequence datasets for patients from each cohort were then merged and analyses conducted using 97% and 99% STs defined with UPARSE [35] across patients. Weighted UniFrac distances between TcGP63I STs among samples were generated and subsequently clustered via a principal coordinates analysis in QIIME [36]. Significance of association between UniFrac clustering, disease status and age was tested in the vegan package in R [37]. Estimates of diversifying selection among TcGP63I STs were made in KaKs Calculator [38] using Yang and Neilson’s 2000 approximate method [39] and tested for significance using a Fisher’s exact test. Prior to selection calculations, sequences were clustered into 99% identity STs and singletons excluded in an attempt to exclude SNPs introduced as PCR artefacts. To test for diversifying selection across putative TcGP63I gene families (TcGP63Ia & Ib—97% cut-off as defined by Cuevas and colleagues [24]), 99% identity STs from each patient cohort were pooled (Table 2). To test for selection within TcGP63I gene families, STs within each 97% category (corresponding to TcGP63Ia & b respectively) were examined separately per cohort (Table 2). Amplicon sequences analysed in this study are available in the data appendix in supplementary information (S1 Appendix).
After quality filtering, trimming, decontamination and removal of unpaired reads, 6,736,749 reads were assigned to the ND5 mitochondrial marker and 871,855 to TcGP63I marker across the 92 clinical samples, perhaps reflecting higher copy number in the former than the latter. After trimming, the overlap between individual mate pairs was marginally too short to be assembled into a single read. Thus paired reads were first aligned against a full-length reference fragment and the central portion excised to remove any gaps and ensure correct alignment. Sequence depth thresholds per sample for inclusion were set for each dataset (Goias—ND5 & TcGP63–10,000; Cochabamba—ND5: 30,000; TcGP63 10,000; see Fig. 1). Reads from samples in excess of this threshold were discarded and samples with read counts below this threshold excluded. Our aim in setting the threshold was: 1) To include as many samples as possible while maintaining a good depth of coverage; 2) To standardise sampling intensity across individuals and thus facilitate comparisons between them.
The ND5 mitochondrial target was sequenced to provide DTU I-VI identification of parasites circulating within and among patients by comparison to existing data [29]. However, with reference to the results from the control samples—and due the necessary truncation of the sequence fragment—only three groups could be reliably distinguished, corresponding to the three major T. cruzi maxicircle sequence classes [40]. The three groups corresponded to TcI, TcII and TcIII-VI respectively. Furthermore, in reference to the control mixes, we found evidence that amplification bias dramatically skewed the recovery of sequence types (STs) towards the TcIII-VI group. Some skew is expected, as these four DTUs (TcIII-VI) share the same maxicircle sequence class, and this class is thus more abundant in the control mix. However, TcI and TcII—which should have in theory been present as 16% (1/6) of all sequences in the controls respectively—were in fact present (on average) at only 2.9% and 0.03% among the four samples where all three STs were recovered (S2 Fig.). Amplicon sequencing from the two most concentrated controls (57 ng/uL and 125 ng/uL genomic DNA respectively) resulted in poor sequence yields and a failure to recover all three STs.
Unsurprisingly perhaps in the light of the control data, most clinical samples were dominated by sequences from a single group, with minor contributions from others (Fig. 2). Indeed sequences recovered from many strains were monomorphic at the 97% identity level—especially in Cochabamba. As such, comparisons based on ND5 are necessarily descriptive and meaningful alpha (within sample) and beta (between sample) diversity statistics were not calculated. Fig. 2 shows the distribution of DTUs among samples as defined by the ND5 locus. Most Cochabamba chronic cases samples were assigned to a single sequence within the TcIII-VI group (likely to be TcV, as we defined with standard genotyping assays [41] with the exception to two TcI cases—PCC 240 and PCC 289 (Fig. 2, Panel B). Sequence-type diversity in Goias was considerably higher (Fig. 2, Panel A). In this case the TcII group, rather than the TcIII-VI group, predominated. Unlike in Bolivia, sequences from other groups were present alongside TcII in multiple patients but at frequencies two orders of magnitude lower. Congenital pairs that originated from Cochabamba resembled chronic cases from the same region in their DTU composition (TcIII-VI group predominant, Fig. 2, Panel C). Strikingly, mother/child pair CIUF65 (B5) and CIUF75 (M5) share similar mixed infection profiles (TcI/ TcIII-VI) at similar relative abundances (c.1:1000), consistent with the minor to major genotype abundance ratios observed in Goias. The same is also true for the Goias congenital pair (Fig. 1) which both showed TcII/TcI mixes. Finally, sequential isolates taken from the same Goias chronic patient at different time points suggest that minor abundance genotypes are not always consistently detectable in the blood (Fig. 2): TcI is absent at first sampling of patient y, but present at the second sampling. For patient z, the TcIII-VI genotype is only present in the first of the two sample points. For both Cochabamba and Goias, reference to the control data suggests that ‘minor genotypes’ could be substantially more abundant in the patients than the amplicon sequence data suggest.
Alpha diversity measurements aim to summarise the diversity of species (in this case STs), within an ecological unit (in this case a host). We summarized the number of STs and their relative abundance in each of our samples, using the Shannon Index (SI) [42]. Among non-congenital cases, our aim was to evaluate possible associations between TcGP63I antigenic diversity and several epidemiological and clinical parameters—age, sex and disease status. We used analyses of covariance (ANCOVA) to test for the effect of these parameters on intra-host antigenic diversity (STs defined both at 97% and 99% for comparison), combining continuous (age) and categorical (sex, clinical forms) data. In Cochabamba, regardless of the order in which parameters were included as factors in the model, there was no evidence for a main effect of age, sex or symptoms on alpha diversity (SI) at either ST divergence level (97% ST Age: p = 0.734; Sex: p = 0.298; clinical form: p = 0.136. 99% ST—Age: p = 0.854; Sex: p = 0.169; clinical form = 0.0988). Similarly, ANCOVAs were non-significant for an association between the SI and age, sex or symptoms in Goias (97% ST—Age: p = 0.382; Sex: p = 0.535; clinical form: p = 0.486. 99% ST—Age: p = 0.319; Sex: p = 0.696; clinical form: p = 0.697). Finally, we undertook linear regressions of SI with age in each population. As one might expect from previous ANCOVAs, no significant correlation was detected (Goias R2 = 0.0233, p = 0.340 (97% ST); R2 = 0.0256, p = 0.3049 (99% ST) Cochabamba R2 = 0.0287, p = 0.429 (97% ST); R2 = 0.0230 p = 0.479(99% ST)).
Congenital comparisons were made pairwise between mother and infant at 99% ST similarity. In addition to the ten matched isolate pairs from Cochabamba, a single pair from Goias was also included (6718 & 6720) in the comparisons. The results of the alpha diversity comparisons are shown in Fig. 3, and read depths were balanced between samples. In terms of the absolute number of STs identified, infants exceeded mothers in most instances (pairs 2, 3, 4, 5, 6, 8 & 9). In the remaining cases however (4/11), the number of antigenic sequence types was greater in the mother. Shannon diversity index comparisons between mothers and infants, which also takes ST abundance into account, suggested that some differences (e.g. pairs 4, 5 &6) might be marginal (Fig. 3).
Individual sample sequence datasets within each of the different study cohorts (Cochabamba congenital, Cochabamba non-congenital and Goias) were merged to facilitate analysis of the distribution of antigen 99% STs among individuals (i.e. beta-diversity comparisons). Pairwise weighted Unifrac distances were calculated within cohorts of chronic cases from Cochabamba and Goias to examine whether the sequence diversity of the TcGP63I antigenic repertoire present in each patient could be associated with disease outcome. Principal coordinate analyses of the resulting matrices are displayed in Fig. 4. Among cases from Goias, repertoires varied considerably among cases, with several outliers. However, repertoires from symptomatic and asymptomatic cases were broadly overlapping in terms of sequence identity, and no clustering was noted among different symptom classes either (Fig. 4, Plot B). Permutational multivariate analysis confirmed the absence of a link between ST clustering and symptoms as well as symptom classes (p = 0.77 & 0.74 respectively). However, ST clustering and age were weakly associated (p = 0.049), consistent perhaps with exposure of individuals among different age groups to different circulating parasite genotypes at their time of infection. TcGP63I read yields permitted comparisons for only two pairs of sequential isolates from the sample patients—x and y (see Table 1)—both of which showed closely clustering, although non-identical, profiles. TcGP63I diversity between Cochabamba chronic cases was arguably lower, with the exception of two outliers unambiguously identified as TcI with reference to the ND5 locus (all others were classified as TcIII-VI—likely TcV). Again, however, symptomatic and asymptomatic cases were broadly overlapping.
Sequence type profile comparisons among Cochabamba congenital cases were made for 99% STs and are displayed in heatmap format in Fig. 5. There are two key features of interest. The first is that profiles in mother an infant can match very closely (e.g. pairs 2&6). The second is that novel STs were present in the infant sample with respect to the mother in half of the cases. Indeed, in pair 9, the infant profile was radically different to that of the mother.
Trimmed TcGP63 reads, pre-filtered for quality and PCR errors, were pooled within each study site (Bolivia, Goias). To further reduce minority SNPs and PCR errors, STs were defined at 99% with each site in UPARSE [35]. Ka/Ks ratio estimates within each study area indicated a significant excess of synonymous mutations among STs (Goias = 0.8354, Bolivia = 0.7515) averaged across sites (Table 2). However, when calculations were based on diversity present among well represented STs of each gene family member (TcGP63Ia and TcGP63Ib, 97% cut-off [24]) a powerful and significant excess of non-synonymous substitutions was noted within each study area (Ka/Ks, Goias, ST1 = 2.6436, ST4 = 6.3415; Bolivia ST3 = 2.8059; Table 2). Again, calculations were based not on individual sequences, but rather 99% STs within predefined 97% clusters. The position of the 97% STs in question is shown in the tree in S3 Fig., with clear similarity between those clusters under apparent diversifying selection (Goias ST1 & 2, and Bolivia ST3) with TcGP63Ia and TcGP63Ib references respectively [24].
In this study our aim was to collect a cohort of T. cruzi samples from clinical CD cases, representative of different endemic regions and of different ages and disease presentations, to explore links between CD epidemiology and multiplicity of infection. To provide a robust, sensitive and quantifiable means of assessing intra-host parasite diversity we first implemented standardized parasite isolation (and enrichment) strategies within each study cohort. Latterly, we developed an amplicon sequencing approach to profile parasite diversity within each patient. Given the relatively short (400–500bp) read lengths generated by next generation sequencing platforms (at the time of experimentation), we chose a rapidly evolving maxicircle gene (ND5) in an attempt to resolve DTU level diversity ([29]). Current multilocus nuclear targets are generally too long (500bp+) to meet our selection criteria [43]). To explore antigenic diversity, we chose a putatively low (5–10) copy number gene family member TcGP63I, expressed on the parasite surface during the amastigote and trypomastigote lifecycle stage and thus exposed to the human immune system [24]. Given that both ND5 and TcGP63I are present as several copies per parasite genome (and potentially show inter-strain copy number variation e.g. [44]), one cannot presume a 1:1 relationship between ST and parasite individual, even if we were able to account for the PCR amplification bias we detected. The identification of a genetically, variable, single copy, surface expressed antigen locus is a major challenge in T. cruzi—antigen genes are by their nature highly repetitive [17,18]. TcGP63I, with its relatively low copy number represents the closest currently available fit, and, as we have shown, provides a useful target for revealing intra-host antigenic diversity. Merozoite surface proteins (MSP) 1 and 2 have traditionally provided useful targets for detecting MOI in P. falciparum (e.g. [45,46]. Furthermore, amplicon sequencing of the MSP locus has been successfully proven to reveal as many as six-fold more variants than traditional PCR-based approaches [15].
The substantial historical interest in defining MOI among P. falciparum owes itself to the strong correlation between MOI and rate of parasite transmission [47]. As such, fluctuations in transmission intensity can be tracked to evaluate the efficiency of vector eradication campaigns, drug treatments, the introduction of insecticide-treated nets etc—without the need to directly estimate the entomological inoculation rate. Evaluation of CD transmission intensity has its own challenges. The presence of infected individuals, triatomine vectors in domestic buildings, incrimination of vectors via human blood meal identification (e.g. [48]) can all help to build the overall picture. However, parasite transmission is likely to occur in only a tiny proportion of blood meals [49,50], and vector efficiency is thought to vary considerably between triatomine species [51]—thus the presence of vectors is no guarantee of transmission. Infection with T. cruzi is lifelong, thus positive patient serology is not a reliable indicator of active parasite transmission either. Traditionally, active T. cruzi transmission has been implied from positive serology among younger age classes. Especially in hyperendemic areas of Bolivia, Paraguay and Argentina the proportion of seroprevalent individuals increases with age [52,53]. MOI in T. cruzi patients should follow a similar trend given a stable force of infection. Furthermore MOI comparisons between disease foci could, controlling for age, facilitate an appreciation of relative transmission intensities—a useful tool for those who wish to track the efficacy of interventions. In the current study, however, we were unable to identify a correlation between MOI and age, even once patient sex and clinical form had been corrected for. Our inability to validate this fundamental prediction has many possible causes. First, patients in each cohort originate from different communities within each study area (Table 1). Micro-geographic variation in T. cruzi genetic diversity is commonly observed (e.g. [11,54,55], and the same is likely to be true for infection intensity. Thus, if patients from different sites share dissimilar histories in the intensity and diversity of exposure to T. cruzi clones, comparisons between them are difficult to make. Secondly, the relationship between MOI and age is not necessarily linear. If a degree of cross-genotype immunity accumulates with exposure, one might expect a slower increase in intra-host antigenic diversity in older age groups. However, this was not the case in our dataset and neither a linear, nor a unimodal relationship could be established.
Amplicon sequencing approaches to the study of transmission patterns in human parasites have so far been restricted to those species that replicate and reach high parasitemias in peripheral blood (i.e. T. brucei [56] and P. falciparum [13,15]). T. cruzi trypomastigote circulating parasitemias, as measured by qPCR, are thought to vary considerably between acute (400 parasites/ml), newborn (150–12000 parasites/ml) and chronic (3–16 parasites/ml) cases [25,57]. Nonetheless, they remain several orders of magnitude lower than those that occur during T. brucei or P. falciparum infections. Low circulating T. cruzi parasitemia presents major problems to studies that aim to achieve molecular diagnosis of CD in chronic cases and ours is no exception. One problem is that much of the parasite diversity present in the host is likely to be sequestered in the tissues at any give time [58], as our sequential samples from Goias also suggest. Thus blood stage parasite genetic diversity may be a poor representation of that actually present in the host. Another confounder is culture bias, by which differential growth of clones in culture, as well as loss of clonal diversity during repassage can both influence diversity estimates. Attempts to generate amplicon sequence data directly from clinical blood samples would likely to be thwarted by low circulating parasitemia [25, 56]. Instead we elected to enrich for parasite DNA via culture—in Goias without further repassage, but in Bolivia with at least one repassage before cryopreservation. Low circulating parasitemia in Chagas patients also means it is possible that amplicon-sequencing strategies might rapidly ‘bottom out,’ if few parasites are present within a sample. In our dataset, for example, at the ND5 locus, minority DTUs at 97% divergence can be present as a proportion of < 1 in 1000 (Fig. 1), with the implication that several thousand parasites must be present in the sample. In both Goias and Bolivia matched instances occurred in congenital cases where TcI exists in mother and infant as the minor DTU at similar relative abundance (i.e. 1 in 1000, Fig. 1). It is highly unlikely that these data directly reflect chronic CD parasitemia levels. Instead, with reference to the data we obtained from the controls, PCR amplification bias is a more likely source of unrealistic major to minor genotype ratios. As such, the fourfold over-representation of a ST in the original sample, for example, can result in 100–1000 fold over-representation after PCR. However, while the relative abundance of sequence types recovered using the amplicon approach may be an inaccurate reflection of those present for both ND5 and TcGP63, similar profiles between mother and infant suggests that this bias is likely to be consistent across samples. Thus comparisons between samples are still valid. Furthermore for ND5 at least it seems that T. cruzi frequently exchanges mitochondrial (maxicircle) genomes with little apparent evidence of nuclear exchange [11,29]. Fusion of maxicircle genomes occurs transiently during T. brucei genetic exchange events [59], and may also do so in T. cruzi. Even though standard maxicircle genotyping of progeny only ever reveals a single parent in both species, it is possible that heterologous maxicircle sequences may persist at low abundance in parasite clones. Such a phenomenon could explain the DTU sequence type ratios observed, and this study is the first to sequence a maxicircle gene to this depth.
There is general consensus in the literature is that the likelihood of congenital CD transmission is not strongly influenced by the genotype of the parasite infecting the mother [60–62]. Nonetheless, the majority of cases are reported in the Southern Cone region of South America, providing a circumstantial link with major human-associated T. cruzi genotypes TcV TcII, and TcVI. In this study, in the one mixed infection we found, major and minor DTUs (TcVI / TcI) detected in the mother at the ND5 locus were recovered from the infant in similar proportions. TcGP63I beta diversity comparisons of STs defined at 99% showed substantial sharing of between mother and infant (Fig. 5). However, both beta diversity comparisons (Fig. 5) and total ST diversity (alpha) comparisons (Fig. 3) at 99% indicate that while maternal diversity sometimes exceeds that of the infant (explicable perhaps by sequestration in the mother and selective or stochastic trans-placental transfer), the reverse is frequently true. The occurrence of STs in the infant, not present in the mother, has several possible explanations. The infants sampled in this study were neonates, thus superinfection can be ruled out as a source of further parasite clonal diversity. A recent study of infected neonates in Argentina estimated mean infant parasitemia at 1,789 parasites/ml via qPCR—far in excess of that one might expect in the mother [57]. Thus the parasite sample size discrepancy between mother and infant perhaps explains the unexpected levels of diversity in the infant. Even though the TcGP63I gene family is apparently under intense diversifying selection, it seems unlikely that point mutation could generate novel variants over such a short time scale to explain genetic diversity in the infant. Structural variants and homologous recombination are a potential source of diversity, although most, if not all of recombinants should have been excluded in the quality filtering stages, and would be hard to distinguish from PCR chimeras in any case.
Many important T. cruzi surface genes belong to large, recently expanded paralogous multigene families [17]. The abundance of these gene copies highlights their likely adaptive significance in terms of infectivity and host immune evasion, especially because trypansomatids exert so little control of gene expression at the level of transcription [63]. In Leishmania major, for example, it has been recently shown that gene amplification may rapidly duplicate segments of the genome in response to environmental stress [64]. As well as expansion, adaptive change is also likely to occur at the amino acid level among members of paralogous gene families, as has been suggested for T. brucei [65]. Despite the relatively small size of the TcGP63I gene family, the amplicon sequencing approach we employed allowed us to explore selection at the level of the gene within the population, i.e. within and between parasite genomes within and between hosts at the population level. Highly elevated non-synonymous substitutions suggest intense diversifying selection within TcGP63Ia and TcGP63Ia STs respectively for those assigned to TcII or TcI. STs from patients infected with TcIII-TcVI (putative TcV) showed few apparent substitutions (Table 2), perhaps consistent with the recent origin of this DTU [66]. The sequence fragment we studied was outside the zinc binding domain of this metalloprotease, indicating selective forces can act on this protein independent of its core proteolyic function, perhaps through repeated exposure to host immunity.
It is important not to overlook the potential importance of multiclonal infections for parasitic disease, both as markers of population level factors such as parasite transmission, but also at the host level, including immunity and disease progression. In this study we have developed an amplicon sequencing approach to probe parasite genetic diversity within and among clinical CD cases to unprecedented depth. While our approach shows the power of this amplicon-seq to resolve diversity in clinical and congenital CD cases, it also highlights the potential biases that might be introduced with the addition of a PCR step. A tool that allows the accurate evaluation MOI would be valuable for tracking transmission rates at restricted disease foci (i.e. villages, outbreaks) in the context of measuring the success of intervention strategies. A similar tool could provide a powerful means of longitudinal tracking of T. cruzi infections in terms of disease progression, treatment failure and immunosuppression. Here we demonstrate that amplicon sequencing could have a role to play in this context. However, as sequencing costs decline and reference genome assemblies improve, whole genome deep sequencing, perhaps even of individual parasite cells, becomes and increasingly viable option as it already has for Plasmodium sp. [7,67].
|
10.1371/journal.pgen.1000321 | Non-Coding RNA Prediction and Verification in Saccharomyces cerevisiae | Non-coding RNA (ncRNA) play an important and varied role in cellular function. A significant amount of research has been devoted to computational prediction of these genes from genomic sequence, but the ability to do so has remained elusive due to a lack of apparent genomic features. In this work, thermodynamic stability of ncRNA structural elements, as summarized in a Z-score, is used to predict ncRNA in the yeast Saccharomyces cerevisiae. This analysis was coupled with comparative genomics to search for ncRNA genes on chromosome six of S. cerevisiae and S. bayanus. Sets of positive and negative control genes were evaluated to determine the efficacy of thermodynamic stability for discriminating ncRNA from background sequence. The effect of window sizes and step sizes on the sensitivity of ncRNA identification was also explored. Non-coding RNA gene candidates, common to both S. cerevisiae and S. bayanus, were verified using northern blot analysis, rapid amplification of cDNA ends (RACE), and publicly available cDNA library data. Four ncRNA transcripts are well supported by experimental data (RUF10, RUF11, RUF12, RUF13), while one additional putative ncRNA transcript is well supported but the data are not entirely conclusive. Six candidates appear to be structural elements in 5′ or 3′ untranslated regions of annotated protein-coding genes. This work shows that thermodynamic stability, coupled with comparative genomics, can be used to predict ncRNA with significant structural elements.
| Recent advances in DNA sequence technology have made it possible to sequence entire genomes. Once a genome is sequenced, it becomes necessary to identify the set of genes and other functional elements within the genome. This is particularly challenging as much of the genomic sequence does not appear to perform any function and is loosely referred to as “junk.” Identifying functional elements among the “junk” is difficult. Experimental methods have been developed for this purpose but they are time-consuming, expensive, and often provide an incomplete picture. Thus, it is important to develop the ability to identify these functional elements using computational methods. Protein-coding genes are relatively easy to identify computationally, but other categories of functional elements present a significantly greater challenge. In this work, we used a computational approach to identify genes that do not encode for a protein but rather function as an RNA molecule. We then used experimental methods to verify our predictions and thereby validate the computational method.
| Non-coding RNA (ncRNA) are functional RNA transcripts that are not translated into protein (i.e., not messenger RNAs). Research, particularly over the last 10 years, has shown that they perform a wide range of functions in the cell [1]–[4]. Despite the growing body of knowledge about ncRNA, it is likely that many ncRNA remain undiscovered. Data from high-throughput experimental methods show that much of the intergenic DNA in eukaryotic genomes is transcribed and may be ncRNA [5]–[10]. Even in Saccharomyces cerevisiae, one of the most thoroughly studied model organisms, there is evidence that only a fraction of the ncRNA is known. Tiling arrays, large-scale cDNA libraries, and serial analysis of gene expression (SAGE) experiments have all shown transcription from many locations in the genome that appear to be unannotated ncRNA genes [11]–[14]. This along with recent identification of new protein coding genes such as YPR010C-A in 2006 shows that even in this best-studied Eukaryote, we still do not know the complete gene set [13].
Computational methods for accurate ncRNA gene prediction remain elusive. The development of such methods are crucial for identifying ncRNA that are difficult to detect experimentally such as those expressed at low levels or under unusual conditions. They are also needed to reduce the time and expense required to perform experimental methods, particularly when considering the large number of species of interest. The challenge of predicting ncRNA genes rests with the fact that they lack common primary sequence features and demonstrate poor cross-species sequence conservation [15],[16]. They do not have start codons, stop codons or open reading frames which serve as key signposts for protein-coding genes and cannot be located using simple sequence searches.
Some success with ncRNA gene prediction has been achieved by focusing on specific sub-classes of ncRNA that share common features. Examples include tRNAs, tmRNAs, snoRNAs (C/D box and H/ACA box), and miRNAs [17]–[32]. In S. cerevisiae, computational screens for C/D box [19] and H/ACA box snoRNAs [20] have identified several new snoRNA genes.
Additional ncRNA screens in S. cerevisiae have included searches for polymerase III promoters, searches in larger than average intergenic regions [33] and searches for ncRNA structural features using the QRNA program. The QRNA program was used to search pair-wise alignments for patterns of compensatory mutations consistent with base-paired secondary structure [34]. These regions were then tested experimentally to determine if they expressed a transcript likely to be ncRNA. Together, these three methods resulted in identification of 6 novel ncRNA that were supported by experimental evidence (RNA170, snR161, snR82, snR83, snR84, RUF5-1/2). In another study, the S. cerevisiae genome was analyzed using the RNAZ program [35]. This program is based on the same principals as the QRNA program and uses multiple, cross-species sequence alignments to search for patterns of compensatory changes suggestive of secondary structure. RNAZ also includes thermodynamic analysis. A total of 572 candidate regions were identified as potentially containing unannotated ncRNA candidates using the RNAZ program [35],[36]. Publicly available data sets were used to provide general support for these predictions but no detailed experimental analysis was performed on individual predictions.
In this work ncRNA genes are predicted in S. cerevisiae based solely on the thermodynamic stability of ncRNA structures as proposal by Maizel in the late 1980's [37]–[39]. Maizel theorized that structural ncRNA are thermodynamically more stable than random sequences. An influential paper by Rivas & Eddy entitled “Secondary structure alone is generally not statistically significant for the detection of noncoding RNAs” suggested that Maizel's approach was generally not effective for structural ncRNA discovery [40]. Based on this conclusion, many investigators turned away from thermodynamic based approaches for ncRNA discovery to methods based on compensatory changes in cross-species alignments[31]. However, a growing body of evidence has been accumulating suggesting that thermodynamic stability is a discriminating feature of many classes of structural ncRNA [41]–[43]. In this work, we build on this result to not only evaluate the thermodynamic stability of known structural ncRNA but also to use it for structural ncRNA discovery.
The work presented here demonstrates the value of thermodynamic structural stability, as summarized in a Z-score, for discovery of structural ncRNA. It also explores the impact of window size and step size on the sensitivity of ncRNA identification. Sets of positive and negative control genes were evaluated to determine the effectiveness of the approach. This approach was then applied to predict ncRNA genes on chromosome six of S. cerevisiae. The analysis was repeated independently in S. bayanus and the gene predictions common to both genomes comprised the final set of gene predictions. Experimental validation of these predictions show that four ncRNA transcripts are well supported by northern blot analysis, rapid amplification of cDNA ends (RACE), and publicly available cDNA data. One additional ncRNA candidate is also supported by experimental data but the data is not entirely conclusive. Six of the predicted candidates appear to be structural elements in 5′ or 3′ untranslated regions (UTRs) of annotated protein-coding genes.
The thermodynamic stability of potential ncRNA candidates was evaluated using a Z-score based on the minimum folding energy (MFE) determined by RNAfold [44]. The Z-score represents the number of standard deviations that the MFE of a native sequence, x, deviates from the mean MFE of a set of shuffled sequences of x (see Materials and Methods).
A key variable in calculating the Z-score for ncRNA discovery (as opposed to evaluating known structural ncRNA) is the length of the sequence to be evaluated. As ncRNA vary in length and structure, no single window size is expected to be optimal for ncRNA gene identification. Short structural elements will probably only be detected with relatively short window sizes while longer structural elements will probably only be detected with relatively longer window sizes. To identify the window sizes most appropriate for ncRNA discovery, values ranging from 20 nt to 200 nt were investigated and incremented in steps of 5 nt (window delta).
A scanning approach was used to computationally search for potential structural elements within a test sequence. A starting minimum window size was selected and this window was used to scan the test sequence starting at the beginning of the sequence and moving each time by the amount of the step size (our analysis used a step size of 5 nt). A Z-score was calculated for each window position. Once the entire test sequence was evaluated using this fixed window length, a new window length was selected by increasing window length by the amount of the window delta (our analysis used a window delta of 5 nt). The test sequence was evaluated in the same manner using the new window size. This process was repeated until all window sizes had been evaluated.
Since the same test sequence was evaluated using multiple window sizes, it was necessary to determine the impact of multiple hypothesis testing. In lieu of a Bonferroni correction, negative control sets were evaluated using the same number of window sizes and step sizes.
Any windows producing a “significant” Z-score during the scanning process were considered candidate regions for structural ncRNA. The Z-score cutoff considered to be “significant” was determined by evaluating positive and negative test sets. It was sometimes the case that multiple, overlapping windows, of several lengths, produced “significant” Z-scores. In such cases, the region encompassed by all the overlapping windows constituted the candidate region.
Once candidate regions were identified, primers were designed within these regions to determine whether they produced a transcript and to identify the transcript boundaries. The primers were designed as close as possible to the middle of the candidate regions. The exact position of the primer was dictated by the need to satisfy the fairly stringent requirements of the rapid amplification of cDNA ends (RACE) procedure (See Materials and Methods).
Positive and negative control sets were compiled to test if the Z-score could be used to distinguish known ncRNA from non-functional sequences as suggested by previous investigators [41]–[43]. The positive control set was drawn from the list of annotated ncRNA in the Saccharomyces Genome Database (SGD) [45] (Table 1). The tRNA and rRNA genes were not included in the positive control set as they can be identified with great accuracy using existing tools [17] and because tRNA are known to produce poor thermodynamic footprints [40],[41],[46]. The positive control set consisted of four snoRNA genes and all of the remaining known ncRNA (Table 2).
Three negative control sets were created to cover the full range of negative control cases. The first negative control set consisted of 20 randomly generated sequences of 300 nt in length. This set was used because it was known not to contain any unannotated genes. The shortcoming of this control set is that it likely fails to capture the nuances of nucleotide distributions in S. cerevisiae. The randomly generated sequences had a GC content of ∼40%, ranging from 35.0% to 49.3%, reflecting the GC content of S. cerevisiae. A second negative control set was created by randomly shuffling the positive control set. Each sequence was shuffled preserving sequence length as well as its mono- and di-nucleotide composition using the “squid” utilities [47]. The third negative control set was generated by selecting six intergenic regions from the S. cerevisiae genome. Intergenic regions were chosen as a control instead of coding regions because the GC content in the S. cerevisiae genome differs between protein coding regions and non-protein coding regions. Since the ultimate goal was to search for ncRNA in intergenic regions, it was best to select a test set representative of these regions. The untranslated regions (UTR) of most genes in S. cerevisiae are not mapped so the actual intergenic regions are generally unknown. In order to minimize the possibility of choosing a region that contained an unannotated structural element, six intergenic regions were chosen that are flanked on one side by a gene with a known, short (<40 nt) 5′ UTR, unlikely to form a structure. A window of 300 nts from the 5′ end of the open reading frame (ORF) of each of these genes was used as a negative control test sequence (Table S1).
Z-score values calculated for the 20 randomly generated negative control sequences revealed that large negative Z-scores are often generated when using window sizes of less than 65 nt. With these short window sizes, many shuffled sequences have a calculated minimum folding energy of zero or close to zero and the Z-score distribution of the shuffled sequences is narrow. This produces a small value for the standard deviation. If the MFE of the original, unshuffled sequence is even slightly above zero, it will be many standard deviations from the distribution mean and produce a large negative Z-score. When examining window sizes of 75 nt or greater, two (Random9 and Random13) of the 20 randomly generated sequences produced a Z-score less than −3.5 (Table S2, Figures S1 and S2). The total length of sequence producing a Z-score ≤−3.5 was 295 nt and represented 5.0% of the nucleotides in the entire randomly shuffled test set (Table 3).
Z-score values calculated for the 6 intergenic sequences of the second negative control set produced a pattern very similar to that of the randomly generated sequences. For window sizes less than about 65 nt, large negative Z-scores were generated. Window sizes longer than 75 nt did not produce any Z-scores less than −3.5 with the exception of the intergenic sequence between genes PTP1 and SSB1. The first 190 nt of this sequence produced Z-scores as low as −4.7 for various window sizes (Table S2). This may represent either a false positive or may suggest the presence of a structural feature (ncRNA or long PTP1 5′ UTR structure). This 190 nt region represents approximately 10.5% of the total length of the intergenic negative control set.
The final negative control set consisted of shuffled sequences of the positive control set (Table 2). Of these, portions of 5 out of 16 sequences (31%) produced Z-scores less than −3.5 (Table S2 and Figures S3 and S4). The total sequence length included in these regions represented 8.1% of the total negative control set length.
All of the sequences in the positive control set produced Z-scores less than −3.5 for multiple window sizes (Table S3, Figures S5 and S6) with the exception of three genes. These genes were snR76, RNA170, and SRG1.
The snR76 gene is a C/D box snoRNA and it is questionable whether structure plays a significant role in the function of this gene. The SnoScan program was written explicitly to predict C/D box snoRNA and has been used successfully to predict these genes in both D. melanogaster and S. cerevisiae [19],[48]. Known C/D box snoRNA were used to identify features shared among this family of ncRNA. Only one of the six criteria identified is related to structure (terminal stem base pairings). This base pairing consists of only 4–8 bps and is not always present [19]. This is in stark contrast to the snoGPS program used to identify H/ACA snoRNA [20]. The snoGPS program was trained using known H/ACA snoRNA examples and includes secondary structure as a key element in H/ACA box snoRNA detection. Results from these snoRNA gene identification efforts strongly suggest that structure is generally not a significant component of C/D box snoRNA genes.
SRG1 is a ncRNA gene that has been shown to repress the expression of its neighboring gene SER3 [49]. Transcription of SRG1 interferes with the binding of SER3 activators in its promoter. This mechanism suggests that SRG1 fulfills its role as a transcriptional repressor through its transcription rather than through a significant structural component.
The RNA170 gene was discovered through a genome-wide search of Polymerase III box A and B consensus sequences [33]. Its function and mechanism of action are unknown. It seems likely that this ncRNA does not require a significant structural component to perform its function.
The total sequence length encompassed by a Z-score less than −3.5 in the positive control set represented 41% of the total sequence evaluated. If snR76, SER3 and RNA170 are removed from the set, 46% of the positive control set produces a Z-score <−3.5 (Table 3). Window sizes of 75 nt to 85 nt were crucial for identifying the short ncRNA such as snR6.
To summarize, three negative control sets were used consisting of a set of randomly generated sequences, a set of intergenic sequences, and a set of shuffled positive controls. The percent of sequence producing a false positive indication (i.e., Z-score ≤−3.5) for each of these sets was 5.0%, 10.5%, and 8.1%, respectively (Table 3). We examined the regions producing Z-scores ≤−3.5 for unusual GC content that might explain the large negative Z-score but found nothing significant in these regions (Table S4). For the positive control set, 13 of the 16 genes produced a Z-score ≤−3.5, encompassing 41% of the total sequence length of the set (Table 3). There is good reason to think that the three genes in this set failing to produce a Z-score ≤−3.5 do not contain structural features.
Analysis of the positive and negative control sets provided the following conclusions, (1) Evaluating window sizes less than 65 nt produces many false positives, (2) A Z-score value of −3.5 is useful for discriminating known ncRNA from non-functional sequence, (3) The percent of false positive sequence was observed to be ∼5.0–10.5% when using a cut-off Z-score value of −3.5.
Evaluation of the positive and negative control sets showed that the Z-score was useful for discriminating known structural ncRNA from non-functional sequence. To apply the approach to de novo gene prediction it is necessary to scan through a large test sequence (i.e., a chromosome) in search of regions that produce Z-score values indicative of structural ncRNA. To test the effectiveness of our approach for ncRNA discovery, and to determine the optimal parameters for the search, we performed two tests. We evaluated our ability to detect known ncRNA (Table 1), then we performed a detailed analysis of optimal search parameters using a small subset of ncRNA.
First, each annotated, nuclear encoded ncRNA (excluding rRNA), along with 200 nt upstream and downstream of the gene, was used as a test sequence. Z-scores were calculated on the ncRNA strand using the following parameters: window sizes = 75 to 200 nt, step size = 5 nt, window delta = 5 nt. The known ncRNA were considered detected if the center of the window(s) producing a Z-score ≤−3.5 overlapped the gene.
100% of the snRNA were detected, 72.4% of the H/ACA box snoRNA were detected, and 23.9% of the C/D box snoRNA genes were detected. Only 16% of the tRNA genes were detected. This result is consistent with previous reports of poor detection of tRNA based on a Z-score-type search criteria [40],[41],[43]. Clote et al [41] suggested that this may, in part, be due to the extensive post-transcriptional modifications that occur to tRNA that are not accounted for in the MFE calculation based on unmodified sequence. The percent of tRNA detected was a function of the tRNA length. 10.4% of the tRNA shorter than 75 nt (192 total) were detected while 34.6% of tRNA greater than 75 nt (83 total) were detected.
This ncRNA data can also be used to show the impact of using a single window size or a large step size on ncRNA detection (Table 4). The table provides the percent of H/ACA box snoRNAs detected when only a single window size was used to perform the analysis. The impact of using different step sizes (5 nt, 25 nt and 50 nt) is also presented. Using a single window size, as opposed to several sizes, reduces the number of snoRNA detected. The number of H/ACA snoRNA detected by evaluating all window sizes from 75 nt to 200 nt was 72.4%, which is greater than the number detected by using any single window size. The number of H/ACA snoRNA detected for a given window size decreases as the step size increases. These results can provide guidance for choosing a subset of window sizes to perform a ncRNA screen. Tradeoffs can be made between the percent of ncRNA detected and the computational investment required to perform the analysis.
A second experiment was performed to further explore the question of optimal values for step size and window delta. Ten tRNA from the Rfam database [50] were embedded at random locations within 300 nt background sequences (Table S5). The selected tRNA ranged in length from 68 nt to 91 nt and generated large negative Z-scores (<−4.0) when evaluated in isolation. The background sequences used were mRNA transcripts that had no significant Z-score along their length. A Z-score was calculated at each position along the total sequence (step size = 1) for each window sizes from 60 to 95 nt (window delta = 1). In most cases it was possible to detect the tRNA in the embedded sequences using a step size of 5 and a window delta of 5 (Figure 1). However, in some cases the window size and window delta needed to be smaller than this to be certain of finding the transcript (Figure 2).
Based on the above results, we chose to use a step size of 5 nt and a window delta of 5 nt for the remainder of our analysis. This provided a high probability of detecting most ncRNA while keeping computational time manageable.
The ncRNA prediction method was applied to intergenic regions of S. cerevisiae chromosome VI using window sizes from 75 to 200 nt, a window delta size of 5 nt, and a step size of 5 nt. The UTRs of most genes in the S. cerevisiae genome are unknown so the term intergenic used here refers to the distance between ORFs of adjacent annotated genes. Genes classified as dubious in SGD [45] were ignored. The UTRs of the flanking genes are thus included in the intergenic region, and those containing structure [51] may be detected. The limited data available on S. cerevisiae 5′ and 3′ UTRs shows that most UTRs are short (3′ UTR median length 91 nt, 5′ UTR median length 68 nt) [12],[13], suggesting that most of the structural signals detected should come from independent ncRNA rather than UTRs. Only intergenic regions greater than 90 nt in length were evaluated.
Forward and reverse DNA strands were evaluated independently since the GU pairing in ncRNA confers different folding potential to the complementary strands. In an attempt to reduce the rate of false positives produced by the screen, the analysis was repeated in syntenic regions of S. bayanus (MCYC623) [52]. For a region to be considered syntenic, it had to have the same flanking genes with the same orientation in both S. bayanus and S. cerevisiae. A total of 66 syntenic regions satisfying these criteria were identified. The percent identity between these regions in S. cerevisiae and S. bayanus varied between 18.0% and 76.5% with an average of 57.0% (Table S6). Predicted structural elements common to both species were taken as ncRNA candidates. There were no constraints placed on the relative position of the structural predictions in syntenic regions, only that they appeared between the same two flanking genes in both species.
There were 23 intergenic regions in S. cerevisiae that produced Z-scores ≤−3.5 and 24 intergenic regions in S. bayanus that produced Z-scores ≤−3.5. Fourteen of these regions were common to both S. cerevisiae and S. bayanus and resulted in a total of 16 high priority candidates (two syntenic regions produced two separate candidates) (Table 5). In many cases, a Z-score below the cutoff criterion was generated from both the Watson and Crick strand. For this reason, experimental testing was performed on both strands independently for all candidates. An example of the Z-score values generated by evaluating the Watson strand for each position in the intergenic region between SEC4 and VTC2 for all window sizes is provided in Table S7. The position of windows producing Z-scores ≤−3.5 within selected intergenic regions are given in Figures S7, S8, S9, and S10.
Northern blots and rapid amplification of cDNA ends (RACE) were used to test the validity of the ncRNA candidates. Since the environmental conditions required for expression of the ncRNA gene candidates were unknown, nine conditions were tested. Conditions were selected that have been shown to generate high overall transcript expression [53],[54]. These nine conditions were: heat shock (25°C to 37°C), diamide treatment, growth in minimal media, saturated growth in minimal media, anaerobic growth, sporulation, schmooing, YPGlycerol (non-fermentable carbon source), and YPD growth. RNA was isolated and northern blotting was performed (see Materials and Methods). Strand specific blotting protocol was used for the northern blot analysis to identify the transcribed strand and to help rule out DNA contamination. Northern blotting confirmed expression of transcripts between SEC4 and VTC2 (RUF20) on the Crick strand and between YFL051C and ALR2 on the Watson strand (Figure S11). The ACT1-YPT1 transcript showed strong expression on the Crick strand under all conditions but later proved to be part of the ACT1 5′ UTR (data not shown).
Rapid amplification of cDNA ends (RACE) was used to measure the 5′ or 3′ end of flanking genes as well as map candidate gene ends (Table 5, Table S8, Table S9). The cDNA was generated using a poly-T primer from RNA collected from anaerobic or heat shock conditions (see Materials and Methods). The RACE analysis proved considerably more sensitive than northern blotting.
In addition to this experimental data, several publicly available data sets were evaluated for their value in substantiating these ncRNA predictions. Tiling array data [11],[12] has been used by several investigators to substantiate computational ncRNA predictions. However, we found this data quite noisy and difficult to interpret with a high degree of confidence. It also remains a point of debate whether all of the transcription measured by microarray tiling experiments represents true functional transcripts or whether some of it represents spurious transcription or experimental artifact [3], [4], [9], [55]–[58]. The sequenced cDNA library data appears to be more useful in verification of ncRNA predictions [13]. The data included information on transcript ends and as such was likely to derive from a functional transcript. A summary of all the experimental data is provided in Table 6.
The candidates in Table 6 are listed in order of increasing experimental support. The top four ncRNA candidates have been assigned names RUF20 (RNA of unknown function) to RUF23 (Figure 3). The RUF name was chosen to follow the naming convention established by previous investigators [34]. These transcripts do not appear to be snoRNA or to encode an ORF (see Materials and Methods). One of the candidates, RUF22, overlaps with an autonomously replicating sequence, ARS607. One other ncRNA candidate, IES1-YFL012W, partially overlaps (120 bp) with the dubious ORF YFL012W-A which is on the opposite strand (Watson). This dubious gene also partially overlaps (120 bp) the IES1 gene. According to SGD, this dubious ORF is unlikely to encode a protein based on available experimental and comparative sequence data [45].
It is reasonable to question whether our computational screen provided an improved ability to identify ncRNA relative to simple random experimental searches. Previous investigators have shown that randomly probing intergenic regions of the S. cerevisiae is unlikely to reveal ncRNA. In the work by McCutcheon & Eddy, 20 intergenic regions were chosen randomly and probed by northern blot [34]. None of these regions produced a transcript. Olivas, Muhlrand and Parker also provided evidence that probing intergenic regions is unlikely to produce a transcript even though they were conducting a directed search for ncRNA [33]. They performed two different screens in an effort to discover ncRNA. In one case, they used a computational approach to identify 10 locations in the genome that contained potential RNA polymerase III binding motifs. When they probed the 10 regions, only one was found to express a transcript. In their second screen, they identified regions within the genome with large gaps between genes. They expected these regions to contain ncRNA transcripts because the high density of genes in the Saccharomyces genome suggested that any large gaps were likely to be occupied by unannotated genes. Probing 59 such regions revealed 15 potential transcripts. It is clear that even probing regions expected to contain ncRNA transcripts is often unsuccessful. Our experimental screen of 16 candidates produced 4 ncRNAs with strong support, 2 potential ncRNA with weaker support, and 6 UTRs likely to contain structure (Table 6). Thus, it appears that our computational method improves ncRNA identification over simple random searches.
To further validate the SEC4-VTC2 ncRNA candidate, RACE was performed in syntenic regions of S. bayanus and the more distantly related hemiascomycete species Ashbya gossypii. This species diverged from S. cerevisiae prior to the S. cerevisiae whole genome duplication. However, A. gossypii still retains many syntenic regions with S. cerevisiae and, in the case of the SEC4-VTC2 gene candidate, gene order and orientation are preserved. RACE products were obtained from both S. bayanus and A. gossypii (Figure 4). The fact that the transcript is preserved over such a large evolutionary distance provides strong evidence that this is a bona fide ncRNA gene.
A computational screen for structural ncRNA in S. cerevisiae was performed using thermodynamic stability to discriminate structural ncRNA from background sequence. The method was tested on positive and negative control sets to determine its effectiveness for identifying known ncRNA and to develop optimal search parameters. These parameters were determined to be a Z-score <−3.5, window sizes 75 nt to 200 nt, step size of 5 nt, and window delta of 5 nt. The parameters were then used to screen for novel ncRNA in the intergenic regions of S. cerevisiae chromosome VI. To reduce the number of false positive predictions, an independent analysis was performed on syntenic regions of S. bayanus. The set of predictions found in common in both species were subjected to further experimental verification. Like all computational ncRNA gene discovery approaches currently available, our method can only provide guidance on regions likely to contain structural elements. It cannot predict the exact location of the ncRNA gene or its precise ends. These must be determined experimentally.
Northern blots, rapid amplification of cDNA ends (RACE), and publicly available cDNA library data were used to test the predictions. Each of these methods was selected for specific reasons. The strength of northern blot analysis is that it does not rely on transcript amplification and hence avoids artifacts that can result from an amplification step. However, it is not as sensitive as other methods and this can be a significant limitation when testing for ncRNA that may be expressed at low levels. RACE provides greater sensitivity than northern blot analysis but may be subject to amplification artifacts. The potential for artifacts is reduced because the 5′ and 3′ ends of the transcript are captured. The presence of a cap and poly-A tail provides strong evidence that the transcript has been processed by the cellular machinery and is a legitimate functional transcript. This makes the approach superior to methods such as tiling arrays that provide information on transcription but for which it is difficult to distinguish transcriptional noise from genuine transcripts. The publicly available cDNA data used here also has the advantage of capturing the transcript 5′ and 3′ ends, providing strong evidence for a legitimate, processed transcript.
The initial computational screen presented here produced sixteen ncRNA gene candidates on chromosome VI of S. cerevisiae. Four candidates are well supported by experimental data and have been given the names RUF20 to RUF23 (Table 5). The RUF20 candidate is also expressed in S. bayanus and in the more distantly related species A. gossypii (Figure 4). All of the transcripts were evaluated for the possibility that they might be snoRNA or encode a protein but this was shown to be unlikely (see Materials and Methods). Two additional candidates are also supported by experimental evidence but further experimental testing is needed to confirm their legitimacy. Six of the candidates were found to be part of the 5′ or 3′ untranslated regions (UTRs) of annotated protein-coding genes. These structures are interesting because they may play a functional role in the UTRs of these genes (Table 5). Additional experimental analysis will be needed to determine the function of the structures as well as the function of the four new ncRNA, RUF20 to RUF23.
There are several possible explanations why experimental data could not be obtained to support three of the ncRNA predictions. These predictions may represent false positives, they may not be expressed under the conditions tested, or they may be expressed at such a low level that they could not be detected. It has been shown that transcript abundance in yeast varies over six orders of magnitude and that some important transcription factors are expressed at levels as low as one transcript per thousand cells [59]. It is also possible that these transcripts are not transcribed by RNA polymerase II, the method used in this study to generate cDNA is dependent on a poly-A tail in the RNA transcript. If the ncRNA candidates are transcribed by polymerase I or III, they would likely not be captured in the cDNA library.
It should be noted that there were three genes in the positive control set (Table 2) that did not generate a Z-score <−3.5 (snR76, SER3, RNA170). It is questionable whether these genes actually contain significant structural elements. One of them, snR76, is a C/D box snoRNA and data from other investigators [19] shows that structural features are only present in a subset of these genes. It is not surprising that this category of ncRNA was not easily detected in this screen based on structural thermodynamic stability. It is clear that some classes of ncRNA will not be identified very well in structural screens. The other two genes in the positive control set were RNA170 (unknown function) and SER3. The SER3 gene suppresses expression of its neighboring gene, SRG1, by blocking access to the SRG1 promoter region via its transcription. SER3 and RNA170 are unlikely to contain significant structural features so the fact that they did not generate Z-scores less than −3.5 tends to validate the method.
Two previous investigators have performed computational genome-wide screens for ncRNA in S. cerevisiae. McCutchen and Eddy, 2003 used the QRNA program to search for structural elements based on observed compensatory changes in pair-wise alignments of S. cerevisiae species. A fixed window size of 150 nts and a step size of 50 nt were used to perform the analysis. Two structural ncRNA candidates were found on chromosome VI. One prediction, between RIM15 and HAC1 (74738–74738), was near one of the candidates predicted in this study between the same genes (74926–75006). They were unable to obtain sufficient experimental support for expression of this transcript. This is consistent with our experimental results as well. The second McCutchen and Eddy prediction, between SMC1 and BLM10, did not correspond to any predictions generated in this study. They obtained northern blot and RACE data to support expression of this second predicted gene.
A second screen for ncRNA was performed by Steigele et al using the RNAZ program [35]. This program searches for compensatory changes in multiple sequence alignments as well as for thermodynamic stability cues indicative of structural elements. The relative contribution of these two factors in the prediction is not specified. A fixed window size of 120 nt and step size of 40 nt was used to perform the analysis. They reported a sensitivity (true positives/total) for identifying snoRNA of 47% (pooling H/ACA box and C/D box snoRNA), sensitivity for identifying snRNA of 66%, and a sensitivity of 72% for tRNA. The screen generated a total of 18 novel intergenic structural predictions on chromosome VI. Of these, 8 were predicted to be on the Crick strand and 8 on the Watson strand. Five of these intergenic regions were shared by our predictions (YFL051C-ALR2, ACT1-YPT1, TUB2-RPO41, GYP8-STE2 and YFR017C-YFR018C). All 5 of the Steigele et al predictions were on the Watson strand in these regions. Two of the predictions overlapped with our predictions (ACT1-YPT1 and YFR017C-YFR018C).
Our experimental data suggested that the YFL051C-ALR2 region is transcriptionally complex and is likely to produce more than a single transcript. This could account for the fact that both studies predicted structural elements in this region. Our RACE analysis of the ACT1-YPT1 region showed that the predicted structural element was contained within the ACT1 UTR on the Crick strand. The Steigele et al prediction overlaps within the ACT1 UTR but is predicted to be on the opposite strand (Watson). For the TUB2-RPO41 region, we experimentally confirmed a transcript on the Crick strand encompassing our predictions. This transcript overlaps with the Steigele et al prediction but is again on the opposite strand (Watson). Our GYP8-STE2 prediction proved to be part of the GYP8 5′ UTR on the Crick strand. The Steigele et al prediction in this region was on the Watson strand and is beyond the region we measured for the GYP8 UTR (although we were unable to map the end of this 5′ UTR). In the YFR017C-YFR018C region, we obtained RACE results that mapped our prediction to the Crick strand as part of the YFR018C 3′ UTR. The Steigele et al prediction, which largely overlaps our prediction, was for a gene on the Watson strand. Hence, while our predictions and those of Steigele et al are close to one another or overlapping in five regions, in all five cases they are on opposite strands.
It is interesting that there is no overlap between the QRNA and the RNAZ predictions of chromosome VI since both programs consider compensatory changes within alignments to identify structural elements. The reason for this is unclear.
There are two primary differences between the search for ncRNA presented here and the work of previous investigators. First, this method does not require sequence alignments in the analysis. Instead, it relies entirely on thermodynamic stability in unaligned syntenic regions of related species to predict ncRNA structure. The approach is capable of finding ncRNA that have moved out of register within syntenic regions and can be applied in situations where accurate alignments may be difficult to obtain.
The second difference in this work is its examination of the impact of various window sizes and step sizes on ncRNA detection. The analysis shows that small step sizes are necessary to ensure that most ncRNA are identified. It also shows that more than one window size is needed when screening for ncRNA. Some ncRNA are detected only when using short window sizes while others are detected when using only long window sizes (Table 7). Limiting the search to a single window size, as has traditionally been done, is likely to bias the screen toward a subset of ncRNA for which that window size is optimal.
The need for multiple window sizes and step sizes in the screening algorithm increases the computational investment necessary to perform the analysis. However, with the rapid increase in computer performance and the availability of computer clusters, these computations are not unreasonable. The increased computational investment will be rewarded by increased sensitivity.
Our analysis suggests that a few carefully selected window sizes will be nearly as effective at detecting ncRNA as the entire set between 75 nt and 200 nt (total of 26 window sizes). For example, when we used the entire set of window sizes from 75 nt to 200 nt, we detected 22 of the 29 known H/ACA snoRNA within embedded sequences (Table 7). If we had used only 4 window sizes (80 nt, 120 nt, 160 nt, 200 nt), we would have succeeded in identifying 90% of these H/ACA box snoRNA (20 of the 22) while reducing computational requirements by approximately 85% (4 of 26 window sizes). If these four window sizes were used with a step size of 25 nt, 77% (17 of 22) of the H/ACA box snoRNA would be detected (Table S10). This becomes 64% (14 of 22) if the step size is increased to 50 nt (Table S11).
Tradeoffs between sensitivity and computational requirements should be evaluated when performing computational screens. We recommend using a range of four window sizes when screening for ncRNA in a genome (one short, one long, and two intermediate values appears to be optimal). Our results suggest that the values of 80, 120, 160 and 200 should provide good results. A step size between 5 and 10 should also provide a good screen. These parameters should provide good ncRNA detection while keeping computational time manageable. The development of an efficient computational algorithm implementing the methodology presented here would also significantly reduce computational run time.
This screen used a simple cutoff Z-score value (≤−3.5) to discriminate ncRNA. The sensitivity of the screen could probably be improved if a more sophisticated cutoff criteria were developed in which the Z-score cutoff was a function of window size. The number of aberrant negative Z-scores dropped as a function of window length in the negative control sets demonstrating that the likelihood of producing large negative Z-score drops with increasing window length. Developing a Z-score cut-off value as a function of window length would probably improve the sensitivity of the screen at longer window sizes.
This work demonstrates that structural thermodynamic stability is an effective tool for predicting ncRNA genes. As examples of ncRNA are accumulated through computational screens such as this, it may become possible to determine ncRNA key features and gain insight into their biological function. Computational methods can complement experimental approaches in the effort to gain a deeper understanding of these genes.
S288C was used for all growth conditions except for sporulation (SK1) and pheromone treatment (BY4741).
Cells grown continuously at 25°C were collected by centrifugation, resuspended in an equal volume of 37°C medium, and returned to 37°C for an additional 20 minutes. The RNA was then isolated as described below. RNA was collected after twenty minutes as it has been shown to be the point of maximum RNA expression [53].
Pheromone treatment stimulates yeast cells to increase the expression of mating genes, arrest cell division in the G1 phase, and form polarizing mating projections directed toward the pheromone source [60]. Overnight yeast cultures grown in YPD at 30°C were treated with 50 nM α-factor (GenScript Corporation). Cells were examined under a microscope to ensure schmooing was induced. Total RNA was extracted 75 minutes after pheromone treatment.
A strong cellular response to diamide treatment has been shown previously [53]. It resembles a composite response to heat shock, H2O2 treatment and menadione treatment. It induces cellular redox genes and genes associated with defense against reactive oxygen species. Diamide (Research Organics) was added to cell cultures grown in YPD at 30°C in late log phase to a final concentration of 1.5 mM. Cells were returned to 30°C for growth for 30 minutes. RNA was then isolated as described above.
This growth condition induces expression of genes involved in meiosis and spore morphogenesis. SK1 yeast cells were sporulated in a synchronous meiosis as described previously [61]. Briefly, yeast cultures were pre-grown in YPD to saturation at 30°C, diluted 200-fold into 100 ml of YPA (1% yeast extract, 2% Bacto-peptone, 2% acetate), and grown to early stationary phase (about 5×107 cells/ml). Cells were then washed with water and resuspended into 100 ml of SPM (sporulation media consisting of 0.3% potassium acetate and 0.02% raffinose). Sporulation was carried out at 30°C under conditions that allowed good aeration. Expression data suggested that metabolic, early, middle and late genes were active 11 hours after transfer to sporulation media so total RNA was collected at this time point [54]. Cells were inspected under a microscope to ensure that sporulation of at least some of the cells had taken place. RNA was then isolated as described below.
S288C yeast cells were grown for approximately 55 hours in 100 ml of minimal media (YNB) in an anaerobic chamber using an anaerobic gas generating system (Mitsubishi Gas Chemical Company, Inc.). Four minimal media plates were also streaked with S288C and grown anaerobically for the same time period. The anaerobic chamber was then opened and the cells growing on the plates were added to the cells in the liquid growth by washing. Total RNA from all of the cells was isolated immediately as described below.
Saturated growth has been shown to activate gene expression, presumably allowing the cells to adapt to nutrient depleted conditions [53]. S288C cultures were grown to saturation (OD 3) in minimal media (YNB). They were also grown to logarithmic phase in rich media (YPD) and on a nonfermentable carbon source, YPGlycerol. All three cultures were grown at 30°C and aerated by shaking at 250–300 rpm.
A phenol-chloroform extraction protocol was used as described previously [62] to extract total RNA from S. cerevisiae, S. bayanus and A. gossypii. All glassware used in the procedure was baked for 4 hours to destroy RNase activity. Reusable plasticware and laboratory bench surfaces were treated with RNAzap (Biohit, Inc.). RNAse-free water was prepared by treating with Diethyl pyrocarbonate for one hour and then autoclaving. Cells were harvested from 50 ml cultures at an OD600 of 1–3 (1 OD = 3×107 cells/ml) unless otherwise specified. The cells were collected via centrifugation (except A. gossypii cells which were collected using a vacuum filter). The cell wall was disrupted by vortexing at high speed with acid-washed glass beads in a solution containing guanidine thiocyanate. Total RNA was purified using multiple washes with a combination of hot phenol and chloroform.
The total RNA was treated with TURBO DNase (Ambion) and incubated at 37°C for 30 minutes prior to using for RACE or northern applications. The DNase activity was destroyed by heating to 70°C for 5 minutes per the standard protocol. RNA quality was assessed by measuring absorbance at a wavelength of 260 nm on a NanoDrop (ND-1000) spectrometer.
A 6%, 7 M urea, 1× TBE denaturing polyacrylamide gel was used to make a northern blot with total RNA as described previously [63]. Total RNA was treated with TURBO DNase (Ambion) and incubated at 37°C for 30 minutes prior to gel loading to ensure that no DNA was present. It was loaded onto the gel and run at 150 V for 3 hours. The total RNA was transferred from the gel to a nylon membrane using the OWL Scientific Panther Semi-Dry Electroblotter (Model # HEP-1) with a current of 200 milliamperes for a period of 1 hr. The RNA was fixed to the blot with UV crosslinking. Radioactive strand-specific probes were produced using the Strip-EZ system with α-P32 dATP (Ambion). Each probe was used to on a separate northern blot. This provided a check that the observed signal derived from only a single strand and was not the result of DNA contamination (which would produce signal from both strands). The probes were detected by exposing the blot to BioMax XAR film (Kodak) at −80°C 24–48 hours.
The SMART RACE cDNA Amplification Kit (Clontech) was used to map transcript ends. Total RNA was isolated from S288C under two different growth conditions: anaerobic growth and heat shock from 25°C to 37°C. It was treated with TURBO DNase (Ambion) prior to making the cDNA.
To obtain RACE products for the ncRNA candidates, a RACE reaction and nested reaction were performed for both the Watson and Crick strand since it was uncertain which strand the transcript might be generated from. The temperature profiles developed to optimize the reaction are given in Appendix A. A hot start approach was used to minimize extraneous amplification by allowing the reaction tubes to reach a temperature of 94°C for 1 minute before adding the Ex Taq (Takara) polymerase. The RACE products were electrophoresed on a 1% agarose gel and the resulting bands were cut out of the gel. They were purified using one of two methods. The first was to use the QIAquick Gel Extraction Kit (QIAGEN), according to the standard protocol. Alternatively, the gel slices were frozen at −20°C and then spun on a tabletop centrifuge at 1400 rpm for 20 minutes. The sample was then drawn from the top of the resulting liquid. This proved a quick and reliable method for obtaining purified product. The purified RACE products were sequenced using standard BigDye chemistry, version 1.3, according to standard protocols (Applied Biosystems).
RACE primers were designed according to guidelines provided in the SMART RACE kit. They were 20–28 nt in length, had a GC content between 50–70%, a melting temperature ≥72°C, and had no more than 2 C's or G's in the last 5 nucleotides of the oligonucleotide. Each primer was confirmed to be unique in the genome using the “fuzznuc” program that is part of the EMBOSS utilities [64].
The Z-score compares the minimum folding energy (MFE) of a sequence, x, to the distribution of MFE generated by permuted versions of x having the same di-nucleotide composition. The di-nucleotide composition must be preserved because of the importance of stacked base-pairs in the MFE calculation [65]. The MFE of each sequence, x, was calculated using the RNAfold program [44]. Each sequence was then shuffled 500 times using the shuffle program provided in Sean Eddy's squid utilities [47] and a mean and standard deviation were calculated for the resulting distribution. The Z-score was then calculated using the equationwhere <·> and σ (·) denote the mean and the standard deviation of the MFEs of the sequences in xshuffled(x). Hence, the Z-score represents the number of standard deviations that the sequence x deviates from the mean MFE of the shuffled sequences.
The genome sequence data used for ncRNA prediction and subsequent evaluation of open reading frame coding potential is listed in Table 8.
It was important to investigate the possibility that the ncRNA candidates might be protein-coding genes. Comparative genomics was used to investigate this possibility for the four ncRNA gene candidates RUF20, RUF21, RUF22 and RUF23 (Table 5). This approach has been applied by other investigators with a high degree of success [66].
There are no conserved ORFs within the three candidates RUF21, RUF22 and RUF23 among the closely related species S. cerevisiae, S. paradoxus, and S. bayanus (sensu stricto). These transcripts are thus unlikely to be protein-coding genes.
The RUF20 candidate contains one ORF consisting of 8 amino acids conserved among S. kudriavzevii, S. bayanus, S. paradoxus, and S. mikatae (sensu stricto). However, the pattern of substitution among these species is not consistent with synonymous amino acid substitutions as would be expected for a protein-coding gene (two mutations are in the 1st codon position, one mutation is in the 3rd position). The ORF is not conserved in Candida glabrata or A. gossypii. This is significant because our RACE data confirmed expression of the transcript in A. gossypii. In addition, the 8 amino acid ORF does not contain any splice signals suggesting that it is spliced to another exon. While a number of short ORFs have been identified in yeast [67], there are none known to be as short as 8 amino acids. Taken together, this data strongly suggests that the short RUF20 ORF conserved among the sensu stricto does not encode a protein.
The SnoScan [19] and SnoGPS [20] programs were used to test if the ncRNA candidates were likely to be snoRNA. The SnoScan program searches for features characteristic of C/D box snoRNA. None of the RUF20 to RUF23 candidate genes have features characteristic of C/D box snoRNA according to the program. The SnoGPS program searches for features characteristic of H/ACA box snoRNA. According to the program, RUF23 is unlikely to be a H/ACA box snoRNA. The program found some features of H/ACA box snoRNA evident in the RUF20, RUF21 and RUF22 candidates, although their overall bit score was low (28.4, 29.3, and 29.9 respectively). A bit score value of 36 is recommended as the cutoff value when searching for new H/ACA snoRNA. To further evaluate the possibility that RUF20, RUF21 and RUF22 might to be H/ACA snoRNA, sequence from two closely related species was used. The homologous gene sequences from S. paradoxus and S. bayanus were evaluated using the snoGPS program. The RUF20 candidates in these species were found to be unlikely to be a H/ACA snoRNA by the program. The RUF21 and RUF22 genes did generate possible H/ACA snoRNA candidates in the related species but there was no common rRNA target identified among the homologous sequences. Hence, the candidates appear to be unlikely H/ACA snoRNA genes.
|
10.1371/journal.pcbi.1003227 | GINI: From ISH Images to Gene Interaction Networks | Accurate inference of molecular and functional interactions among genes, especially in multicellular organisms such as Drosophila, often requires statistical analysis of correlations not only between the magnitudes of gene expressions, but also between their temporal-spatial patterns. The ISH (in-situ-hybridization)-based gene expression micro-imaging technology offers an effective approach to perform large-scale spatial-temporal profiling of whole-body mRNA abundance. However, analytical tools for discovering gene interactions from such data remain an open challenge due to various reasons, including difficulties in extracting canonical representations of gene activities from images, and in inference of statistically meaningful networks from such representations. In this paper, we present GINI, a machine learning system for inferring gene interaction networks from Drosophila embryonic ISH images. GINI builds on a computer-vision-inspired vector-space representation of the spatial pattern of gene expression in ISH images, enabled by our recently developed system; and a new multi-instance-kernel algorithm that learns a sparse Markov network model, in which, every gene (i.e., node) in the network is represented by a vector-valued spatial pattern rather than a scalar-valued gene intensity as in conventional approaches such as a Gaussian graphical model. By capturing the notion of spatial similarity of gene expression, and at the same time properly taking into account the presence of multiple images per gene via multi-instance kernels, GINI is well-positioned to infer statistically sound, and biologically meaningful gene interaction networks from image data. Using both synthetic data and a small manually curated data set, we demonstrate the effectiveness of our approach in network building. Furthermore, we report results on a large publicly available collection of Drosophila embryonic ISH images from the Berkeley Drosophila Genome Project, where GINI makes novel and interesting predictions of gene interactions. Software for GINI is available at http://sailing.cs.cmu.edu/Drosophila_ISH_images/
| As high-throughput technologies for molecular abundance profiling are becoming more inexpensive and accessible, computational inference of gene interaction networks from such data based on well-founded statistical principles is imperative to advance the understanding of regulatory mechanisms in various biological systems. Reverse engineering of gene networks has traditionally relied on analysis of whole-genome microarray data; here we present a new method, GINI, to infer gene networks from ISH images, thereby enabling exploration of spatial characteristics of gene expressions for network inference. Our method generates a Markov network, which encapsulates globally meaningful statistical-dependencies from vector-valued gene spatial patterns. In other words, we advance the state-of-art in both the usage of richer forms of expression data, and the employment of principled statistical methodology for sound network inference on such new form of data. Our results show that analyzing the spatial distribution of gene expression enables us to capture information not available from microarray data. Such an analysis is especially important in analyzing genes involved in embryonic development of Drosophila to reveal specific spatial patterning that determines the development of the 14 segments of the adult fly.
| In multicellular organisms such as the metazoans, many important biological processes such as development and differentiation depend fundamentally on the spatial and temporal control of gene expression [1], [2]. To date, the molecular basis and regulatory circuitry underlying metazoan gene regulation remains largely unknown. Numerous statistical or algorithmic approaches have been attempted to infer “networks” of regulatory elements from high-throughput experimental data, based on various computational techniques like Bayesian networks [3]–[5], undirected Gaussian graphical models [6], [7], graph mining [8], ordinary differential equations [9], partial correlations [10], and others. Comparisons of different methods used for reverse engineering gene networks have been performed [11], [12], and predictions made by automatically learned gene networks have been experimentally validated [13], [14], thus increasing the credibility of such approaches.
This progress notwithstanding, a key deficiency of existing approaches is that they rely almost exclusively on univariate characteristics of gene states, such as a continuous-valued abundance measurement from microarray, or a binary on/off status derived from discretization of microarray data. However, microarray profiling of mRNA abundance can often be ill-suited for spatial-temporal analysis of gene expressions in multicellular organisms such as Drosophila, or in tissues/organs with natural or pathological progressions, because it captures only the “average” pattern of a sample. For any sample of interesting heterogeneous cell populations, the averaging operation would cause severe information loss and inaccuracy in downstream analysis (see Figure 1 in [15] for an intuitive illustration of how two genes with completely different spatial patterns over time yield near identical “average” temporal patterns.)
Recent advancements in image-based genome-scale profiling technology such as whole-body mRNA abundance micro-imaging via in situ hybridization (ISH) have begun to reveal a more holistic view of the activities and functions of genes in rich spatial-temporal contexts. ISH has been used to characterize whole genome expression patterns for different species such as Drosophila embryos [16], [17], C. elegans [18], and adult mouse brain [19], and at smaller scales for Arabidopsis flowers [20], testicular germ cell tumors [21], and others. The availability of this form of gene expression data calls for development of next-generation image analysis systems to facilitate not only efficient pattern mining such as image clustering or retrieval, but also in-depth reasoning of complex spatial-temporal relationships between gene expression patterns, which will be essential for functional genomics and regulatory network inference in higher organisms. In this paper, we focus on a particularly interesting, but previously unaddressed challenge along this direction: inferring a statistically sound gene network from gene expression micro-imaging data, in the same sense of inferring a gene network from microarray data as widely studied in the literature. Analyzing ISH data allows us to infer a network by computing similarities in the spatial distributions of gene expressions in Drosophila embryo. Another important source of information is the temporal changes of the spatial distributions of genes, which could reveal how a gene regulation network evolves over time during dynamic biological processes such as embryogenesis [22]. We will defer the spatio-temporal network building based on time series of ISH data for future work as it requires the technique developed in this paper as a building block.
A major motivation of our work is the extensive imagery documentation of all the genes expressed during Drosophila embryogenesis via ISH imaging by the Berkeley Drosophila Genome Project (BDGP) [16]. BDGP is an ongoing effort to determine gene expression patterns during embryogenesis for Drosophila genes. In February 2013, the data contained more than 110,000 ISH images capturing the expression pattern of 7516 genes. Each image is annotated with time information, indicating the development of the embryo in six development stage ranges. Each image documents the gene expression pattern of a single gene in an embryo. Most images have a single embryo, however some images capture partial views of the embryo, others have overlapping or touching embryos. This is an extremely interesting but difficult dataset that reveals unprecedented details of gene activities during metazoan embryogenesis, but at the same time posts large unanswered challenges on methodologies for systematic and principled analysis. Specifically, we recognized the following main challenges that are unique to micro-imaging data versus the classical microarray data, which must be properly addressed before a genome-scale gene network can be derived from such data.
There has been some earlier work on automatic annotation of ISH images with annotation terms [24], [25], clustering of gene expressions [17], determination of the development stage of embryos [26], etc., some of which have been applied on the BDGP dataset. In this paper, we propose a machine learning system to infer gene interaction networks from spatial similarity of gene expressions captured via ISH images. The system is called GINI, for Gene Interaction Network from Images. With such a system, we were able to address satisfactorily the challenges mentioned above, and systematically performed a genome-scale network learning and analysis on the BDGP dataset.
GINI first extracts the gene expression pattern from each image using a computer version driven image analysis pipeline [15]. These expression patterns are spatially aligned and normalized to enable spatial comparison of gene expression across multiple images. Next, the expression patterns are represented by suitable standardized features through a process called “triangulation”, followed by feature normalization and selection. Since each gene may have a different number of images in the data, each gene can now be represented by a “bag” or a set of feature vectors - one feature vector per image. Thus, our task is to estimate the gene network, given bags of images per gene (Figure 1). We cast the problem of estimating a gene interaction network as the task of estimating the graph structure of a Markov random field (MRF) over the genes. The underlying graph encodes conditional independence assumptions between the genes, that is, two genes are said to not interact in the network if their gene expressions are conditionally independent of each other, conditioned on the expression of all other genes in the network. We employ multi-instance kernel technique using different order statistics to compute similarity between bags of images. We then estimate a sparse network of gene interactions by modeling the data as a multi-variate multi-attribute Gaussian, and estimating the sparse inverse covariance matrix of the model. A schematic diagram of the system pipeline can be seen in Figure 2.
GINI is a bioimage informatics system based on a computer vision pipeline for ISH micro-image processing and a statistical learning algorithm for network inference. The main contributions of this work are summarized below.
First, the image analytic pipeline used by GINI offers a rigorous and universal approach to extract a standardized representation of spatial patterns of gene expressions. Comparing to the popular SIFT features [24], which is based on detecting interest points with heavy assumptions on object shape, texture, and other physical properties originally meant for natural objects, our approach is more suitable for ISH staining in Drosophila embryos which do not resemble natural objects and require preservation of overall spatial shape and overall intensity information in a canonical way for intra-gene normalization and inter-gene comparison.
Second, GINI infers a network that enjoys the Markov network property: it gives globally consistent conditional-independency interpretation for every edge, and therefore is biologically more meaningful. It is known that marginal correlation (as often used in estimating an ad hoc network), which is computed for every gene-pair in isolation (i.e., ignoring all other genes in the system), confounds direct and indirect dependencies, and could result in a clique-like dense graph or subgraph among genes that are not directly dependent, but have a long-distance interaction. Studying conditional independencies in a network allows us to predict interactions between a pair of genes in the context of other genes, allowing a distinction to be made between direct and indirect relationships between the genes, and reducing false positives.
Third, our formulation based on Gaussian Markov random field and multi-instance kernel for the GINI network is convex, hence the globally optimal estimator of the network is computed, no approximations are involved. Furthermore, under suitable conditions, our graphical model learning algorithm is sparsistent, i.e. as the amount of available data increases, the algorithm is statistically guaranteed to predict the correct interactions between the genes. While Bach et.al. [27] have previously proposed learning the structure of graphical models from data using Mercer kernels, their approach is based on a non-convex local greedy search to find edges in the graph. Our approach represents the first work that uses Mercer kernels and Gaussian Graphical Models to predict kernelized graphical models using a convex formulation.
Finally, with the GINI system, we were able to systematically perform a genome-scale network learning and analysis of the genes expressed during 2 time points of Drosophila embryogenesis recorded by ISH imaging from BDGP [16]. In both time points, we find that the GINI networks are modular and scale free, which are properties predicted to hold true in gene interaction networks. Further, different regions of the networks are enriched for spatial annotations, thus GINI is able to cluster spatially similar genes. The hubs of the networks, i.e., the genes with the largest number of predicted interactions are functionally enriched for important cellular functions. We demonstrate that the networks predicted by analyzing microarray data does not have either spatial or functional enrichment, thus these results could not have been obtained by analyzing microarray data.
To the best of our knowledge, GINI represents one of the first efforts to reverse engineering gene networks from ISH image data. In both extensive simulation studies and empirical biological analysis, we demonstrate the effectiveness of GINI in predicting networks, and show that the statistical assumptions behind GINI are reasonable, and the biological analysis enabled by GINI merits close examination and further exploration.
We begin by introducing the key algorithmic innovations needed to compute the gene network from the ISH images, assuming that each gene has a bag of images, with the images processed to be represented by informative and canonical feature vectors. This is followed by a discussion on the image processing procedures needed to extract informative features from the images.
We first show how GINI estimates a gene network, when each gene has only one image. The next subsection extends the GINI algorithm to deal with multiple images per gene.
Let denote the set of genes being studied, so that is the gene, where , and is the number of features extracted per image. Each feature represents the gene expression in a spatial location of the embryo.
Note that algorithms that analyze microarray data typically treat samples drawn from different time points as independent samples [28], even though expressions of the same gene across time is expected to be auto correlated. We similarly assume that the different spatial features are independent of each other. The spatial independence assumption has also been implicitly made by [29], [30] while modeling transcription networks in Drosophila embryos. In the results section, we use simulated data to demonstrate that this assumption does not affect the accuracy of the algorithm significantly.
By modeling the gene interactions as invariant across the spatial locations in the embryo, we can assume that each feature is independently and identically drawn (i.i.d.) from the same distribution. Inferring gene interactions is then equivalent to modeling the dependence between the expression values of different genes at the same spatial location. Expression of the genes in each spatial location is assumed to be drawn from some (multi-variate) distribution, independent of all other spatial locations. Each spatial feature () may be modeled as a vector of length , with capturing the expression value of the gene in this location . This gives us independent samples with which the parameters of the underlying distribution may be learned. Formally, let each spatial location be drawn independently from a multi-variate Gaussian , where is the mean vector, and is the positive semi-definite covariance matrix between the genes.
In a multivariate Gaussian distribution, the entry of the inverse covariance matrix is zero if and only if the corresponding genes are conditionally independent given the rest of the graph. Thus, the non-zero entries of the inverse covariance matrix correspond to edges in the corresponding Gaussian Markov random field, giving rise to the gene interaction network. The Gaussian Markov random field is also known as a Gaussian graphical model (GGM) [31]. Since we expect a small number of interactions per gene, the estimated graph must be sparse, i.e. the number of non-zero entries of the inverse covariance matrix must be small.
Thus, the gene interaction network may be estimated by learning a Gaussian distribution from the observed images, such that the inverse covariance matrix is sparse. The mean of the Gaussian is estimated by the observed sample mean,(1)
Then, the inverse covariance matrix can be estimated by minimizing the negative log-likelihood of the data, over all possible positive semi-definite matrices. To enforce sparsity, the norm of , which counts the number of non-zero elements, is added to the negative log likelihood. Since optimizing the norm is non-convex and NP hard, the norm is used as a convex relaxation to the norm. The norm of a matrix is the sum of the absolute values of the elements of the matrix, and also enforces sparsity in the solution. Adding the norm regularization also ensures that the minimizer of the objective function exists, and is well defined. Thus, our objective function is(2)where is the second moment matrix about the mean(3) is a tuning parameter, by which we determine the strength of the penalty. As we increase the value of , we increase the penalty on the absolute values of , and hence, the graph induced by becomes more sparse. The edges in the graphical model are then estimated as(4)
Multiple images of the same gene at the same time point should have the same gene expression pattern. However, in practice, the expression patterns in different images may differ considerably, for three main reasons.
Firstly, there is a wide interval of time considered as a single time point while collecting such data. For instance, the BDGP data divides embryonic development into 6 time stages. The last stage 13–16 corresponds to development of the embryo 9.3 to 15 hours after fertilization, which represents more than a third of the time taken for embryonic development. Hence, the true gene expression pattern may be dynamic within the time period of a single development stage, and the gene expressions captured for the same gene at the same time may not look similar to each other. Secondly, we might expect that for any organism for which ISH data is collected, there will necessarily be some ambiguity in how the development stage of the organism is labeled by human annotators. Finally, noise in the expression patterns due to excessive staining, lighting conditions and similar other reasons will also be observed. For all of the above reasons, any network-learning algorithm should leverage the existence of multiple images per gene per time point in improving its estimates of gene similarity.
The problem of multiple images per gene is reminiscent of multi-instance learning [35], [36]. Multi-instance learning is a form of supervised learning, where instead of labeling each instance, a bag of instances is labeled. A popular solution to the multi-instance problem is to define a multi-instance kernel, that can compute the similarity between bags of instances. Let be a collection of order statistics of the set , for example, mean, median, minimum, maximum etc. In dimensions, is computed on each dimension independently, to form a vector of order statistics. If we use order statistics, then the length of will be . The similarity between gene with a set of images and gene with images can then be computed as(5)where is an appropriate kernel function between vectors and . Such a kernel is called the statistic kernel.
The choice of the order statistics used in the kernel depends on the data collection procedure of the ISH. One concern in ISH data is that images may be overstained. In such a scenario, the median may be an appropriate choice of order statistic. If over-staining is not a concern, the maximum statistic may be more appropriate to ensure that information about presence of gene expression is not lost.
For the BDGP data, we use the covariance kernel , and the mean statistic . The choice of using a single statistic to represent information from multiple images was due to the presence of noisy images in the data set. Thus,(6)
Thus, our choice of kernel is equivalent to computing the mean similarity of all pairs of images in bags and . This specific kernel is also known as the normalized set kernel, and has been shown to perform very well in multi-instance classification [37].
Any kernel function may be written as the dot product in some higher dimensional feature space, i.e. [38]. Hence, if we assume that the data is drawn from a distribution such that is a zero-mean Gaussian, we can learn the gene interaction network by treating as the sample covariance matrix. Since estimating the inverse covariance matrix by solving equation 2 requires only the sample covariance matrix and not the data itself, we can kernelize it by using the kernel matrix defined in equation 6 as the required sample covariance matrix. Thus, the objective function is(7)which can be solved as discussed in the previous section.
We convert the ISH images into canonical feature vectors suitable for analysis by our algorithm described above in a three-step manner. First, the precise expression pattern found in each image is extracted and aligned spatially to make all images spatially comparable. Next, each image is represented by a feature vector using Delaunay triangulation. Finally, features are normalized and feature selection is performed to extract meaningful features, that can be then used to compute the multi-set kernels to obtain gene similarity and learn the gene network.
Putting everything together, we conclude the method section with a summary of the GINI system for network inference from ISH images. Each ISH image is converted into a standardized expression pattern using , and then triangulated to extract a low-dimensional spatial feature vector. Next, feature values are normalized, uninformative features are removed, and genes with insufficient information available are rejected. Finally, the multi-set kernel is used to compute the similarity between the bags of image vectors available for each gene, and the gene network is estimated using Equation 7. The algorithm is summarized in Algorithm 2.
We first demonstrate that the independence and Gaussian assumptions are reasonable for ISH data, and that GINI explains the ISH data well, with small fitting errors, and no bias in the residues. Next, we show the performance on a small subset of 12 images for 6 genes to verify that the network predicted by GINI is reasonable. We then run GINI on two datasets of ISH images from 2 time points in the BDGP data, and study the networks. We find the networks are modular and scale free as expected. Furthermore, different regions of the networks are enriched for spatial annotations, and the hubs of the networks are functionally enriched for important cellular functions. Finally, we demonstrate that these results could not have been obtained by analyzing microarray data.
GINI assumes that the gene expression in each triangle can be assumed to be independently drawn from a multi-variate Gaussian. However, the true gene expression in adjacent spatial locations is correlated and not independent. To verify that this dependence of adjacent samples does not affect the accuracy of the estimated network, we simulate synthetic data where the underlying network is known, but the data points are not independent of each other, and test whether GINI can recover the correct network in such a scenario. The data samples depend on each other via a parameter that captures degree of dependence between data samples. When , all data samples are drawn i.i.d. from the known distribution. As increases, data samples are drawn from the same distribution, but they depend on the adjacent samples.
For a high-dimensional distribution, it is not feasible to test if the data is truly Gaussian. However, a consequence of Gaussianity is that for each gene, the gene expression can be expressed as a weighted linear sum of the expression values of a few other genes, which form the edges of the network. To test if this assumption holds true in ISH data, for each gene, we fit a linear regression between the gene and its neighbors found by GINI and look at the absolute value of the error i.e. the mean absolute difference between the predicted and the known gene expression. When the maximum expression value is 1, for more than 90% of the genes we looked at, the absolute error was less than 0.02; 99.5% of all genes had absolute error less than 0.05, confirming that the GINI generative model explains the ISH data.
We also confirm that the prediction error is not systematic with respect to the spatial location. For each gene, we compute the prediction error (residue) when the gene is predicted by regressing it on its neighbors. For each spatial location, we plot the mean residue at that location for all genes. As can be seen in Figure 5, there is no systematic bias in the spatial positions that are hardest to predict for any gene.
Before running our algorithm on a large sized dataset, we construct an artificial small data set to verify the results. We input 12 images, shown in Figure 6(a) from 6 genes to the GINI algorithm (each gene has 1–3 images in the data set). With , 4 edges are predicted in the network, shown in Figure 6(b). As can be seen, the three genes hunchback(hb), four-jointed(fj), and Blimp-1, which are expressed in the dorsal, ventral and procephalic ectoderm, are connected in a single cluster. Similarly, the genes organic anion transporting polypeptide 74D(Oatp74D) and bicoid(bcd) are connected by an edge, since both show expression in the foregut and the anterior endoderm. Finally, the expression of sloppy paired-1(slp1) was considered to be sufficiently different from the other genes, hence it is not connected to any other gene in the network.
Thus, the gene interaction network found by GINI can be verified to be reasonable for the above small data set.
We now turn our attention to the ISH images from the Berkeley Drosophila Genome Project data set. We have obtained around 67400 ISH images of 3509 protein-coding genes from the BDGP data released in September 2009, captured at key development stages of embryonic development. Each image captures embryonic gene expression of a single gene using RNA in-situ hybridization. Each image was labeled manually with the age of the embryo, categorized into six distinct embryonic stages : 1–3, 4–6, 7–8, 9–10, 11–12, and 13–16. Genes are also annotated with ontology terms from a controlled vocabulary of around 295 terms, describing the unique embryonic structures in which gene expression is observed during the various stages of embryonic development. analyzes these image automatically, rejecting unsuitable images, to produce 51593 expression patterns of 3347 genes.
As proof of concept, we focus on images viewed from a lateral perspective from two development stage ranges of this data : 9–10 and 13–16. For the stage 9–10, we have 2869 expression patterns of 2609 genes, and for stage 13–16, we have 6350 expression patterns of 3258 genes. We extracted features as described in the methods section. For each development stage, we ran a separate analysis.
Using a value of 0.775 for stage 9–10, we ran GINI and obtained a network having 258 genes, and 516 interactions (edges) between them. For the development stage 13–16, we used , and obtained a network with 1202 genes and 3666 interactions between them. The value was selected for each network by running GINI for 21 values between 0.5 and 1, and picking a value such that the mean-degree for the network is reasonable (approximately 2–3) - see Supplementary Figure S1 for a plot that shows how the number of edges in the network decreases as increases.
Some of the interactions predicted by GINI have already been reported in the literature. For example, in the network for stage 9–10, GINI predicts that DCP-1 (CG5370), an effector caspase which is involved in apoptosis, will interact with the thread gene (CG12284), a known inhibitor of apoptosis protein [43]. GINI also predicts that Snf5- related 1(CG1064) interacts with echinoid (CG12676), both of which are known to be involved in epidermis development, muscle organ development, as well as imaginal disc-derived wing vein morphogenesis. In the 13–16 development network, GINI predicts that the capping protein beta gene (CG17158) interacts with the Glycogen phosphorylase gene (CG7254), and Tpc1 (CG6608) interacts with CG2812, which has been previously reported in [44].
The next five subsections do a detailed analysis of the 2 networks.
A network is said to be scale free if its degree distribution asymptotically follows a power law. That is, the fraction of genes that have at least interactions with other genes is(10)where is the scale free parameter, and is the normalization constant. It has been hypothesized that gene regulatory networks are scale free [10]. We looked at the characteristic of our interaction networks by plotting the number of interactions per gene (Figure 7), and found that the networks found by GINI are scale free. The parameter obtained is 2.3 and 2.5 for the 9–10 and 13–16 networks respectively, which corresponds well to the values observed for a large variety of power law graphs. The scale free nature of the network was found to be independent of the tuning parameter of the algorithm.
Unlike the gene regulatory network obtained for Human-B cells [10], we found that the scale-free nature of the gene network we obtain has a good fit, without observing a deviation from the expected at low connectivity values. However, this could be a side-effect of the larger number of genes they analyzed.
Using spectral clustering, we construct 12 regions or clusters within each network, and visualize the five biggest clusters of each of the networks in Figure 8. All 12 clusters in both networks are very well separated. The ratio of within-cluster edges to total number of edges is 70% and 87% for the 9–10 and 13–16 development stage networks respectively, indicating that the estimated networks are highly modular. From a biological perspective, different parts of gene networks may be responsible for different pathways or biological functional components of the cell, thus modularity is a good prediction for real interaction networks.
Given the scale-free nature of the network, a small number of the genes have a large number of interactions. We analyze the Gene Ontology functions of the genes having the largest number of interactions, i.e. the hubs of the network. The question we wish to address is: if we pick the top 5% of the genes having the maximum connectivity with other genes, what kind of functional enrichment do these genes have? Our background population is of the 2609 and 3258 genes for which we have at least one ISH image describing its expression for the 9–10 and 13–16 stages respectively. We use the hypergeometric test, with Bonferroni correction used to correct for multiple hypothesis tests [45]. As can be seen in Table 1, we observe enrichment of a wide variety of functions that are essential to cell growth and functioning, including metabolic processes, cellular respiration, transport of electrons and ions, protein modification, ribosome biogenesis etc.
Next, we examine a few high-degree hubs in the two networks in detail, along with their neighborhood genes in the networks. Figure 9 shows the hub neighborhood for two genes in the 9–10 development stage network. CG3969 is a Activated Cdc42 kinase-like gene known to be involved in protein phosphorylation [46] and cell death [47], and CG9984 (TH1) is known to be involved in regulation of biosynthetic process [48] and nervous system development [49]. Both genes interact with many genes having functions related to the primary metabolic process, and single-organism cellular process. In stage 13–16, we examine the hub neighborhood of CG5904 and CG6501. The mitochondrial ribosomal protein CG5904 has been previously predicted to be a structural constituent of ribosome [50], and we find that it interacts with many genes involved in the ribosome biogenesis. Gene CG6501 (Ns2) has been previously predicted to be involved in phagocytosis, engulfment [51], and ribosome biogenesis [46]; CG6501's neighborhood has multiple genes that are also involved in ribosome biogenesis and single-organism cellular process.
Each gene in the BDGP data has been labeled manually by annotations describing the spatial gene expression, using 295 annotation terms. We expect that since the gene interaction network is constructed via spatial similarity, genes that are connected to each other in the network will have similar spatial annotation terms.
To test this, we cluster the gene network using spectral clustering [52] into 12 clusters, and analyze the enrichment of each cluster for annotation terms using the hypergeometric test, with Bonferroni correction used to correct for multiple hypothesis tests. In the gene network for the 9–10 stage, 11 of the 12 clusters are enriched for 63 total annotation terms (Figure 10). The only cluster not showing any enrichment in the 9–10 stage network is also the smallest cluster, having only 4 genes. For example, in cluster 8, 92% of the genes have expression in the ventral nerve cord primordium P3 , while only 8% of the genes in the data have expression in this region. Similarly, 73% of the genes in cluster 11 have expression in the trunk mesoderm primordium, while only 16% of the genes in the data have expression in this region. For the 13–16 stage network, all 12 clusters are enriched for a total of 81 enrichments, a part of which is visualized in Figure 10. Tables S1 and S2 in the supplementary material report the complete enrichment analysis.
We learn a network from microarray data collected by the BDGP project over 12 time points in embryonic development [16], over the same genes that are being studied in the 9–10 and 13–16 networks, using covariance between the microarray expression as the kernel. We find that the overlap in edges between the 2 networks is very small, only 1% of the edges are common to both networks. If we assume that spatial expression annotations are a proxy for functional enrichment, then we can check if the microarray network is enriched for the spatial annotation terms. Figure 14 shows that the percentage of enriched clusters in the microarray network is small, independent of the number of clusters analyzed.
We can also test functional GO enrichment of the hubs of the network. Table 2 shows that the hubs of the microarray network for stage 13–16 are enriched for only a single function, where 4 of the 145 hub genes are involved in the “aromatic compound catabolic process”, while the microarray data network for stage 9–10 has no enrichments.
Thus, we find that the network learned from ISH images is clearly different from a network learned from microarray data. The ISH image network is enriched for spatial annotation terms, as well as functional enrichment of the hubs of the network, which does not hold true for the microarray network. This suggests that analyzing ISH images could support different scientific conclusions, which should be studied in greater detail.
GINI predicts gene interaction networks by analyzing Drosophila embryo ISH images. While the experiments above have been reported on the ISH data from BDGP, the GINI algorithm can be applied to all image data, by suitably modifying only the image processing pipeline. Using synthetic and image data, we establish that GINI fits the ISH data well, with low error residues, and that it can learn the true network correctly even if the data is not completely i.i.d. The analysis of the BDGP data shows that the hubs of the predicted gene interaction network are enriched for essential cellular functions, and that different regions of the interaction network are enriched for different combinations of annotation terms describing the gene expression. Thus, the predicted gene interaction network is capturing essential spatial and functional information about the expression pattern of the genes. We found that the gene interaction network learned from ISH images differs significantly from a network learned from microarray data.
The current work focuses on extracting gene networks from spatial data. The next step is combining information from multiple time stages to improve predictions, thus learning spatial-temporal gene networks. The problem of time-varying networks has been studied extensively for microarray data, by using different statistical penalties to estimate the network. For example, Ahmed et. al. [22] construct time varying networks by using a temporally smoothed -regularized logistic regression formulation, while Danaher et. al. [53] propose a fused lasso and group lasso based approach to combine information across time. Extensions of such algorithms for image data require stronger assumptions on data quality, such as having the same number of genes and image quality across time. Further, certain development stages may be less informative than others; for example, very few genes are active at development stage 1–3, and expression data from this stage is not as informative as expression data from development stage 13–16, when the embryo is much more mature. Developing algorithms that can account for such variations in data quality, while combining information across time, remains an interesting future direction to explore.
|
10.1371/journal.pgen.1004983 | K-homology Nuclear Ribonucleoproteins Regulate Floral Organ Identity and Determinacy in Arabidopsis | Post-transcriptional control is nowadays considered a main checking point for correct gene regulation during development, and RNA binding proteins actively participate in this process. Arabidopsis thaliana FLOWERING LOCUS WITH KH DOMAINS (FLK) and PEPPER (PEP) genes encode RNA-binding proteins that contain three K-homology (KH)-domain, the typical configuration of Poly(C)-binding ribonucleoproteins (PCBPs). We previously demonstrated that FLK and PEP interact to regulate FLOWERING LOCUS C (FLC), a central repressor of flowering time. Now we show that FLK and PEP also play an important role in the maintenance of the C-function during floral organ identity by post-transcriptionally regulating the MADS-box floral homeotic gene AGAMOUS (AG). Previous studies have indicated that the KH-domain containing protein HEN4, in concert with the CCCH-type RNA binding protein HUA1 and the RPR-type protein HUA2, facilitates maturation of the AG pre-mRNA. In this report we show that FLK and PEP genetically interact with HEN4, HUA1, and HUA2, and that the FLK and PEP proteins physically associate with HUA1 and HEN4. Taken together, these data suggest that HUA1, HEN4, PEP and FLK are components of the same post-transcriptional regulatory module that ensures normal processing of the AG pre-mRNA. Our data better delineates the roles of PEP in plant development and, for the first time, links FLK to a morphogenetic process.
| Unlike animals, angiosperms (flowering plants) lack a germline that is set-aside early in embryo development. Contrariwise, reproductive success relies on the formation of flowers during adult life, which provide the germ cells and the means for fertilization. Therefore, timing of flowering and flower organ morphogenesis are critical developmental operations that must be finely regulated and coordinated to complete reproduction. Arabidopsis thaliana FLOWERING LOCUS WITH KH DOMAINS (FLK) and PEPPER (PEP) encode two KH-domain RNA-binding proteins phylogenetically related to human proteins characterized by their high developmental versatility. FLK and PEP modulate the mRNA expression of the MADS-box gene FLOWERING LOCUS C, key in flowering control. In this work we have found that FLK and PEP also play a pivotal role in flower organogenesis by post-transcriptionally regulating the MADS-box floral organ identity gene AGAMOUS (AG). Interestingly, FLK and PEP physically interact with proteins involved in AG pre-mRNA processing to secure correct AG function in the floral meristem and flower. Taken together, our results reveal the existence of a post-transcriptional regulatory activity controlling key master genes for floral timing and flower morphogenesis, which might be instrumental for coordinating both developmental phases.
| Development of multicellular organisms relies on exquisitely controlled transcriptional and post-transcriptional regulatory actions to govern gene expression and accurately respond to endogenous and environmental fluctuations. As exemplified in the reference plant Arabidopsis thaliana (Arabidopsis hereafter), reproductive success in angiosperms largely depends on two developmental events that initiate the reproductive phase: floral timing and flower morphogenesis. Upon flowering, the shoot apical meristem (SAM) transforms into an inflorescence meristem (IM) which will give rise to floral meristems (FMs) [1]. FM identity genes, such as LEAFY (LFY) [2] and APETALA1 (AP1) [3], are crucial in activating the floral homeotic genes that specify identity of concentric whorls of organs in the Arabidopsis flower [1]. According to the ABC(E) model [4–6], the class A genes AP1 and AP2 specify sepals and, together with the B function genes PISTILLATA (PI) and AP3, contribute to petal identity. Co-expression of B-genes and the C-function gene AGAMOUS (AG) confer male stamen identity, while AG alone specifies female carpels, defining the pistil or gynoecium situated in the innermost whorl. The model also establishes mutual antagonism between A and C activities and requirement of the E activity, represented by the redundant SEPALLATA function [4–9]. With the exception of AP2 (an AP2/EREBP) [10,11], all floral homeotic genes encode type II MADS-box transcription factors, a lineage comprising central regulators in most aspects of plant development [9,12,13].
In addition to floral organ identity, AG plays a crucial role in FM determinacy by repressing the homeobox stem-cell-identity gene WUSCHEL (WUS) [14,15]. WUS and LFY activate AG, which in turn, represses WUS both directly and through the activation of the transcriptional repressor KNUCKLES (KNU) [16], resulting in consumption of the stem cell niche [16–21]. Otherwise, continuing cell proliferation leads to an indeterminate pattern of alternating whorls of sepals and petals, as described in strong ag mutants [22].
Whereas transcriptional control of gene expression is key to development, it is nowadays widely accepted that post-transcriptional operations are crucial to secure proper gene regulation. For example, mounting evidence indicates that mRNA processing steps, such as splicing and polyadenylation, usually proceed co-transcriptionally in a tightly coordinated manner to ensure correct gene activity [23–25]. RNA-binding proteins from multifunctional ribonucleoprotein (RNP) complexes coat nascent transcripts to regulate different aspects of mRNA synthesis, affecting thus, the final levels of gene expression [26, 27].
It has been shown that, in addition to its transcriptional control, post-transcriptional regulation is essential to secure correct AG function during flower development, in particular AG intron 2 processing [28]. So far three Arabidopsis RNA-binding proteins (RNPs) were found to facilitate this process: HUA1, a nuclear CCCH-type zinc-finger protein [29], the RPR-domain (Regulation of nuclear pre-mRNA) protein HUA2 [30], and HUA ENHANCER 4 (HEN4), containing 5 K-homology (KH) domains and one of the few KH proteins functionally characterized in Arabidopsis [31,28]. Interestingly, hua1 hua2 hen4 triple mutants displayed stamen and carpel homeotic transformations, and loss of flower determinacy as a result of the reduced levels of mature AG mRNA. The fact that HUA1 binds to the AG pre-mRNA and physically associates with HEN4, suggests that both proteins belong to the same RNP regulatory complex [28].
Named after the human heterogeneous nuclear ribonucleoprotein K (hnRNP K) [32], the KH domain is an ancient RNA-binding module present in proteins whose disruption causes important developmental alterations in animals, including human syndromes as fragile-X [33,34], metastasis and cancer progression [35]. The hnRNP K is also representative of the remarkably versatile poly(C)-binding proteins (PCBP), characterized by a stereotypical triple-KH-domain configuration. PCBPs play roles in multiple developmental processes in animal systems, from erythropoiesis to neuronal differentiation [36–40]. The KH domain also provides a structural basis for protein-protein interactions, which most likely contributes to the multifunctionality of PCBPs [36,41].
In contrast, very little is known about plant PCBP-type hnRNPs and their relevance to plant development or morphogenesis is largely unexplored. So far, only two canonical PCBP-type hnRNP encoding genes, FLOWERING LOCUS WITH KH DOMAINS (FLK) [42,43] and PEPPER (PEP) [44], have been characterized in Arabidopsis to some extent. FLK promotes flowering in the autonomous pathway by negatively regulating the MADS-box floral repressor FLOWERING LOCUS C (FLC) [42,43,45]. PEP was originally described to interact with element(s) of the WUS pathway [44] and more recently we found that PEP is a positive regulator of FLC activity, hence antagonizing with FLK [46]. In line with this, the late flowering phenotype of flk plants (due to elevated levels of FLC) is rescued in the flk pep background [46]. However, in spite of the fact that PEP is expressed in FM and developing flowers, pep, flk or pep flk double mutants lack conspicuous floral defects, probably reflecting the compensation by overlapping activities [43,44,46].
In this work, we have functionally investigated the magnitude of PEP and FLK roles in flower patterning. Our genetic and molecular analyses place PEP as a positive regulator of the floral C-function by facilitating AG pre-mRNA processing and preventing premature polyadenylation in the large second intron. Here we also show that FLK also contributes to maintain the C-function. Furthermore, we provide evidence that PEP and FLK interact with the previously identified AG mRNA processing factors HUA1 and HEN4, strongly suggesting that all these proteins likely work together as components of a common post-transcriptional regulatory activity. Identifying PEP and FLK as new regulators of AG, broadens the scope of the developmental functions played by plant PCBPs, as they impinge upon the control of master regulatory genes, in this case AG, central during reproductive development.
As mentioned above, PEP is expressed in FM and developing flowers, but pep flowers are largely normal. Thus, to test whether the role of PEP in floral patterning was masked by redundant gene activities, we combined the null pep-4 allele [44] (pep hereafter) with mutations in HEN4, HUA1 and HUA2, genes that encode post-transcriptional regulatory proteins [28].
HEN4 is a KH paralog relatively distant to PEP [31]. Unlike hen4-2 (hen4 hereafter) and pep single mutants (S1A Fig.) [28,44], ∼10% of hen4 pep flowers exhibited petaloid stamens (Fig. 1A, 1B, and S2B-S2D Fig.). Similarly, hua1-1 mutants (hua1 hereafter) appeared normal (S1B Fig.) [30], but hua1 pep double mutants displayed abundant petaloid transformations in the third whorl (40% of the flowers examined; Fig. 1C). We could not obtain hen4 hua1 pep triple homozygous mutants implying that PEP becomes essential in the hen4 hua1 background. This was noteworthy since hen4 hua1 double mutants flowers look wild-type [28]. Strikingly, introducing only one pep allele into hen4 hua1 plants (hen4 hua1 pep/+) led to conspicuous floral alterations including petaloid stamens in all flowers (Fig. 1D).
Loss of HUA2 does not cause any obvious floral phenotype [30] and, although HUA2 and PEP interact during floral timing, hua2-4 pep flowers are normal [46]. However, this might not be surprising as our data indicate that the hua2-4 allele is leaky (S3 Fig.). We therefore used the null hua2-7 allele (hua2 hereafter, unless it is specified otherwise). Double mutants hua1 hua2 showed a variety of flower defects, including stamen-to-petal transformations (Fig. 1E and S1 Table), as reported for hua1 hua2-1 [30]. Unexpectedly, we were unable to isolate hua2 pep or hua1 hua2 pep individuals and only hua1 hua2 pep/+ plants were identified among the progeny. This background was sterile and showed a significant enhancement of the hua1 hua2 floral phenotype, including stronger petaloid transformations (Fig. 1E-I and S1 Table).
The fruit derives from the fertilized gynoecium carpels, whose formation, in turn, almost entirely depends on C-function [7,47]. We therefore decided to use carpel and fruit development as readout of how pep, hua, and hen mutant combinations affect C-function.
Although fruits in some of the mutant backgrounds were slightly shorter but normal looking (100% in hen4 pep, 20–60% in hua1 pep), we detected pistils with very distorted development, such as unfused carpels, and reduced style and stigma (S1D Fig.). In certain combinations, the apical portion of carpels was pointed with areas of white or pale green tissue conformed by smaller fringe cells as those in the apex of wild-type sepals (Fig. 1K, 1L, 1N, 1O and S2F, S2G Fig.).
The hua1 hua2 double mutant presented shorter pistils broadened at the tip [30] (Fig. 1M and S1D Fig.). However, hua1 hua2 pep/+ pistils were on average much shorter and crumpled (S1D Fig.). Indeed, close inspection of severely affected gynoecia in hua1 hua2 pep/+ by scanning electron microscopy (SEM) revealed that the carpel epidermis, rather than the wild-type characteristic vertical files of smooth cells (Fig. 1P), showed a wide range of epidermal cell sizes with epicuticular wax crenulations, including sepal-like giant cells [48–50] (Fig. 1N, 1Q, 1R). These alterations are typical of carpel-to-sepal transformation and were also seen in additional pep mutant combinations (Fig. 1K, 1L and S2H-S2K Fig.).
We detected that a significant percentage of hua1 pep pistils (40%) developed supernumerary valves (S1 Fig. and S2L Fig.). This trait is typical of loss of meristem determinacy and it was further enhanced in hua1 hua2 pep/+ (Fig. 1O and S1 Table). Terminal hua1 pep flowers, and at least a quarter of the hua1 hua2 pep/+ flowers exhibited conspicuously long gynophores and gynoecia that, strikingly, contained additional flowers inside. These basically consisted of petals and sepaloid gynoecia recapitulating the sepaloid features seen in the fourth whorl (Fig. 1S, 1T and S2M, S2O Fig. and see below). This phenotype, never observed in hua1 hua2 flowers (S1 Table), was reminiscent of that of ag mutants and also resembled the loss of HEN4 in the hua1 hua2 background [28]. In hen4 hua1 pep/+ a significant fraction of flowers (25%) contained supernumerary sepaloid valves (S2N Fig.), reflecting certain loss of determinacy in this genotype.
Overall, these results indicate that PEP, in collaboration with HUA and HEN genes, act as a positive regulator of the floral C-activity to, therefore, secure the downstream developmental programs depending on this function, such as fruit development.
The mutant combinations described above exhibit very similar developmental defects. Moreover, gene dosage effects in hua1 hen4 pep/+ and hua1 hua2 pep/+ plants illustrate the sensitivity of such backgrounds to PEP activity. These findings strongly suggest that PEP shares redundant developmental functions with HUA1, HUA2 and HEN4 despite their protein structural disparity. Accordingly, hen4 hua1 hua2/+ pep/+ plants showed very dramatic floral alterations (Fig. 1U-W and S2P-S2R Fig.). Hence, these factors were tentatively included in a common gene activity abbreviated as HUA-PEP along this work.
The MADS-box regulatory gene FRUITFULL (FUL) [51] is crucial for valve formation during ovary patterning, and it does so, in part, by preventing valves from adopting valve margin identity through the negative regulation of valve margin identity genes [52–56]. Upon fertilization, ful lignified valve cells remain small, arresting stomata development and silique growth. However, replum cells develop normally leading to a characteristic zig-zag configuration of this tissue in ful fruits [51] (Fig. 2A, 2B and S4A Fig.). Additionally, ful siliques show elongated styles [57] (Fig. 2A and S4A, S4K Fig.).
To get more insights into the role of HUA-PEP activity during pistil development, we decided to characterize the behavior of hua-pep activity mutants in the ful background.
The ful-1 hua1 fruit was virtually identical to that of ful-1 plants [51] (ful hereafter; S4A Fig.). In contrast, ful pep, ful hua2 and ful hua1 pep siliques were progressively longer and showed shorter styles (S4A Fig.). In such backgrounds, valve epidermal cells were elongated and streaked, along with interspersed stomata. These phenotypes indicated that valves took onto sepaloid identity (Fig. 2B and S4B Fig.). In ful hua1 hua2, gynoecia were smaller and replum cells remained small as in wild-type unpollinated pistils [58], abolishing the characteristic zig-zag shape (Fig. 2A, 2B and S4A Fig.). Fertility in ful hua1 hua2 plants was severely reduced.
ful hua1 hua2 pep/+ plants were phenotypically identical to hua1 hua2 pep/+. In such combinations, we also found new floral organs developing inside swollen gynoecia that were often seating on long gynophores (Fig. 2E and S4A, S4E, S4I, S4L, S4M Fig.).
The glucuronidase (GUS) reporter harbored by the ful-1 transposon reflects the native expression pattern of FUL [51]. Pistils of ful or wild-type-looking heterozygous ful/+ plants displayed characteristic GUS activity in the valves, style and nectaries [51] (Fig. 2C and S4D, S4G Fig.). In ful hua1 hua2, strong GUS signal was detected in nectaries and apical territory preserving style identity, whereas valves presented a more irregular pattern (Fig. 2D and S4H Fig.). The GUS-staining pattern of ful hua1 hua2 pep/+ in the fourth whorl organs had little resemblance to that of a gynoecium, except in nectaries and style vestiges, notably evoking FUL expression in the sepal vasculature [51] (Fig. 2E, 2F and S4E, S4F, S4I Fig.).
Next, we treated flowers with the lignin-specific dye phloroglucinol. Mature wild-type fruits showed preferential staining in the valve margin, whereas in ful mutants valves were ectopically lignified [52,53] (S4J, S4K Fig.). Nonetheless, in equivalent flowers from ful hua1 hua2 pep/+ plants, lignification in the presumptive gynoecium was nearly restricted to branched red lines with striking resemblance to lignified sepal vasculature (S4L-S4N Fig.). Altogether, genetic and histochemical analyses indicate that the HUA-PEP gene activity is required to prevent gynoecium tissues from adopting sepaloid fate independently of their original identity (valve, valve margin), highlighting the role played by PEP to preserve carpel identity.
To determine whether PEP impinges upon AG regulation and therefore C-function, we measured mRNA levels from wild-type and mutant flower buds by quantitative PCR (qPCR). In consonance with the phenotypes described above, relative expression of AG decreased significantly in hua1 pep and hua1 hua2 double mutants, and reduced even further in hua1 hua2 pep/+ plants (Fig. 1X).
To investigate whether somehow PEP (and HUA) control A and B function, we measured the transcript levels of the homeotic A- and B-class genes AP1 and PI, respectively. Results were inconclusive because, although expression of both genes declined moderately in some mutant strains (S5 Fig.), no morphological evidence of altered A- or B-floral functions was observed in any of the hua-pep mutant combinations examined. Therefore, our molecular and genetic data suggest that, in contrast to the C-function, it appears that HUA-PEP gene activity has little or no role in regulating A- and B-functions.
AG triggers several reproductive developmental programs in part by activating additional regulators that perform different subsets of its functions. For example, SPOROCYTELESS (SPL) stimulates stamen development, including organ identity [59–61], whereas the zinc-finger gene KNU cooperates with AG to repress WUS [19,21]. Consistently, SPL and KNU expression decreased markedly in hua1 pep and hua1 hua2 double mutants, and hua1 hua2 pep/+ plants (S6A, S6B Fig.). Accordingly, KNU gene expression monitored by a GUS-reporter construct was found to be less intense in hua1 pep developing flower organs, as compared to the wild type (S6C-S6H Fig.).
Interestingly, we found that the mutant phenotype of hua1 pep plants was completely rescued by increasing the dosage of AG gene with a genomic construct able to complement ag mutants [62] (S7 Fig.), thus reinforcing our hypothesis that AG functions depend on HUA-PEP activity genes. Collectively, these results might explain the organ identity and determinacy defects seen in pep hen hua combos and further support PEP as a positive regulator of AG.
One of the functions of AG is to prevent AP1 expression in the two inner whorls of organs where stamens and carpels normally form [8]. To examine the expression of AP1 we used the genomic GFP (green fluorescent protein)-based reporter gAP1::AP1-GFP, that largely mirrors endogenous AP1 expression [62]. As expected, in the wild type AP1-GFP signal was detected in sepals but absent in pistils (S8 Fig.). However, a number of hua1 pep pistils showed AP1-GFP fluorescence (Fig. 3A-D). These results are coherent with earlier work showing AP1 mRNA ectopic expression in inner whorls of hua1 hua2 and hua1 hua2 hen4 [28,30], and underscore the importance of PEP as a regulator of the C-function during flower organogenesis.
Our loss-of-function genetic analyses show that components of the HUA-PEP function are redundantly required for the floral C-function. So we asked whether PEP alone could compensate for the deficiency in members of this activity. To test this idea, a 35S::PEP overexpressing construct [46] was introduced into the hua1 hua2 background. Strikingly, PEP overexpression, instead of rescuing, dramatically enhanced the hua1 hua2 mutant phenotypes. Homozygous hua1 hua2 35S::PEP flowers were sterile, and exhibited much stronger stamen-to-petal and carpel-to-sepal transformations than in hua1 hua2, as well as frequent severe indeterminacy defects, a trait never observed in hua1 hua2 plants (Fig. 4A-D, 4F-H, S9A-S9E Fig. and S1 Table). In line with the strong phenotypes observed, the levels of AG, KNU and SPL mRNAs in hua1 hua2 35S::PEP plants were significantly lower than those of hua1 hua2 mutants (Fig. 4I and S9K, S9L Fig.).
We ruled out any RNA silencing effect as hua1 hua2 35S::PEP plants showed much higher PEP mRNA levels than wild-type individuals (S10D Fig.). Rather, PEP protein overproduction might exceed a certain critical threshold, leading to the strong phenotypes observed. Consistent with this idea, hemizygous hua1 hua2 35S::PEP/+ plants produced PEP mRNA levels higher than those of the wild type, yet much lower than in homozygous hua1 hua2 35S::PEP plants (S10D Fig.), and did not show the severe floral phenotypes of the latter, being indistinguishable from hua1 hua2 individuals (S10E-J Fig. and S1 Table).
Although, PEP overexpression in hen4, hua1 and hua2 single mutant backgrounds did not result in noticeable morphological alterations (S10A-C Fig.), we speculated whether excess of PEP was critically detrimental in more compromised conditions. In line with this interpretation, PEP overexpression in the wild-type looking hen4 hua1 plants [28] led to the same developmental abnormalities previously described for the strong deficient hua-pep backgrounds. A significant number of hen4 hua1 35S::PEP flowers (∼65%) displayed severe indeterminacy, closely resembling ag flowers (Fig. 4E and S9F-J Fig.). It is worth mentioning that this phenotype never occurred in hen4 hua1 pep/+, indicating that PEP gain-of-function has a stronger impact on floral determinacy in hen4 hua1 than reducing PEP activity, similarly as described for hua1 hua2 background (S1 Table).
Mutations in HEN4, HUA1 and HUA2 led to a gradual decrease of AG mRNA levels concomitant with the accumulation of aberrant transcripts incorrectly terminated at the large second intron [28]. To test whether PEP impacts on this process, we carried out qPCR assays using intronic primers situated near the exon2/intron2 junction (Fig. 4J and S2 Table). The relative abundance of a PCR product increased progressively in various hua-pep mutant strains, notably in hua1 hua2 pep/+ and hua1 hua2 35S::PEP individuals, whereas it was barely detectable in the wild type (Fig. 4J). These values negatively correlated with the levels of correctly spliced AG transcript in the mutant backgrounds under study, and unambiguously indicated that altering levels of PEP has an important impact on the accumulation of these transcript species.
To examine transcript structure, polyadenylated RNA from hua1 hua2 pep/+ and hua1 hua2 35S::PEP plants was subjected to 3’ RACE (Rapid Amplification of cDNA Ends). Several products were obtained corresponding to transcripts comprising correctly spliced exons 1 and 2 followed by a variable stretch of nucleotides of intron two (105–368 nt), after which premature cleavage and polyadenylation events took place (S11A, S11B Fig.). These transcripts miss the last 6 exons, lacking the ability to encode a functional AG polypeptide. In plants, three polyadenylation signals define the site of processing: the far upstream element (FUE), the near upstream element (NUE), and the cleavage element (CE) [63]. Inspection of such RACE products revealed the presence of FUE, NUE and CE elements properly situated, strongly suggesting their implication in the premature termination event [63] (S11A, S11B Fig.).
FLK is expressed in all major organs, yet its loss of function did not cause any visible defect [42,43]. FLK interacts with PEP and HUA2 during flowering time regulation [46] but its possible role in flower morphogenesis has not yet been studied.
To explore FLK activity during flower development and to determine whether FLK participates in the HUA-PEP function, the null flk-2 mutant [43] (flk hereafter) was crossed to different hua-pep mutant combinations. flk hen4 double mutant flowers were wild-type in appearance (S12A, S12B Fig.). Unlike hua1 pep (Fig. 1 and S1 and S2 Fig.), flk hua1 and flk pep double mutant flowers also looked essentially normal (S12C Fig.) [46]. In contrast, flk pep hua1/+ plants showed some aberrant gynoecia, and petaloid stamens (S12D-F Fig.). Interestingly, stamen identity in the flk pep background, therefore, is sensitive to HUA1 gene-dosage since this trait is never observed in pep hua1/+, nor in flk pep flowers.
Next, the flk mutant was crossed to hua1 hua2 plants, a sensitized background repeatedly used to uncover gene activities involved in flower organ identity and determinacy [28,64–66, this work]. The resulting flk hua1 hua2 triple mutants were easily identified because of their conspicuous flower defects. flk hua1 hua2 flowers had two sets of petals and were “stamenless” (Fig. 5A and S12G Fig.), thus lacking fertilization and fruit set. Besides, flk hua1 hua2 gynoecium development was severely distorted with obvious sepaloid attributes (Fig. 5B, 5C). Nevertheless, the most defining feature was again the occurrence of indeterminate flowers (>50%) (Fig. 5B, 5D, 5E and S12H Fig.). As indicated above, hua1 hua2 flowers never show this severe developmental alteration, underscoring the contribution of the flk mutation to debilitate the floral C-function.
Our qPCR gene expression data backed up the hypothesis of FLK as part of the HUA-PEP activity. In flk the expression levels of AG, KNU and SPL remained unaltered when compared to those of the wild type, whereas in flk hua1 hua2 significantly dropped, being even lower than in hua1 hua2 individuals (Fig. 5F and S12I Fig.). This result substantiates the floral defects detected. Conversely, levels of AG transcripts containing intron 2 sequences increased in flk hua1 hua2 (Fig. 5G), suggesting an influence of FLK on AG post-transcriptional regulation. Indeed, we performed 3’ RACE assays for RNA from flk hua1 hua2 and identified new aberrant transcripts indicating premature cleavage and polyadenylation within the large intron 2. As described above, polyadenylation signals were found around the presumptive maturation site (S11C Fig.). Altogether, these results strongly support FLK as an additional component of the HUA-PEP activity.
As mentioned in the introduction, RNA binding proteins participate in multimeric RNP complexes to perform their regulatory functions [36,41]. Our genetic and expression analyses indicated that genes of the HUA-PEP activity act in concert during floral organogenesis, which makes reasonable their interplay at the protein level. Nuclear localization of their products has been demonstrated [28,42,43,46,67] and, importantly, physical interaction between HEN4 and HUA1 has been already established [28]. Moreover, HEN4 was also computationally predicted to interact with PEP and FLK [68]. We therefore, conducted in vivo bimolecular fluorescence complementation (BiFC) assays in tobacco leaves using PEP, FLK, HEN4 and HUA1. Reconstituted yellow fluorescent protein (YFP) was detected in leaf cell nuclei when FLK-PEP, HEN4-PEP and HUA1-PEP interactions were assayed, respectively (Fig. 6 and S13 Fig.). Similarly, robust nuclear interaction was seen when FLK was tested against HUA1 and HEN4 (Fig. 6 and S13 Fig.). The HUA1-HEN4 BiFC interaction was used as a positive control (S13 Fig.). All associations were tested in both directions, thus endorsing specificity of the interactions (Fig. 6 and S13 Fig.).
We were also able to confirm in vivo protein homodimerization of PEP and FLK in our assays, corroborating the publicly available in silico data [68] (S13 Fig.). Homodimer formation was also seen in HUA1 BiFC experiments (S13 Fig.). These associations were further verified in yeast-two-hybrid assays (Y2H; S14 Fig.).
In a subset of our BiFC assays we detected, in addition to clear signal in the nuclei, specific cytoplasmic fluorescence (Fig. 6 and S13 Fig.). KH-domain containing proteins, particularly PCBPs, are known to participate in numerous RNA processing events in the nucleus and in the cytoplasm (RNA transport, stability, translation) [36–38]. Therefore, this extranuclear signal might reflect additional regulatory roles for PEP and FLK in this cell compartment.
Taken together, these results strongly suggest that PEP, FLK, HUA1 and HEN4 proteins physically associate likely reflecting their participation in common multimeric complexes involved in pre-mRNA processing. Additionally, these data further reinforce the assumption of FLK as a new partner of the HUA-PEP activity.
PEP and FLK were previously identified to control flowering time through regulation of the FLC gene [46,42]. Now, our analyses show that PEP and FLK also play a key role in the specification of flower organ identity as components of the post-transcriptional machinery that ensures normal processing of the AG pre-mRNA. Genetic, functional and molecular interactions with additional RNA-binding proteins previously established as AG regulators [28] led us to define HUA-PEP as a common gene activity comprising HUA1, HUA2, HEN4, PEP and FLK.
We have demonstrated that PEP is functionally linked to the AG pre-mRNA processing pathway. Whereas hua1, hua2 and hen4 single mutants are phenotypically wild-type [28, this work], when these same mutants were combined with pep, we observed developmental abnormalities consistent with reduced C-function activity. Moreover, hua1 hua2 double mutant flower defects [30] were dramatically enhanced when combined by plants that were heterozygous for a mutation in PEP (hua1 hua2 pep/+), illustrating dosage-effects among HUA-PEP genes as previously reported for HUA1, HUA2 and HEN4 [28]. The intensity of these floral phenotypes correlated with a reduction in AG mRNA levels. As a result, the A-function gene AP1, which is normally expressed in whorls 1 and 2, was ectopically expressed in the inner whorls of hua1 pep flowers, consistent with a compromised C-function and the A-C antagonism [4,8]. The sepaloid transformations seen in gynoecium tissues when HUA-PEP genes were mutated in the ful background provided further evidence for the critical contribution of PEP to carpel identity.
Loss of PEP contributed to reduce the floral C-function activity. Surprisingly, PEP overexpression in hua1 hua2 and hua1 hen4 also caused a dramatic enhancement of flower mutant phenotypes. Although this might seem unexpected, there are many examples in which loss- and gain-of-function result in the same phenotypical alterations. Loss and overexpression of bancal, encoding a Drosophila homologue of vertebrate hnRNP K, generates appendage developmental defects [69]. In Xenopus embryos, both reduction and overexpression of the KH gene Mex3b, involved in neural plate formation, led to downregulation of target genes [70]. In Arabidopsis, increasing or reducing the expression of kinase-encoding genes FAB1A/B elicits the same pleiotropic alterations, which are attributed to perturbations in the protein complexes in which they participate [71].
However, PEP overexpression in wild-type or single hua-pep mutant backgrounds rendered normal flowers, suggesting certain buffering capacity against PEP excess. Nevertheless, simultaneous inactivation of various HUA-PEP components (hua1 hua2 or hen4 hua1) when PEP is overexpressed might aggravate a detrimental excess of PEP, by likely disrupting protein stoichiometric equilibria [72]. In line with this hypothesis is the fact that hemizygous hua1 hua2 35S::PEP/+ plants, expressing higher levels of PEP than the wild type but much less than homozygous hua1 hua2 35S::PEP plants, do not differ from hua1 hua2 double mutants.
Our analyses have also uncovered a role for FLK in plant morphogenesis. FLK participates in the HUA-PEP activity during C-function maintenance. The genetic interaction between flk, pep, hua1 and hua2, the phenotypic similarities between flk hua1 hua2 (Fig. 5) and hen4 hua1 hua2 [28], the gene expression analyses, as well as FLK physical associations, firmly support this conclusion.
FLK represses FLC and thus promotes flowering whereas PEP and HUA2 are FLC activators [42,43,46,73,74]. During flower morphogenesis, however, FLK and PEP promote flower morphogenesis through the positive regulation of AG (this work). Taking into consideration the promiscuity of RNA-binding proteins, it is very plausible that components of the HUA-PEP activity might be participating in functionally distinct complexes. This is not unprecedented. For example Arabidopsis SR (serine/arginine rich) factors and the hnRNP AtGRP8 exhibit antagonistic and cooperative effects during circadian regulation [75]. Also, closely related MADS-box genes AGAMOUS-LIKE 24 (AGL24) and SHORT VEGETATIVE PHASE (SVP) accelerate and delay flowering, respectively. Later, AGL24 and SVP cooperate with AP1 to downregulate AG during first stages of floral development [76–78]. Similarly, FUL-SVP replaces FLC-SVP heterodimers counteracting the repressive effect of the latter on flowering time [79]. Moreover, AG and AP3/PI participate in the same protein complexes to specify stamen anlagen. However, many genes promoting carpel development that are induced by AG are, on the contrary, repressed by AP3/PI [80]. Functional versatility of the HUA-PEP activity, in turn, might be very advantageous to provide regulatory flexibility to modulate the highly dynamic and complex networks governing reproductive development.
PEP and FLK physically associate, as well as with HUA1 and HEN4, indicating that, probably, they all participate in common regulatory complexes. HUA2, however, might affect AG independently since no physical interaction between HUA2 and any other HUA-PEP component described here could be detected in a recent Y2H screen [81]. Formally, HUA2 molecular interactions might be mediated through HUA-PEP factors yet to be identified. We observed stronger phenotypes in hua-pep backgrounds when HUA2 was mutated. These results might be explained with the existence of two complementary subactivities: one incorporating the HUA2 function and another one comprising the remaining identified HUA-PEP factors. Simultaneous disruption of both complexes might account for more profound phenotypic defects. Lethality in hua2 pep mutants substantiates this notion.
Our molecular analyses of hua-pep mutants are coincident with previous work showing accumulation of transcripts retaining intronic sequences at the expense of the functional AG mRNA [28]. A large intron where important regulatory motifs reside is a feature shared by AG, FLC and other MADS-box genes, that is conserved across species [82–87]. However, nascent transcripts are vulnerable to premature processing and large introns might increase the risk of cryptic signals recognizable by the splicing and/or polyadenylation machineries [88,89].
Transcript maturation mainly proceeds co-transcriptionally, increasing the fidelity of the process [24,90,91]. Altering PEP and FLK expression in the hua1 hua2 background had a profound effect on the accumulation of AG intron-retaining transcripts. Remarkably, FLC intron-containing transcripts also increased in pep plants [46]. We propose that the HUA-PEP proteins assist transcription elongation by “hiding” cryptic signals in the nascent RNA (Fig. 7A-C). Otherwise, these sites could be accessible to the corresponding processing machinery, giving rise to non-functional or prematurely terminated transcripts (Fig. 7D). Our hypothesis is consistent with the recent characterization of mammal PCBPs as global regulators of alternative polyadenylation. Knock down of PCBPs actually favors usage of cryptic intronic sites [40]. Interestingly, hnRNP K suppresses usage of a premature polyadenylation site for NEAT1, a long non coding RNA (lncRNA) operating in nuclear paraspeckles (ribonucleoprotein bodies) formation, thus increasing the ratio of the long effective transcript [92].
By sequestering intronic polyadenylation motifs, PEP (and the remaining HUA-PEP factors) may also facilitate correct splicing, as documented for other PCBPs [36,41]. The U1 snRNP (U1), in addition to its splicing role, protects pre-mRNAs from premature termination at intronic polyadenylation sites [88,89], raising the attractive possibility of a connection with the HUA-PEP gene activity. The carboxyl-terminal (CTD) domain of eukaryotic RNA polymerase II coordinates transcription and transcript maturation [93]. The Arabidopsis KH protein SHINY1 (SHI1) interacts with a phosphatase that dephosphorylates particular residues in CTD, downregulating transcription of abiotic stress-related genes by preventing 5’ capping [94,95]. Uncovering new functional and molecular relationships among distinct HUA-PEP components will certainly provide a better understanding of the developmental programs regulated by this activity (floral timing; flower patterning) and the importance, at the regulatory level, of multifunctional plant PCBP-type hnRNPs.
This work was carried out with the Arabidopsis thaliana Columbia (Col-0) accession as the wild type. Strains previously obtained in other accessions were backcrossed at least five times into Col-0 before any further experiment. Plant materials used in this study were pep-4 [44], flk-2 [43], hua2-4 [73]; hua2-7 [74], 35S::PEP [46], and ful-1 [51]. gAG::AG-GFP and gAP1::AP1-GFP [62] were provided by Gerco Angenent and Richard Immink (Wageningen University, The Netherlands). hen4-2 [28], hua1-1 and hua2-1 [30] were provided by Xuemei Chen (UC Riverside, USA). KNU::GUS [16] was provided by Anna M. Koltunow (CSIRO, Adelaide, Australia). Information about all primers used in this work and molecular genotyping can be found in S2 Table. Plants were grown in MS plates or soil as previously described [44].
Phloroglucinol lignin staining [96,97] and GUS assays were performed essentially as described [44,96,97]. All GUS analyses, except in the case of ful-1/+, were performed in homozygous lines. Whole-mount pictures were taken under a Nikon SMZ1500 stereomicroscope. Histological sections (8 μm) were photographed under bright-field or dark-field illumination using a Nikon E800 microscope. In both cases Nikon Digital Camera DXM1200F was used operated by the ACT-1 2.70 program. Scanning electron microscopy (SEM) was according to [44]. For confocal laser scanning microscopy, all analyses were performed in homozygous lines. Samples were pre-treated with methanol/acetone (1:1 v/v) solution for 30 minutes at -20°C, and subsequently rinsed in PBS buffer (1.94 mM K2PO4; 8.06 mM Na2PO4; 2.7 mM KCl, 0.137 mM NaCl, pH 7.4) to be observed under a Leica TCS SPE confocal microscope. Pictures were taken with the LAS AF program.
For quantitative RT-PCR (qPCR), 5 μg of total RNA was extracted from young flower buds until stage 9, treated with DNase I, and used for cDNA synthesis with an oligo(dT) primer and RevertAid Premium Reverse Transcriptase (Thermo Scientific) following the manufacturer’s instructions. Subsequently, for each qPCR reaction, 0.5 μl of the cDNA was used as template. Relative changes in gene expression levels were determined using the LyghtCycler 1.5 system with the LightCycler FastStart DNA amplification kit according to the manufacturer (Roche Diagnostics). RNA levels were normalized to constitutively expressed genes OTC (ORNITINE TRANSCARBAMILASE) [98] and ACT2 [99], and the corresponding wild-type levels, as previously reported [46,100]. Each experiment was undertaken using three biological replicates with three technical replicates each. The standard deviation was calculated in Microsoft Excel. Statistical significance was estimated by the Student’s t-test according to [101] (*P < 0.05, **P < 0.01).
For 3’ rapid amplification of cDNA ends (3’ RACE), 5 μg of young flower bud total RNA was reverse transcribed using Maxima Reverse Transcriptase and the adaptor oligo d(T)-anchor (kit 5’/3’ RACE, Roche Diagnostics) as a primer. Then, AG cDNAs were amplified with High Fidelity PCR Enzyme Mix (Thermo Scientific) using forward primers situated in the exon 2 (S2 Table) and the PCR anchor (Roche Diagnostics) as a reverse primer hybridizing with the adaptor sequence, thus ensuring that only polyA-containing sequences were amplified. Amplified products were cloned into pSC-A plasmids and sequenced with M13F and M13R primers. Sequences were analyzed using CLUSTAL-W aligning [102].
For bimolecular fluorescence complementation (BiFC), coding sequences of all genes under study were amplified from their respective cDNAs using Phusion Taq-polymerase (NEB). The corresponding primer sequences (S2 Table) were designed for cloning the resulting PCR amplicons via Gibson DNA assembly method [103], and cloned into both the pBJ36-SPYNE and pBJ36-SPYCE plasmids, containing N-terminal (nt) and C-terminal (ct) halves of the yellow fluorescent protein (YFP), respectively (YFPnt and YFPct) [104]. The 35S::SPYNE and 35S::SPYCE cassettes were then cloned via NotI into the binary vectors pGreen0229 and pGreen0179 [105], respectively. Transformed AGL-0 Agrobacterium tumefaciens cells were used to infect Nicotiana benthamiana leaves. YFP reconstituted fluorescence was visualized 72 h after inoculation under a Nikon Eclipse TE2000-U epifluorescence microscope. The reciprocal BiFC assays were also performed obtaining the same results as shown in Fig. 6 and S13 Fig. As negative controls, Nicotiana leaves were co-infiltrated with the corresponding recombinant YFPct construct and the empty YFPnt version, yielding no signal in any case.
For yeast two-hybrid assays, the cDNAs for PEP, FLK, HEN4 and HUA1 genes were amplified with the proof-reading Phusion Taq-polymerase (New England Biolabs, Inc.) using the corresponding primers (S2 Table). The resulting products were cloned into the pB42AD (+Trp) and pGilda (+His) vectors via Gibson DNA assembly procedure [103]. The integrity of constructs was checked by sequencing. The yeast strain EGY48 (-Ura) was cotransformed with the corresponding combinations of pGilda and pB42AD constructs. Empty vectors were used as negative controls. Positive colonies were selected on solid media (-Ura, -His, -Trp +glucose). Induction for testing protein-protein association was assayed growing the resulting yeast strains on plates or liquid in the presence of galactose and raffinose (DB Falcon). X-gal was used for colorimetric assays on plates, and ONPG (2-Nitrophenyl β-D-galactopyranoside, SIGMA) for β-galactosidase liquid experiments. The Clontech protocol book was followed for all these procedures.
|
10.1371/journal.pntd.0004471 | Dog Demography, Animal Bite Management and Rabies Knowledge-Attitude and Practices in the Awash Basin, Eastern Ethiopia | Rabies is a viral zoonosis that has been described in limited numbers of studies in Ethiopia at large and among pastoralists in particular. This study assessed dog demography, bite wound prevalence and management, potential risk factors of disease transmission and knowledge attitude practice towards rabies among urban dwellers, pastoralists and health workers in Awash, Eastern Ethiopia.
Information was collected by means of structured questionnaires and interviews and through medical and official records from the Agricultural and Health bureaus.
Respondents totaled 539 (471 urban, 49 pastoralists, 19 medical). Dog(s) were owned in 33% urban and 75.5% pastoralist households respectively. Mean dog number per dog owning household was 1.50 (95%CI: 1.40–1.60) in urban and 2.05 (95%CI: 1.51–2.60) in pastoralists sites. Human Dog Ratio in Metahara was 4.7:1. No bite wounds records were kept in medical facilities, where staff recalled around 100 bites per year, 2/3 being in adults. Over 90% of the respondents claimed knowing rabies but up to 79.2% pastoralist did not know how dogs acquire the disease; 37.3% urban and 23% pastoralist did not know the symptoms of rabies in dogs; 36% urban and 44% pastoralists did not know rabies symptoms in people. Eighty percent of pastoralists did not know that the disease was fatal in people if untreated. Over half (58.7%) of pastoralist respondents go to traditional healers if bitten, despite a health extension worker program in place in the study area. Knowledge gaps were also shown amidst medical staff.
The study highlighted overall poor disease knowledge, severe under-reporting of human rabies cases, lack of record keeping and poor collaboration between the public and animal health sectors and communities in rabies control.
| Rabies is a fatal viral disease of animals and people. People usually get infected via bites from an infected animal (e.g. dog). Post exposure prophylaxis (PEP) has to be initiated immediately after bite wounds of suspected rabid animals in order to avoid fatalities. The situation of rabies is poorly known in Ethiopia, particularly in the pastoral context. We conducted questionnaire surveys in urban, pastoral and medical health worker communities around Awash National Park (Ethiopia) in order to capture information on dog demography, bite wound prevalence and management, potential risk factors of disease transmission and knowledge- attitude- practice (KAP) towards rabies among these communities. Disease knowledge was generally poor. Dog demography varied depending on the community which would affect control strategies. Health facilities did not keep bite records and there was poor recording and reporting of rabies cases. Delivery of PEP was inadequate. Communication and collaboration between the public and animal health sector was poor to inexistent regarding reporting and control of rabies cases.
| Rabies is a viral zoonotic neglected disease caused by a negative stranded RNA virus from the Genus Lyssavirus [1].
Although a wide range of animals can become infected and transmit the disease, only mammals from the Carnivora and Chiroptera (bats) Order act as reservoir for the disease [2]. Domestic dogs are considered to be the main source (>90%) for human rabies in Africa [3, 4]. Once the symptoms have appeared, the disease ends almost always fatally. Transmission to people occurs predominantly via infected animal bite or scratch as well as via their saliva through mucosa and broken skin [3].
Therapy has to be initiated immediately and relies on Post-Exposure prophylaxis (PEP), which consists of rapid and thorough washing of the wound, completion of post-exposure vaccination schedules plus inoculation with rabies immunoglobulin (RIG) for severely exposed bite-victims.
The disease claims 24,000 human deaths annually in Africa alone [4, 5]. Rabies burden in Tanzania was 4.9 human death/100,000 based on active surveillance data on bite incidence and 0.62 human deaths/100,000 when based on national bite statistics [4]. Ethiopia is thought to be a high-burden country for rabies [6]. However, hard data on dog demography and ecology as well as true rabies incidence in dog and people are lacking. Information is based on estimations and extrapolations, small scale studies and limited record reviews [6–8]. The Ethiopian Public Health Institute (EPHI) is the only laboratory facility in the country to diagnose rabies and produce PEP. Ferni vaccines (adult sheep brain nervous tissue vaccines) were used at the time of this study, despite the WHO recommendation in 2006 to completely replace nerve tissue vaccine with cell-cultured based anti-rabies vaccines [9].A retrospective review from EPHI records between 2001 and 2009 showed that 1026–1580 patients per year in and around Addis Ababa were taking PEP and that the total fatality human cases was 35–58 per year [6]. A one year follow-up in Gondar based solely on clinical diagnosis revealed an incidence rate per year of 2.3/100,000 [8]. Rabies incidence however, is likely to be much higher considering the lack of accurate data and underreporting of cases [4, 5, 10]. Published data of rabies from rural areas of Ethiopia, including pastoralists are however entirely lacking.
The aim of this study was to try to gain a picture of the rabies epidemiology (prevalence, risk factors) at the animal-human interface in the Awash Basin. Dog demography, bite history and knowledge-attitude-practice (KAP) regarding the disease was assessed amongst pastoralists, urban dwellers and health workers by means of questionnaires and/or interviews. Constraints to rabies control and prevention in the study area are discussed.
The study was carried out between February and June 2012 in Metahara, the administrative center of Fentale woreda (Oromia region) with a population of 25,670, its neighboring town Addis Ketema and Merti (Metahara Sugar Cane Plantation) as well as in the neighboring pastoral villages of the Oromia and Afar region (Fig 1). Urban dwellers were from various national ethnic backgrounds. Pastoralists were Ittu-Oromo, Kereyu-Oromo or Afar. Climate is semi-arid with bi-annual rainfalls. The area has an elevation ranging from 800 to 960 meter above sea level (National Meteorology Agency; Metahara Agricultural Bureau).
This cross-sectional study collected solely in depth interview/questionnaire data and no clinical samples. Questionnaires, with closed and open questions were translated into the local languages Amharic, Oromifa and Afarinia and back translated into English for consistency checks. The questionnaires were pre-tested in the study site. A trained interviewer administered all interviews in the local languages. Two sets of questionnaires were prepared, one for urban/ pastoralist dwellers and one for health workers. The first questionnaire attempted to capture information on dog population structure and husbandry as well as Knowledge Attitude Practice (KAP) of the interviewees regarding rabies. The question categories included: general information on the interviewee, questions related to dog husbandry and demography, contact of dogs with livestock and wildlife, questions related to bites (person bitten, bite location, bite wound treatment), questions related to disease awareness/attitude/practice (e.g. transmission to animal and people, symptoms and outcome in animals and people, source of disease knowledge, rabies therapy, bite wound therapy, rabies in livestock), and willingness to buy vaccine for dogs if available.
The second questionnaire aimed at assessing the knowledge of rabies amongst health workers. All types of medical facilities were included: health posts (N = 3), Health Centers (N = 1), private clinics (N = 3) and Merti hospital. Questions included: information on the interviewee and his work (age, sex, religion, work place, professional background, years spend in the profession), the treatment of bite wounds (kind of treatment, information on the bitten patient, bite location, number of bitten patient seen), whether or not he/she had rabies patients in the last 12 months and information on these patients, rabies knowledge (transmission and symptoms in animals and people), rabies therapy, and suggested intervention for rabies prevention in people.
The respondent was never offered answer options but had to talk freely on a question. The question was first asked in broader term to assess general knowledge and subsequently narrowed down to more detailed questions. This was to ensure no unwilling biased guiding from the interviewer.
In addition, verbal (oral evidence) and recorded data and general information on rabies were collected from both, the Health Bureau and the Agricultural Office in Metahara as well as from all medical facilities.
We performed a non-probability sampling. A list of all pastoralist settlements was obtained from the Woreda Agricultural Bureaus of both Woredas; inclusion criteria were logistic feasibility (accessibility, security), proximity to the National Park, and willingness of pastoralists to participate in the interviews. All families of the chosen pastoral settlements and all medical staff present in the medical facility at the day of the interview were included in the study. For the urban study a house to house visit was made through the city and all families willing to answer the interview were included.
All data was entered into Microsoft Access tables and analyzed descriptively using Stata software (version 10.1, StataCorp, Texas, USA). A scoring between 1 and 3 was given to KAP parameters. For instance if a respondent could describe all main symptoms of rabies in dogs he would get a score of 1, if the respondent could not describe them but rather minor symptoms or remaining generally vague, he would get a score of 2. A score of 3 was attributed to respondents who claimed not knowing any symptoms. Pearson’s chi square statistics test was used to compare group differences for categorical variables in the KAP analysis as well as dog ownership by areas in Metahara. Associations and assessment of determinants for KAP were considered statistically significant if p<0.05.
The study received institutional ethical clearance from the AHRI/ALERT Ethical Review Committee (AAERC), number PO04/12. Heads of the Woreda Agricultural and Health Bureau in Metahara and Awash town were informed and permitted the project. All interviewees gave verbal informed consent.
Respondents totaled 539 (471 urban dwellers in Metahara, 49 pastoralists and 16 health workers, all medical facilities included). Amongst the health staff, all were nurses except for one health officer and 2 medical doctors. Nurses from health posts were also working as health extension workers.
Dogs totaled 236 in the 471 urban household and 76 among the 49 pastoralist households. Considering a population of 25,670 and 11,000 households (City Administration Metahara, annual census report 2011), the crude extrapolation for human dog ratio in Metahara was 4.7:1. Dogs were evenly represented throughout the city quarters (p: 0.138). Table 1 shows the demography and ownership of the dog population in Metahara and the pastoralist areas.
Dog ownership was not influenced by religious background (p: 0.12); 112/285 (39.3%) and 77/209 (36.8%) owned dog(s) among orthodox Christian and Muslim respondents respectively. Dog ownership differ statistically between urban and pastoral residents (p<0.001) (Table 1).
All dogs were kept as guard dogs and/or for livestock protection. None had received preventive rabies vaccination, was neutered/castrated, received veterinary care or had collars on. Puppies were never intentionally killed but kept or given away. They were all fed with left-overs from human consumption and/or left to roam free for food. The majority of the dogs were free-roaming during the day (stated by N = 114/157 (72.6%) urban and N = 33/37 (89.2%) pastoralist households) as well as during the night (N = 90 (57.3%) urban and N = 31/37 (83.8%) pastoralist households).
Dogs had regular contact with livestock in 94.3% and 94.6% of the urban and pastoralist households. Two urban and 14 (37.8%) pastoralist households had their dogs going regularly into the nearby National Park. Direct regular contact between dogs and wildlife was observed by N = 88/157 (56%) and N = 28/37 (75.7%) urban and pastoralist respondents respectively. Main wildlife species reported were hyenas, jackals and less often rabbits, monkeys and leopards.
None of the 471 urban and 9 (18.4%) pastoralist respondents experienced bite wounds in his/her household in the last 5 years. Among the pastoralists 1 was a child under 5 years, 5 were children between 5 and 15 years and 3 were adults. Bites were recalled to be located in the foot and leg (N = 7) and arm and hand (N = 2). Four out of the 9 bitten patients flushed the wound with water and soap while the others went to a medical facility in town. When bitten by an animal, 58.7% of the pastoralists, as opposed to 1% of the urban responded that they would go to traditional healers. The rest of the interviewed pastoralists either do wound cleaning themselves or go to medical facilities. Ninety-nine percent of the urban interviewees would go directly to a medical facility.
Bite wound information (e.g. patient age, sex, bite location, severity of bite) was not recorded in any of the medical facilities assessed. Overall, interviewed staff recalled 7 bites that they personally treated in the last 12 months, 2 from hyenas and 5 from dogs. The bite locations from dogs, as recalled, were most often in the legs followed by feet, arms and rarely the face. Hyenas on the other hand bit in the face. Total number of bite patients recalled from all health facilities was estimated to be around 100 per year, of which approximately 75% were adults. None of the health posts did treat patients with bite wounds and referred them directly to Health Centers or private clinics. Reasons given were the unavailability of water, soap and other disinfectant and medical supply. In the other visited facilities, all treating staff used disinfectants such as providone iodine to treat bite wounds as well as antibiotics. Initial flushing with water and soap/detergent has been done by 2 out of 16 respondents.
This study was done during a rabies outbreak, which started in the rural areas South-West of the Awash National Park in May 2012. A strychnine poisoning campaign, in a bid to stop the disease spread in the urban areas, killed in a couple of days 275 dogs (personal communication, Merti hospital). The pastoralist communities stated that many people had been bitten and had died of rabies during the outbreak. Exact numbers, however, were unavailable. The Agricultural Bureau stated (recalls only as no written records) that the outbreak lasted 2 months and that 11 livestock at least had died of rabies and 4 people were on PEP therapy. The Health Bureau officially recorded 2 patients undergoing PEP treatment and stated the outbreak did last 1 week only. In May only, Merti hospital on the other hand, treated 32 patients for bite wounds and 15 patients with PEP.
Dogs that are biting people are not further diagnosed as whether they have rabies or not. The decision to start PEP treatment on a bitten patient relies solely on the decision of the health worker.
Control of zoonosis and prevention of human cases is usually most cost-efficient by controlling the disease in the animal reservoir [11]. Knowledge of the dog population structure, dynamics and ecology is, however, an essential pre-requisite to achieve proper required preventive vaccination coverage in the dog population (critical vaccination coverage varies with animal density), not to waste scarce financial and logistic resources and avoid large-scale campaign failure [12, 13]. In Ethiopia, published data on dog population number, structure and dynamics is lacking. In our study, 33% of the urban (grossly 1 dog per 5 people) and 75.5% of the pastoralist households kept dogs, regardless of religious background. Despite it being a small geographical study area, it could be observed that dog ownership and demography differed between the urban and the rural households but also between the pastoralist groups (Afar versus Oromo). The observed sex-and age- based dog population imbalance is in line with previous findings [14,15]. Respondents in our study, with the exception of the Afar, generally preferred male dogs, a trend reflected in many developing countries [15]. Dogs were not intentionally killed, therefore a steady population increase would be assumed. However, in general, free roaming dogs have short life expectancy and 2/3 die in their first year [15]. In Kenya, a study showed that life expectancy for males was 3.5 years and for females 2.4 years [12]. In our study, veterinary care was inexistent and dog husbandry poor, factors likely to lead to high dog mortality.
A common perception of all respondents was the large number of stray dogs, hence un-owned dogs. This study did not investigate the number of stray dogs but we need to keep in mind they also contribute to human rabies. Pastoralists stated that they were coming from town. However, the majority of owned dogs were free-roaming (day and night time), and in constant quest for food since they receive only little left-overs at home. In addition, they were not wearing collars, thus easily mistaken for stray dogs. Studies have shown that in reality most dogs are owned [16,17]. This is an opportunity for health intervention campaigns since the fact that dogs have owners would facilitate regular vaccination campaigns in Metehara. To support this, the majority of the urban and pastoralist respondents stated that they would be willing to regularly pay for their dog’s vaccination if the vaccine was indeed available.
Large scale lethal poisoning with strychnine is a common preventive measure undertaken in dogs in Ethiopia. Four out of 16 health workers reported killing of stray dog as a means of eliminating dogs. Mass dog elimination, however, besides being unethical and hazardous to the environment, has been shown to be counter-productive as it will not affect the dog population size and will not stop the spread of rabies as dogs will engage in compensatory breeding and migrate into newly vacated territories, thus facilitating disease transmission [18,19]. Morters et al (2013) recently showed as well that rabies transmission is not density-dependent [20].
Rabies is one of 20 reportable diseases in Ethiopia. Our study showed however, that rabies notification was poor compared to other diseases. The authors observed a severe discrepancy between the orally recalled rabies cases, the number of used PEP bottles logged into medical facility pharmacies, and the number of cases actually officially recorded and reported. Mainly adults were recalled to have been bitten. Generally, children are known to be more at risk for being bitten [9,21,22]. Our results raises serious concern as whether children were indeed less at risk or whether they were not brought to health facilities when bitten, thus showing that rabies in children is likely to be underreported and their treatments severely neglected in these communities.
Bite locations were most often in the legs followed by feet, arms and rarely the face. This picture differed from a study in Tanzania where dogs bit mostly hands and face [4]. The location of the bite and the wound severity (scratch versus deep skin penetration) is likely to affect the outcome into clinical rabies [4]. Our study unfortunately could not collect enough details on bite wounds. Bite records (location, severity, patient identification) were not kept in any of the assessed health facilities. All information on bite wounds was collected only through verbal recollection of health workers. The authors assume however that the wounds must have been severe for patients to come to health facilities considering the cost and time involved for patients. Reliable record keeping of bite wounds has been shown to be a useful epidemiological tool as a proxy for human exposure and rabies incidence in animals [2,10,23,24]. In our case, calculation of rabies cases—and exposure- incidences in people and animals was also impossible due to the lack of accurate record keeping, severe underreporting, lack of population census amongst pastoralists and the dog population in particular.
Pastoralists of the study area are likely not to go to a medical facility when bitten; reasons include distance to health facility/logistics, mistrust in the medical system, and poor knowledge of the disease fatal outcome. Pastoral respondent showed to be a determinant for poorer knowledge of rabies particularly for rabies in people. Despite the presence of an extension health worker system, over half (58.7%) of the respondents said they would go to traditional healers if bitten. It is estimated that the majority of human rabies deaths occurs in rural rather than in urban areas [10,25].
The main constraint to human rabies prevention in the study area was the quantitative lack of PEP, and its immediate inaccessibility (the vaccine is available in Adama, 134 km away from Metahara) thus delaying severely the start of the prophylaxis. From the moment a person was bitten to the start of PEP injections, delays of several days (up to 5 days) were not unusual, particularly if patients came from rural areas. Wound cleaning was rarely performed as first aid, neither by the patients themselves nor the medical staff, which is a behavior described in other developing countries in Africa and Asia [22, 26,27]. However, immediate flushing of a bite wound for 15 minutes with water and soap can be lifesaving, as the virus is mechanically removed from the site or is rendered unable to invade tissue [28]. Neglecting immediate wound flushing was shown to increase the risk of developing rabies by fivefold [21]. Health posts, that are often the first health facility patients, particularly pastoralists, would visit, were not offering this simple, cheap and important service. Hence, health posts and health extension workers visiting remote rural villages are in a unique position to initiate this procedure before transferring a patient to a larger facility that could start PEP, to educate people at large and pastoralists in particular about the importance of immediate wound cleaning, as well as the outcome of an untreated patient with rabies. However, this study also highlighted knowledge gaps about rabies among the health staff. Patients, particularly pastoralists who are not knowledgeable about the disease will rely on the health practitioner’s advice for adequate treatment, as whether PEP should be administered or not, as also seen in a study in Tanzania [22]. In our study, 25% of health staff did not know that livestock can transmit the disease to people. Hence, patients bitten by rabid livestock are likely not going to be treated for rabies. Also only 2 out of 16 health workers knew that bats can transmit rabies. The role of bats in carrying and transmitting rabies in Ethiopia is not known. However, its role should be mentioned in any future awareness programs. The fact that 25% of health workers thought that rabies is a bacterial disease raises the question as whether they thought PEP is an antibiotic and/or if they would treat the patient with antibiotics. Unfortunately we did not look further into this point. These results overall show an urgent need of improved training of health workers in rabies epidemiology and treatment.
On the other hand, the lack of diagnosis in dogs implies that likely a high number of patients are unnecessary treated, putting a strain on the already difficult logistics, economics and availability of PEP. Deressa et al. (2010) showed that only 10% of the dogs that had bitten people and were brought to EPHI for quarantine were actually rabid [6]. There are currently, new rapid and simple rabies diagnostic tests on the market that can be used directly under field condition, such as the Anigen rabies test, an immunochromatographic test (ICT) giving results within minutes, not requiring expertise nor special facilities/equipment and thus helping in the decision of PET use [29]. This would require, however, close collaboration and communication between the Agriculture Office, the Health Bureau, the veterinarians, the clinicians, and the communities. In our study, we could observe that collaboration and communication between the different stakeholders was poor to non-existent.
This study highlighted overall poor disease awareness amongst non-medical respondents although over 90% stated knowing the disease. A high percentage of respondents, particularly pastoralists (79.2%) did not know how dogs acquire rabies. Water shortage, wind and consumption of rotten food were often given as reason. Over a third of respondents did not know the symptoms of rabies in dogs and in people. The fate of a rabid person was not known by the majority of the pastoralist respondents (up to 80%). This highlights that the seriousness of this fatal disease in people was poorly known.
In a country that lacks hard data on epidemiology, vaccines for dogs and PEP for humans, and that is tight by logistic and financial constraints to efficient medical rabies control, disease awareness takes an important place in disease prevention. Our study showed that family nucleus and the community particularly amongst pastoralists played a central role in passing down knowledge about the disease whereas school and media played a minor role. Urban respondents may have better access to media, but pastoralists gather a lot of information from the pastoral community during transhumance with their animals. The latter were also aware of the economic burden of the disease since livestock can be affected.
The use of media can be instrumental to increase rabies awareness in a population, similarly as was- and is being done with other diseases such as tuberculosis in Ethiopia; it targets large audience and can also help promoting responsible dog ownership.
The overall rabies knowledge amongst health practitioners was variable and showed some knowledge and diagnosis/intervention gaps, particularly among staff from private clinics. Rabies is still sometimes misdiagnosed for malaria.
This study was the first in its kind to look at a rabies situation in a defined study site from a holistic approach including as well the animal as human side, urban and pastoralist respondents, medical and non- medical people. The in-depth interviews as well as the analysis of official records gave a valuable insight in the many existing gaps on rabies knowledge, prevention and health delivery. These information are valuable before embarking in a control and or awareness program in the study area. The limitations of the study lay in the non-probability sampling method used and the small sample size as well as the limited information for KAP determinant analysis. Also all data on bites and rabies patients (animal and human) were solely based on recalls and not on official records calling for inevitable biases.
Rabies is a 100% preventable disease but numerous challenges and constraints in Africa render its control and elimination difficult, hence relegating it to the neglected diseases [2, 30, 31]. In this study, it was observed that preventive dog vaccination was non-existent due to lack of vaccine availability at the time. In such a situation, disease awareness takes even more so, an important place in disease prevention as well in urban as in pastoral communities. However, this study highlighted overall poor disease knowledge and gaps in recognizing and treating rabies in people, the likelihood of severe under-reporting of cases, poor medical facility registries regarding bite wounds and rabies and a lack of collaboration between the animal and public health sectors. These factors are likely to hamper any future efficient rabies control campaign in the study area.
|
10.1371/journal.pgen.1006563 | Evolutionary analysis reveals regulatory and functional landscape of coding and non-coding RNA editing | Adenosine-to-inosine RNA editing diversifies the transcriptome and promotes functional diversity, particularly in the brain. A plethora of editing sites has been recently identified; however, how they are selected and regulated and which are functionally important are largely unknown. Here we show the cis-regulation and stepwise selection of RNA editing during Drosophila evolution and pinpoint a large number of functional editing sites. We found that the establishment of editing and variation in editing levels across Drosophila species are largely explained and predicted by cis-regulatory elements. Furthermore, editing events that arose early in the species tree tend to be more highly edited in clusters and enriched in slowly-evolved neuronal genes, thus suggesting that the main role of RNA editing is for fine-tuning neurological functions. While nonsynonymous editing events have been long recognized as playing a functional role, in addition to nonsynonymous editing sites, a large fraction of 3’UTR editing sites is evolutionarily constrained, highly edited, and thus likely functional. We find that these 3’UTR editing events can alter mRNA stability and affect miRNA binding and thus highlight the functional roles of noncoding RNA editing. Our work, through evolutionary analyses of RNA editing in Drosophila, uncovers novel insights of RNA editing regulation as well as its functions in both coding and non-coding regions.
| Many important modifications are made to RNA to fine-tune genomic information. One type, Adenosine-to-Inosine (A-to-I) RNA editing, changes certain adenosines to inosines and is essential for the neurological well-being of many animals. Although RNA editing occurs at thousands of sites across the genomes of various animals, the functions of nearly all editing events–particularly those in non-coding regions–have not been studied, and what determines whether particular adenosines across the genome are edited has not been fully explored. Here, using the Drosophila genus as model organisms, we analyze the evolution of A-to-I RNA editing to identify a large fraction of both coding and non-coding editing events that are under evolutionary constraint and therefore likely functionally important. We find that non-coding editing events in the 3’UTRs of genes could affect miRNA binding and are associated with a decrease in gene expression levels.
| Adenosine-to-inosine (A-to-I) RNA editing converts adenosine to inosine in RNA, which is then read by the cellular machinery as guanosine (G) [1–3]. This process is co-transcriptionally catalysed by adenosine deaminase acting on RNA (ADAR), which recognizes double-stranded RNA (dsRNA) structures as editing substrates [4]. A-to-I RNA editing plays a critical role in neuronal function and integrity through the fine-tuning of editing [5–8]. In humans, changes in editing are associated with neurological disorders such as amyotrophic lateral sclerosis [9] and autism [10]. In Drosophila, the knockout of ADAR leads to severe neurological defects such as locomotion impairment, heat sensitive-paralysis, and age-dependent tremors [11].
The widespread presence of A-to-I editing at thousands to millions of sites in various organisms, from flies to humans, has been recently revealed (e.g. [4,12–18]). However, only a handful of sites have been functionally studied [6] and it is unknown what fraction of editing events is functionally important. Recent studies have found that, although human RNA editing events are generally non-adaptive [19] possibly because the vast majority are in primate-specific Alu repeats, high-level nonsynonymous editing events, which cause amino acid changes, are most likely beneficial in humans [20,21]. In species such as squid and Drosophila, a significant fraction of editing events are located in coding regions, and many are likely to be beneficial [22–24].
While these studies have highlighted the potential functions of coding editing events, little is known about noncoding events. Few instances of noncoding RNA editing functions have been found: For example, RNA editing in the intron of mammalian ADAR2 alters splicing and thus truncates the protein [25]. In this study, we systematically examine both coding and noncoding editing events by comparing different species in the Drosophila genus and find that a surprisingly large fraction of both types are under selective pressure and therefore likely functionally important. Furthermore, we examine the potential functions of editing events in 3’ UTRs using a fly with catalytically inactive ADAR.
Although there are thousands of editing events in Drosophila, this only represents a small fraction of the adenosines in the genome, and why particular adenosines are edited is not fully explored. Our analysis of the evolution of RNA editing gives us an opportunity to study how the evolution of RNA editing substrates alters editing. By making incremental changes to editing substrates, evolution has conducted a natural experiment allowing us to test the effects of various changes on RNA editing. Therefore, we can examine the characteristics of editing substrates. dsRNA is a well-known requirement for A-to-I RNA editing [26–28]; we recently found that genetic variants associated with editing level changes are enriched in areas with secondary structure, and the structures around editing sites were more stable in the allele with higher editing levels [29,30]. Furthermore, an RNA sequence motif associated with editing has been found [31]. However, it is unknown how these two factors, dsRNA structure and RNA sequence, work together to determine editing levels. Here, we explore the combined contributions of dsRNA structure and sequence to the evolution of RNA editing in the Drosophila genus.
We first assembled a master list of editing sites in the Drosophila genus to use for our analyses. Although we and others have recently identified a large number of RNA editing sites in D. melanogaster (D.mel) [4,13,16], it is unlikely that their discovery has been saturated. To discover more editing sites in D.mel and identify additional editing sites unique in other Drosophila species, we applied our previously developed methods [16] to RNA-seq data (male and female whole body) from 13 Drosophila species (S1 Table). We identified 627 novel exonic editing sites, including 545 novel ones edited in non-D.mel species only (Fig 1A and 1B, S1A and S1B Fig, S2 Table). We combined this list with three other sets of sites identified in D.mel [4,13,16]. Using RNA-seq data from a D.mel ADAR null mutant [16], we estimated that the false positive rates of these four sets range from 1 to 7% (S1C Fig). We removed the false positive sites to obtain 2,380 exonic editing sites (counting orthologous sites only once) (S2 Table). These sites are located in 909 genes, preferentially in coding regions and 3’UTRs. 56% alter amino acid coding (S1D and S1E Fig).
We next used publicly available RNA-seq data, supplemented with our targeted sequencing assay covering ~600 editing sites, to measure the editing levels of the 2,380 exonic editing sites in D.mel and 5 other Drosophila species that spanned a range of evolutionary distances from D.mel and had well assembled genomes and deeper RNA-seq coverage (Fig 1A). To accurately measure editing levels in the RNA-seq data, only sites covered by ≥20 or 50 reads (medium or high stringency) were included, which leads to an average of 95 or 127 reads per editing site, respectively. While RNA-seq has been used to quantify RNA editing [12,16], its inaccuracy in lowly and moderately expressed genes hinders the accurate measurement of a consistent set of sites. Therefore, to validate and extend measurements in the RNA-seq data, we used the microfluidic multiplex PCR and deep sequencing (mmPCR-seq) assay [32] to quantify the editing levels of an average of ~600 sites in male and female whole body samples of the 6 selected species (S3 Table, S4 Table). We obtained ~2,400 reads per editing site per sample (S2A and S2B Fig). Editing levels are highly consistent between biological replicates (S2C Fig) and with the RNA-seq quantification when sufficient RNA-seq reads are available (S2D Fig). Finally, we combined the RNA-seq and mmPCR-seq datasets to obtain a more comprehensive and accurate RNA editing level measurement (Fig 1C).
When we compared editing levels between all pairs of the six species (Fig 1D, S2E Fig) and, separately, eight D.mel strains (S2F Fig), we found that there was a clear positive correlation between editing level divergence and species divergence time (Fig 1E). Furthermore, the editing level divergence tree recapitulates the topology of the phylogenetic tree (Fig 1F). These observations suggest that editing level evolution is generally neutral.
Since ADAR expression levels were very similar across species (S3 Fig), we reasoned that changes in cis regulatory elements may account for editing level variation at individual sites. Cis regulatory elements, namely the primary ADAR sequence motif and dsRNA structure around editing sites, are believed to play an important role in editing regulation across species, as demonstrated in a few case studies [23,28,33–35] and our recent larger scale studies [29,30]. Therefore, we first examined whether the underlying mechanism behind RNA editing divergence between pairs of species involved both RNA sequence motif and dsRNA structure. We also distinguished between sequence differences that removed editing entirely and ones that merely tweaked editing levels.
ADAR’s preferred sequence motif, in particular the triplet containing nucleotides immediately adjacent to the edited adenosine [31], is essentially identical in our six selected species (S4 Fig). To evaluate the effect of ADAR motif changes on editing level variation, we first deduced the ADAR motif weight matrix and used it to calculate motif scores for each of the 16 possible nucleotide triplets (S5 and S6 Tables). We compared “Presence/Absence” sites, whose editing was present (≥10% level) in the “anchor” species but absent (≤1.5% level) in the other species under comparison to the control “Presence/Presence” sites, which were edited (≥10%) in both species. With D.mel as an example anchor species, we found that ~25% of Presence/Absence sites had higher motif scores in D.mel, while only 8% of them had higher motif scores in the other, non-edited species, indicating an excess of better ADAR motifs in the “anchor” species (p = 0.065, 1.9e-3, 5.7e-5, and 6.4e-3 for D.mel vs. D. yak, D. ana, D. pse, and D. vir, binomial test) (Fig 2A). Strikingly, <1% of the “Presence/Presence” control sites had motif differences between species, although many of the sites were edited at different levels between two species. We observed similar results when performing the same analyses using the other 5 species as the anchor (S5A Fig). Therefore, ADAR motif preference plays an important role to determine whether a site is edited rather than how much a site is edited.
To examine the effect of secondary structure on editing level variation, we used a computational pipeline we previously developed [29] to predict editing complementary sequences (ECSs), the sequences base-paired with editing sites, for 1,664 (70%) editing sites in at least one species. Using these ECS predictions, we found that editing differences generally correlate well with secondary structure changes across species as exemplified in Fig 2B. This observation is also supported by our recent analyses of secondary structure changes within D.mel strains [29].
To systematically evaluate how secondary structure changes contribute to editing variation across Drosophila species, we dissected the dsRNA structures into 8 structural features and examined their roles separately (Materials and Methods). Using the set of Presence/Absence editing sites and its control set of Presence/Presence sites described above, we observed that Presence/Absence of RNA editing best correlated with free energy, stem length, and number of paired bases (Fig 2C, S5B Fig). This suggests that the stability and length of the RNA duplex play an important role for editing.
While we showed that both motif and dsRNA structure changes are associated with editing changes, their relative importance is unknown. To study how these two regulatory elements work together, we determined the relative importance of the features for the establishment of editing in Presence/Absence sites and the variation in editing levels observed in Presence/Presence sites. We used random forests, a machine learning technique [36], to predict the changes of editing between species using the motif and structural feature changes as predictor variables (Materials and Methods). The difference in the prediction accuracy before and after permuting the predictor variable is used as an importance measure (Fig 2D). For both scenarios, structural features such as folding energy, base-pairing, and stem length were important. While motif score and maximum bulge size were important in determining the presence or absence of editing, they were less informative for explaining editing level variation. Thus, our observations suggest that both motif and structural features determine whether ADAR can bind and edit a substrate, while changes in structural features contribute to editing level variation. It was previously shown that mutations in the ADAR motif can increase or decrease editing levels in vitro [26]. But our data suggests that structural features are generally used to tune editing levels in vivo, perhaps because such tuning allows for greater flexibility.
To estimate the contribution of cis-regulatory elements to editing divergence, we examined how well these features could be combined together to predict changes in editing using random forests [36]. The difference in the predictive accuracy between two models, with and without ADAR motif and dsRNA structure features, is used as a measure of the contribution of cis-elements to editing divergence. In particular, given the editing levels in D.mel, we predicted the editing levels of orthologous sites in other species. The predictive accuracy of the model with motif and structural features are substantially higher than those of the model without these features (Fig 2E). Therefore, a large portion of the editing level divergence across species can be accounted for by changes in cis regulation. Taken together, our cis regulatory element analyses reveal the structural landscape of editing and demonstrate how cis regulatory element changes lead to editing changes.
Our cross-species data allowed us to examine how RNA editing events may be born or persist during evolution. We used a combination of phylostratigraphy and transcriptome data to deduce the evolutionary age of each D.mel editing site (Fig 3A, Materials and Methods). We found that evolutionarily long-lived (older) editing sites have higher editing levels than younger sites (Fig 3B), suggesting that the older sites with higher editing levels might be optimized for ADAR binding and editing. We observed similar results when using the other Drosophila species as the anchor (S6 Fig). Our observation that older sites are much more likely to be clustered with other nearby editing sites (Fig 3C) further supports the idea that the sequences around older editing events can be edited at multiple positions and thus might be optimized for editing. Additionally, we found that the correlation of editing levels between species for old editing sites is higher than that for young editing sites (S7 Fig), suggesting that older sites are under stronger purifying selection. Furthermore, we examined the functional enrichment of genes containing the younger and older editing sites. We observed no significant enrichment for genes containing the younger editing sites. However, the older sites gradually became enriched in genes with functions that are mostly neuron-related (Fig 3D). This data agrees with the mostly neurological and behavioral phenotypes of ADAR knockout flies [11], and suggests that older editing sites may have neurological functions.
To determine whether editing tended to affect slowly or quickly evolving genes, we also examined the evolution rate of edited versus unedited genes. We only analyzed RNA editing sites identified from D.mel RNA-seq data alone [4,13], in order to avoid potential bias from using sites identified from cross-species comparisons. The evolution rate was measured using Omega (Ka/Ks), the ratio of the nonsynonymous site substitution rate (Ka) to the synonymous site substitution rate (Ks) [37]. We found that genes with nonsynonymous editing sites tend to evolve more slowly than unedited genes (Fig 4A, S8A Fig). This pattern persists when comparing genes with and without editing sites within the same functional categories (Fig 4B). Furthermore, the number of nonsynonymous editing sites per gene is negatively correlated with the evolution rate of edited genes (Fig 4C, S8B & S8C Fig). In contrast, no significant correlation was observed for genes with intronic editing sites (S8D Fig). Consistent with the findings above, we found that younger sites reside in genes evolving at a rate similar to the rest of the genome, while older editing sites tend to be in slowly evolved genes (Fig 4D). Additionally, neuronal genes evolved more slowly (S8E Fig). Taken together, these results suggest that individual editing events arise in genes of diverse functions and are subsequently selected to remain in slowly evolved genes, particularly with neuronal functions.
We next separated exonic D.mel editing sites based on whether they resulted in nonsynonymous, synonymous, or 3’UTR editing (5’UTR sites were not examined because only a few were available) and examined their evolution with respect to both editing levels and surrounding DNA sequences. First, we examined the editing levels of the eight D.mel strains. The editing levels of nonsynonymous sites are least variable, suggesting the strongest selective pressure. Unexpectedly, the editing levels of the 3’UTR sites are much less variable than those of the synonymous sites, indicating the presence of purifying selection around the 3’UTR sites (Fig 4E, S9A Fig). We observed a similar pattern between species, except that the editing levels of 3’UTR sites were more variable, probably because of the increased sequence divergence rate in noncoding regions between species (S9B and S9C Fig). Second, we examined the conservation levels of DNA sequences surrounding editing sites. We used RNA editing sites identified from D.mel RNA-seq data alone [4,13], in order to avoid the potential bias of using sites identified from the cross-species comparisons. We found that, compared to more distal flanking regions, the regions spanning editing sites and separately, their ECS sequences, had a decreased sequence divergence rate for nonsynonymous and 3’UTR sites, but not for synonymous sites (Fig 4F, S9D Fig). This is likely due to the evolutionary constraint in maintaining the dsRNA structure for many nonsynonymous and 3’UTR sites. Additionally, we observed higher editing levels of nonsynonymous and 3’UTR sites (Fig 4G). These data suggest that a large number of nonsynonymous and 3’UTR, but not synonymous, sites are functionally important.
Our observations prompted us to identify sites that were under evolutionary constraint and thus more likely to be functional (S2 Table). By examining the conservation of the regions spanning the editing sites and the more distal flanking regions, we categorized each editing site as highly constrained, moderately constrained, or unconstrained (Materials and Methods). We found 514 highly constrained and 466 moderately constrained sites (S2 Table). As expected, the surrounding DNA sequences of evolutionarily older sites tended to be highly or moderately constrained (S10 Fig). In addition, the highly constrained nonsynonymous and 3’UTR sites had the highest editing levels, and the moderately constrained sites had higher editing levels than the unconstrained ones (Fig 4G), further supporting the functional importance of highly and moderately constrained sites. If one assumes that lowly edited (<1.5%) events are functionally neutral and uses the fraction of lowly edited sites in highly or moderately constrained regions as a baseline, one can roughly estimate that 14.2% (= 75.4%–61.2%, Fisher’s exact test p-value = 5.5e-05) of nonsynonymous editing events and 12.0% (= 50.5%–38.5%, Fisher’s exact test p-value = 0.048) of 3’UTR editing events are in highly or moderately constrained regions and therefore likely functionally important. For synonymous events, the difference is not significant (Fisher’s exact test p-value = 0.42, difference = 2.7% = 45.1%–42.4%). (The fraction of editing events that are functionally important could be greater than this estimate if some lowly edited events are functionally important.) This further supports the idea that a fraction of nonsynonymous and 3’UTR editing events are likely functionally important.
While nonsynonymous amino acid changes affect protein sequence, the function of 3’UTR editing is unclear and under studied. 3’UTR regions often play a regulatory role in gene expression [38]. Therefore, we examined gene expression changes when RNA editing is removed, using RNA-seq data of an ADAR deaminase inactive (EA) mutant fly we generated, which contains a point mutation (E374A) that abolishes the catalytic activity of ADAR but retains the protein [39]. This allows us to examine only the editing-dependent function of ADAR, rather than other possible functions of ADAR. We found that the expression levels of 3’UTR-edited genes increased more in the ADAR mutant than genes edited elsewhere (p = 0.016) (Fig 5A). This effect was more dramatic in genes that were edited in more than one site in the 3’UTR (p = 0.004) (Fig 5B). Additionally, we found that the expression levels of genes with high 3’UTR editing levels or constrained 3’UTR sites increased more than those of genes with low editing levels or unconstrained sites, respectively (Fig 5C and 5D). Thus, 3’UTR editing is associated with changes in gene expression, and this further supports the functional importance of 3’UTR editing.
The association of the presence of 3’UTR editing and decrease in gene expression prompted us to hypothesize that RNA editing in the 3’UTR may lead to mRNA degradation. To test this, we compared editing levels from nascent RNA-seq to those from polyA+ RNA-seq in publicly available data4 (S1 Table). Because RNA editing is co-transcriptional [4], one would expect the editing levels of nascent RNA to be lower than those of mature polyA+ mRNA. For CDS sites, as expected, RNA editing levels are indeed lower in nascent RNA than polyA+ mRNA (Fig 5E). On the contrary, 3’UTR editing sites have very similar editing levels in nascent RNA and polyA+ mRNA (Fig 5E). This suggests that, among other possibilities, the edited 3’UTR transcripts may be degraded post-transcriptionally, thus lowering the editing levels in the mature polyA+ mRNA. Further experiments are needed to provide direct evidence for this hypothesis.
We further explored the possible roles of miRNAs in regulating gene expression via differential binding to edited and unedited 3’UTRs [40,41]. We examined if editing sites create or destroy putative miRNA targets for genes with expression differences between WT and catalytically inactive EA flies (log2 Fold change >0.2 or <-0.2, 14 genes in total). We only used miRNAs that are co-expressed in fly male head samples for target prediction. We identified two editing events that create binding sites for miR-282-5p and miR-1017-3p; in both cases, the host genes have increased gene expression levels in the EA mutant compared to the WT flies (Fig 5F). Therefore, RNA editing in the 3’UTR may affect miRNA binding, thus at least partially explaining the gene expression.
In this work, we examined the birth, death, and persistence of editing events as well as the quantitative changes of editing levels during ~60 million years of evolution in the Drosophila genus. We used changes in both RNA primary sequence and secondary structure to predict differences in editing level between species. This highlights the importance of local cis-regulatory changes in the evolution of individual RNA editing events, consistent with the findings of recent studies of RNA editing variation across more closely related Drosophila [29,30], and shows that what is known about editing cis-regulation can account for a large fraction of the variation observed. We gained insights into the cis regulatory architecture of RNA editing and demonstrated how these features can be used to explain and predict the occurrence of an editing event as well as differences of RNA editing across species. For example, unexpectedly, we found that the ADAR binding motif is a feature that is a lot more important to determine whether a site can be edited compared to what level a site is edited (Fig 2D). Our work lays a foundation for future work towards a cis regulatory code of RNA editing.
In addition, our data suggests that RNA editing is selected to be enriched in slowly evolved neuronal genes. While neuronal genes evolve slowly, the nervous system acquires high complexity at least partially through the means of RNA editing. Thus, RNA editing, rather than nucleotide substitution at the genomic DNA level, may be the preferred evolutionary means of fine-tuning neuronal functions.
While this manuscript was under preparation, another study on Drosophila RNA editing evolution was published [42]. Yu and colleagues examined the evolution of RNA editing in the Drosophila genus and found that editing events conserved in at least two members of a gene family were enriched in genes with neurological functions and in regions subject to purifying selection. These conserved editing events tended to cause amino acid changes, which is consistent with the idea that nonsynonymous editing events are more likely functionally important. This is echoed by our analyses that employ a targeted deep sequencing method to achieve much more accurate measurements of RNA editing levels. We both find that highly conserved editing events tend to be nonsynonymous, in neurological genes, and in regions or genes with a lower substitution rate.
In addition to examining coding events, our work delves further into the often ignored non-coding RNA editing events. We found that 3’UTR editing is evolutionarily constrained at the DNA level and highly conserved at the RNA level. This is surprising to us given that the non-coding 3’UTRs are generally not evolutionarily conserved compared to coding regions. While the nonsynonymous editing events may have a more obvious function because of their ability to alter amino acid sequences, our work suggests that many 3’UTR sites are under negative selection. Thus, we identify a large fraction of both nonsynonymous and 3’UTR editing events that are likely functionally important. Our list of candidate functional sites will be valuable for further identification and characterization of truly functional, individual editing sites among the large number of sites that have been recently identified (e.g. [4,12–14]).
Our work further provides evidence of potential functions of 3’UTR editing events. Our finding that 3’UTR editing is associated with mRNA degradation in vivo suggests a new role for RNA editing in the regulation of RNA expression levels. We also pinpoint examples of 3’UTR editing sites that alter miRNA binding, thus affecting gene expression. Our findings would add another element to the array of tools by which RNA editing could fine-tune gene function in complex neurological systems. Further experimental work needs to be done to establish whether 3’UTR editing causes gene expression changes, and to explore the mechanism of action. In sum, our analyses provide strong evidence for a set of functional RNA editing sites and indicate an association between 3’UTR editing and mRNA regulation.
By comparative evolutionary analyses, we found that nonsynonymous and 3’UTR sites are edited at high levels and located in highly or moderately constrained genomic regions. These data suggest that nonsynonymous and 3UTR edited sites are functionally important. However, it should be noted that little is known about the dominance of post-editing phenotypes, and editing divergence is only moderately correlated with the editing level (S11 Fig), so caution should be exercised when using editing level alone as a proxy for fitness or negative selection. The approaches used in this study can be applied to identify functionally important editing events in other species, such as the primates where RNA editing is a lot more abundant particularly in non-coding regions [17].
We obtained RNA-seq data for 13 Drosophila species from the NCBI Sequence Read Archive (SRA) (http://www.ncbi.nlm.nih.gov/sra) and modENCODE project (http://www.modencode.org/). We obtained the Anopheles gambiae whole body RNA-seq data from NCBI SRA. A list of datasets is shown in S1 Table. We obtained D.mel yellow white (yw) strain nascent RNA-seq, ADAR null mutant nascent RNA-seq, yw genomic DNA-seq, ADAR wild type and null mutant RNA-seq from NCBI SRA (GSE37232, GSE42815).
We adopted a pipeline that can accurately map RNA-seq reads to the genome [12]. In brief, we used BWA [44] to align individual RNA-seq datasets to a combination of the reference genome and exonic sequences surrounding known splicing junctions from available gene models (obtained from the UCSC genome browser). We chose the length of the splicing junction regions to be slightly shorter than the RNA-seq reads to prevent redundant hits. After mapping, we used SAMtools [45] to extract uniquely mapped reads, merged uniquely mapped reads of individual datasets from the same sample, and detected nucleotide variants between the RNA-seq data and reference genome. We took variant positions in which the mismatch was supported by ≥2 reads and both base and mapping quality scores were at least 20. We used additional filters to remove wrongly assigned mismatches as previously described [12].We inferred the strand information of the sites based on the strand of the genes. Regions with bidirectional transcription (sense and antisense gene pairs) were discarded. ANNOVAR was used to annotate the editing sites [46].
The LiftOver tool was used to convert genomic positions between different species. Since LiftOver does not provide strand information between two species (for example, a sense strand in one species may correspond to a reverse stand in another species), we obtained the strand information between different species using pairwise alignment data from UCSC genome browser. For the 6 species without genome alignments available on the UCSC genome browser (D.eug, D.tak, D.fic, D.ele, D.kik, and D.bip), the sequences were aligned using Lastz with the following parameters: H = 2000, Y = 3400, L = 4000, K = 2200, O = 400, and E = 30 [47]. Then, these were chained, netted, and converted to axt files using the tools axtChain (with a minimum chain score of 3000), chainPreNet, chainNet, and netToAxt from UCSC [48].
RNA-seq data of the wild-type strain and ADAR null mutant were obtained from two recent studies [4,16]. Sequences were mapped as described above. We examined all identified A-to-G sites that are edited in the wild type strain (defined as having more than two altered reads and editing levels ≥ 1.5%). Sites that do not have altered reads in the ADAR null mutant were considered to be genuine A-to-I RNA editing sites. For sites that have altered reads in both the wild-type and ADAR null mutant, statistically significant editing sites in the wild-type were determined by applying Fisher’s exact test to compare the A-to-G occurrences between the wild-type and ADAR null sample [16]. P values were corrected using the Benjamini-Hochberg method, and a confidence level of 0.05 was used as the cutoff.
We have recently developed a cross-species transcriptome comparison method that accurately identifies exonic RNA editing events [16]. We modified this method and applied it to the transcriptomes of 13 Drosophila species (S1A Fig). RNA variants were called separately for each species as described above. Shared RNA variants that are present in any of the species pairs were obtained using various frequency cutoffs. As expected, the majority of the resulting RNA variants are A-to-G mismatches, indicative of A-to-I editing (S1B Fig). To achieve a high accuracy in editing site calling without a substantial reduction in sensitivity, we chose a frequency cutoff such that the fraction of variants that were A-to-G was at least 80% [26]. In total, we identified a list of 1,564 exonic A-to-I editing sites (Fig 1B, S2 Table).
We next combined this list with three other lists [4,13,16]. We estimated the false positive rates of all studies using RNA-seq data from the D.mel wild type strain and from the Adar null mutant that eliminates RNA editing (S1C Fig). We then removed the sites we found to be false positives to obtain 2,380 exonic editing sites.
We note that we did not include a recently published D.mel editing site list [15] because all our experiments and data analyses had been done before the publication of the study. In addition, the number of novel sites from that study only accounted for 13% of the sites collected in our current list. Therefore, the exclusion of sites from this recent study is unlikely to affect our conclusions.
A list of all species and strains used for mmPCR-seq are listed in S3 Table. All stocks were grown on standard molasses media (Stanford fly media center). Ten whole bodies were collected from adult flies (eclosion + ~5 days) for each sample. The ADAR E374A mutant (“Adar EA”) contains 2 point mutations: C1725T (synonymous) and A1733C (leading to the E374A mutation), was produced using CRISPR reagents described in [49], and is on a y sc background, backcrossed to white Canton-S (CS) for 7 generations. For the RNAseq, stocks were grown at 18°C, and heads were collected from 3 day old male adult flies.
Total RNA was extracted with the RNeasy Kit (Qiagen). After DNase I treatment, 3 ug of total RNA was used to synthesize the cDNA using iScript™ Advanced cDNA Synthesis Kit (Bio-Rad). cDNA was purified with MinElute PCR Purification Kit (Qiagen).
For the ADAR EA mutant RNAseq, total RNA was extracted using RNAdvance magnetic beads (Agencourt), treated using TURBO DNase (Thermo Fisher Scientific), depleted of ribosomal RNA [50], and treated again using TURBO DNase.
We used a microfluidic multiplex PCR and sequencing method [32] to quantify the RNA editing levels of selected sites. We selected sites that are edited at ≥5% editing levels in adult D.mel using RNA-seq data for primer design. Using the D.mel genome, we designed 48 pools of 12 to 15-plex multiplex PCR primers [32], which covered a total of 605 loci, allowing us to examine all of these loci on a single chip. The sizes of the amplicons range from 150 to 350 bp. For distantly related species D.ana, D.pse, and D.vir, we also designed additional multiplex primers, which covered 192, 211, and 232 editing loci, respectively, that could not be amplified using the primers designed for D.mel. All primer sequences are listed in S4 Table.
We next loaded cDNAs and primer pools into the 48.48 Access Array IFC (Fluidigm) and performed target amplification as previously described [32]. PCR products of each sample were then subject to 15 cycle barcode PCR and pooled together. All pools were combined at equal volumes and purified via QIAquick PCR purification kit (Qiagen). The library was sequenced using Illumina HiSeq with 101 bp pair-end reads.
We used the FASTX Toolkit (FASTQ/A short-reads pre-processing tools, hannonlab.cshl.edu/fastx_toolkit) to demultiplex the raw reads. We used Tophat [51] to align the pair-end reads to the corresponding genome.
For editing level quantification, sites covered by ≥50 mmPCR-seq reads were used.
We performed two rounds of targeted RNA sequencing. In the first round, we used mmPCR-seq to amplify and sequence all selected editing loci for all samples from six species with the primers designed for D.mel. For the non-D.mel species, because of the presence of the mismatches between some of the D.mel primers and templates, we expected that the amplification efficiency of different loci would be variable; in addition, some loci could not be amplified. As expected, we found a higher variability in coverage of various loci in non-D.mel data (S2B Fig). Since we are only comparing the ratios of two alleles of the same amplicon, we expect that the amplification efficiency difference should not affect the accuracy of editing level measurement. Both the high reproducibility of biological replicates (S2C Fig) and the general consistency between RNA-seq and mmPCR-seq (S2D Fig) support this notion.
In the mmPCR-seq data, for closely related species D.sim and D.yak, we were able to amplify about 80% and 70% of the selected sites, respectively. For distantly related species D.ana, D.pse and D.vir, we were only able to amplify about 56%, 50% and 48% of the loci, respectively. Therefore for these 3 species, we designed additional multiplex primers for ~210 loci that could be amplified in D.mel, D.sim and D.yak but not in D.ana, D.pse and D.vir. We used these primers for the second round of targeted RNA sequencing. We amplified these loci using a regular PCR machine. All samples were then barcoded, pooled and sequenced in one Illumina MiSeq run with 150 bp paired-end reads.
For editing level divergence related analyses, sites covered by ≥20 reads were used unless otherwise specified. To maximize the number of sites used for analysis, we combined the whole body RNA-seq data with the mmPCR-seq data. For mmPCR-seq, sites with ≥50 reads and editing level differences ≤10% between biological replicates were used. For sites with both RNA-seq and mmPCR-seq measurements, we used the mmPCR-seq measurement. Editing level divergence was defined as 1 –Spearman’s rho, where rho was the correlation between editing levels in two species.
The motif weight matrix was deduced using the nucleotide triplets (the editing site and its immediately adjacent nucleotides) of highly edited sites (≥50%). The motif weight matrices are very similar across species so we used the average motif weight matrix for analysis. Motif scores of each of the 16 possible nucleotide triplets were calculated by FIMO [52] and then scaled to the range of 0–10.
To maximize the number of sites used for “Presence/Absence” site analysis, we combined the male and female whole body data. For an editing site to be considered as absence, we required that the observed A-to-G frequency was less than 1.5% in all datasets with editing level measurements for the site. For an editing site to be considered as presence, we required that the A-to-G frequency was at least 10% in at least one dataset. To estimate if an absent site has evidence of editing, we examined if the editing level of this site was significantly higher than the typical A-to-G sequencing error rate (1%) [32] based on the read coverage of this site using Binomial Test. For example, with D.mel as an anchor species, we found 492 Presence/Absence sites. 90% of the Presence/Absence sites had statistically significant higher editing levels in D.mel (p<0.05, Fisher’s exact test). And 99% of the absence sites had similar editing level to the typical A-to-G sequencing error rate, indicating no evidence of editing.
We used the proximal and distal ECS prediction pipelines previously described [29]. Briefly, to predict proximal ECSs, we predicted the secondary structure of the region within 200 bp of each editing site using the programs partition, MaxExpect, and ct2dot from the RNAStructure package [53], and identified ECS-like sequences with at least 20 bases paired in the stems and a max bulge of 8. The same method was used to predict proximal ECSs in all 6 Drosophila species. To predict distal, intronic ECSs, we predicted conserved ECSs located in intronic regions. We smoothed phastCons scores using a sliding window of 51 bp [34]. We selected regions that were within 2,500 bp of the editing site and at least 20 bases long with a smoothed phastCons score of at least 0.90 (determined using known intronic ECSs). Next, we folded candidate sequences and identified ECSs as with the proximal ECSs. Since the intronic ECS predictions could bias our analyses towards D.mel, the species in which they were identified, we used only the proximal ECSs for our cross-species evolutionary analyses. Both the intronic and proximal ECSs were used for comparisons within the D.mel population.
We used our ECS predictions for two types of analyses to examine how differences in various features of orthologous sites are related to differences in editing level. First, we examined changes of editing in two scenarios. In the “Presence/Absence” scenario, we examined features of sites that were edited (≥10%) in an “anchor” species and unedited (≤1.5%) in another species. As a control, comparisons were also made in the “Presence/Presence” scenario, in which both the anchor and other species were edited (≥10%). Sites with predicted ECSs that passed the filter in the anchor species were used. (The ECS was not required to pass the filter in the other species since it may or may not be present.) Secondly, we predicted the editing level in a second species, given the editing level in the first species. In this analysis, sites with ECS predictions that passed the filter in at least one of the two species were used.
We dissected the dsRNA structures into 8 structural features (free energy, stem length, max bulge size, percent of the stem that is base-paired, distance between editing site and closest stem edge, number of paired bases in the entire stem, and, separately, number of paired bases in the stem downstream and upstream of the editing site). To determine the free energy of the dsRNA stem, we joined the editing stem with the complementary region stem by a 100 base linker of adenosines and ran the RNA secondary structure prediction program fold from the RNAStructure package.
The relationship between editing changes and the ADAR motif and 8 structural features was modeled using random forests [36] (R package: randomForest; parameters: ntree = 10000, mtry = 3, importance = TRUE). The random forests algorithm is particularly resistant to overfitting and uses the out-of-bag error (for each iteration, calculated using the unused samples (roughly one-third)) to obtain an unbiased estimate of the accuracy rate [54].
To predict whether a site is edited (“Presence/Absence”), classification trees were used. (Classification trees are decision tree models in which, at each branch, decisions based on features (ie: RNA structure or sequence) are made to classify items (ie: sites).) To predict the editing level changes (“Presence/Presence”), regression trees, which are decision tree models that use continuous values, were used. The relative importance of the features (accuracy importance for classification, percent increase in mean squared error for regression) was calculated by randomly permuting each feature, taking the average difference between the accuracy of predictions using the permuted data and the accuracy using the non-permuted data, and then dividing by their standard deviation [36]. The percent variation (R2) = 1 –(the mean squared error of the editing level prediction / the variance in editing levels).
For predicting the editing level in one species using the editing level in D.mel without motif and structural features, there was only one feature, so only one was sampled at each split (parameters: ntree = 10000, mtry = 1, importance = TRUE).
To calculate the conservation levels of DNA sequences surrounding the editing sites, we used a 30 bp sliding window size because the editing stem regions are usually 30–50 bp long [24]. For nonsynonymous and synonymous sites, we did not use flanking regions if they fell into a non-CDS region. For 3’UTR sites, we did not use flanking regions if they overlapped with any annotated CDS regions.
To categorize each editing site as highly constrained, moderately constrained or unconstrained, we examined the conservation of the region spanning the editing sites (S_region, 15 bp upstream and 30bp downstream of the editing site) and the more distal flanking region (F_region, 46bp before and after the S_region). We first defined sites with high conservations (average PhastCons score > 0.9) for both S_region and F_region as highly constrained sites. For the remaining sites, we defined sites with a statistically significant higher conservation in the S_region compared to the F_region to be moderately constrained sites, and the ones with similar or lower conservation of S_regions as unconstrained sites. The significance was determined using two-sample Kolmogorov-Smirnov test (fdr corrected p value < 0.05).
We obtained a total of 59 D.mel RNA-seq datasets, including 30 RNA-seq datasets from 30 developmental stages spanning the whole life cycle and 29 tissue-specific RNA-seq datasets in various tissues dissected from various stages (http://www.modencode.org/). For each dataset, we calculate editing levels for editing sites covered by ≥20 reads. For each site, the maximal editing level across all samples was used as the representative editing level of a site.
We defined the age of each D.mel editing event using a combination of the male and female whole body data. We used 5 fly species (D.sim, D.yak, D.ana, D.pse, D.vir) and 1 mosquito species (Anopheles gambiae) representing 6 time points compared to D.mel. We posited that editing events rarely independently originate from different lineages so the age of each D.mel editing site was based on the most distantly related species in which sites were also edited (parsimony principle).
To determine if a site is edited, we required the editing level to be ≥2% in at least one dataset. With this cutoff, 89% of the editing sites have significantly higher editing levels than the typical A-to-G sequencing error rate (1%) [32] (p < 0.05, Binomial Test). To determine if a site is unedited, we required that the site be covered by ≥20 reads and that the editing level be ≤1.5% in all datasets. With this cutoff, <1% of the editing sites have significantly higher editing levels than the typical A-to-G sequencing error rate (p < 0.05, Binomial Test). A site that was not an A at the DNA level was defined as unedited. Only sites that could be well defined as edited or unedited in at least 5 species were used for analysis.
Gene ontology (GO) term analysis was done using DAVID [55]. P-values were corrected for multiple hypotheses testing using the Bonferroni method. All statistically significant GO terms (p ≤ 0.01) with at least 5 genes were shown for all age groups except the D.vir age group for which we showed the top 10 GO terms from the total of 33 significant GO terms.
The ω values of each gene were obtained from Larracuente and colleagues [56]. In brief, all single-copy orthologs in the melanogaster group were obtained. Paralogs were excluded because of difficulties in computationally verifying the accuracy of phylogenies and of alignments. We further selected genes that have orthologs in all fly species compared for analysis. The orthologous gene information was obtained from flybase. Only the melanogaster group was used because divergence at silent sites is too great (saturated) beyond the melanogaster group, which prevents an accurate estimation of dS and thus would erode the power to accurately estimate both rates of evolution (dS and ω) [56]. PAML (Phylogenetic Analysis by Maximum Likelihood) was used to calculate ω [37].
RNA sequencing libraries of the Adar EA mutant and wild type control were produced using the KAPA RNA-Seq Kit (Kapa Biosystems). Libraries were sequenced on an Illumina NextSeq 500 Sequencer using paired-end 75-bp cycles. Reads were mapped using TopHat [51] and DESeq2 [57] was used to obtain gene expression levels. Only genes with editing levels above 5% and at least 20 reads at the editing site were used in the analysis.
We obtained nascent and polyA RNA-seq data for yw male head samples from a recent study [4]. Reads were mapped using BWA to a combination of the reference genome and exonic sequences surrounding known splicing junctions, as described above. Only sites covered by at least 20 reads were used in the analysis.
We used the TargetScan algorithm [58] to predict miRNA binding sites for the unedited and edited form of 3’UTRs using head expressed D.mel miRNAs. miRNAs were obtained from miRBase database Release 20. Male head miRNA expression data was from GSM322543. Only miRNAs expressed in male heads (with >0.1% of the total miRNA reads, 62 miRNAs in total) were used for analysis.
|
10.1371/journal.pgen.1001068 | Epigenetically-Inherited Centromere and Neocentromere DNA Replicates Earliest in S-Phase | Eukaryotic centromeres are maintained at specific chromosomal sites over many generations. In the budding yeast Saccharomyces cerevisiae, centromeres are genetic elements defined by a DNA sequence that is both necessary and sufficient for function; whereas, in most other eukaryotes, centromeres are maintained by poorly characterized epigenetic mechanisms in which DNA has a less definitive role. Here we use the pathogenic yeast Candida albicans as a model organism to study the DNA replication properties of centromeric DNA. By determining the genome-wide replication timing program of the C. albicans genome, we discovered that each centromere is associated with a replication origin that is the first to fire on its respective chromosome. Importantly, epigenetic formation of new ectopic centromeres (neocentromeres) was accompanied by shifts in replication timing, such that a neocentromere became the first to replicate and became associated with origin recognition complex (ORC) components. Furthermore, changing the level of the centromere-specific histone H3 isoform led to a concomitant change in levels of ORC association with centromere regions, further supporting the idea that centromere proteins determine origin activity. Finally, analysis of centromere-associated DNA revealed a replication-dependent sequence pattern characteristic of constitutively active replication origins. This strand-biased pattern is conserved, together with centromere position, among related strains and species, in a manner independent of primary DNA sequence. Thus, inheritance of centromere position is correlated with a constitutively active origin of replication that fires at a distinct early time. We suggest a model in which the distinct timing of DNA replication serves as an epigenetic mechanism for the inheritance of centromere position.
| Centromeres form at the same chromosomal position from generation to generation, yet in most species this inheritance occurs in a DNA sequence–independent manner that is not well understood. Here, we determine the timing of DNA replication across the genome of the human fungal pathogen Candida albicans and find that centromeric DNA is the first locus to replicate on each chromosome. Furthemore, this unique replication timing may be important for centromere inheritance, based on several observations. First, DNA sequence patterns at centromeres indicate that, despite high levels of primary sequence divergence, the region has served as a replication origin for millions of years; second, formation of a neocentromere (a new centromere formed at an ectopic locus following deletion of the native centromere DNA) results in the establishment of a new, early-firing origin of replication; and third, a centromere-specific protein, Cse4p, recruits origin replication complex proteins in a concentration-dependent manner. Thus, centromere position is inherited by an epigenetic mechanism that appears to be defined by a distinctively early firing DNA replication origin.
| Centromeres are essential components of eukaryotic chromosomes required for proper chromosome segregation to daughter cells. Lack of a functional centromere, or the presence of multiple centromeres, renders chromosomes unstable and prone to mis-segregation and breakage. This genome instability is associated with carcinogenesis and can also result in cell death. An intriguing property of most eukaryotic centromeres that remains poorly explained is their mode of inheritance. In principle, the functional identity of a single locus on a chromosome, such as a centromere, requires that locus to have at least one unique property that distinguishes it from other loci on that chromosome. While a primary DNA consensus sequence would be sufficient to define a single locus per chromosome, most eukaryotic centromeres are not defined at the DNA sequence level (reviewed in [1]–[5]). Thus, for instance, centromeres on different chromosomes in any one species do not share primary DNA sequence between them; furthermore, centromeric DNA sequence diverges between closely-related species while centromeric loci remain syntenic; in rare cases, centromere proteins move to new loci that do not normally function as centromeres. These neocentromeres, which remain stable at their new ectopic loci, have been observed in humans as well as in several model organisms [6]–[8].
Rather than a specific DNA sequence, a unique, conserved histone H3 variant, termed CENP-A/CenH3 (Cse4 in yeasts), distinguishes eukaryotic centromeres from the rest of the chromosome and has inspired the majority of current models of epigenetic centromere inheritance. These models propose that centromeric chromatin structure contains the information necessary to form and maintain centromeres at a given locus. One model suggests that CENP-A and histone H3 are expressed and/or deposited at different times during the cell cycle [9]–[12]. Consistent with this, S. pombe CENP-A expression reaches maximal levels just before the G1-S boundary [11] and can be deposited in either G1/early S-phase, or via a different pathway in G2 [12]. A related hypothesis suggests that centromeric DNA might replicate at a distinct time during S-phase, and that this may be coordinated with the timing of CENP-A deposition [13] (reviewed in [14]). However, studies in flies and mammalian cells that utilized microscopy measurements of BrdU incorporation to follow the replication timing of centromeric DNA failed to detect a distinct time of replication at centromeres [9], [15]–[17] and thus such models were abandoned. It is important to note that these experiments may lack the resolution and precision necessary to detect replication events within specific regions of the genome. Replication timing microarrays could provide the necessary precision, however centromeres in most eukaryotes typically span hundreds to thousands of kilobases of highly repetitive DNA and are usually not fully sequenced, thereby obviating such analyses. In contrast, centromeres in the pathogenic yeast Candida albicans exhibit all the hallmarks of epigenetically inherited centromeres, yet are short (∼ 3 kb), simple and fully sequenced [7], [18]–[21], making this yeast an attractive model system for the study of centromeric DNA properties.
In this study, we used C. albicans to study the replication of centromeric DNA. We found that on all chromosomes, the centromere was associated with a replication origin that fires first, and well before all other origins, on that chromosome. Manipulating Cse4 binding by either deletion of a canonical centromere locus or by placing it under the control of a conditional promoter revealed that centromere determinants attract replication components and specify early DNA replication. In addition, we describe a sequence feature of C. albicans centromeres - asymmetrical nucleotide composition - that is indicative of stable replication activity over evolutionary time scales. Using phylogenetic comparison, we provide evidence linking the epigenetic inheritance of centromere position with replication activity. Thus, DNA replication timing can serve as the basis for the inheritance of functional centromeres at specific chromosomal sites, representing a novel mechanism of epigenetic inheritance.
We determined the genome-wide DNA replication timing profile at high temporal and spatial resolution for all eight Candida albicans chromosomes (Figure 1A). Briefly, asynchronous log phase cells were sorted by fluorescence activated cell sorting into G1 and S phase fractions. DNA from 2×106 cells from each fraction was extracted, differentially labeled with fluorescent dyes and hybridized to genomic tiling arrays. The data were smoothed and replication timing was displayed as a function of chromosome position. In these replication profiles, peaks represent replication origins and the height of each peak reflects the relative timing and/or efficiency of replication from that origin. When applied to S. cerevisiae, this method identifies known replication origins at a resolution of ∼5 Kb ([22]; Materials and Methods).
The replication profiles exhibited a striking feature: centromeres were always proximal to a predicted replication origin and remarkably, these centromeric origins were the first to fire on each chromosome. Furthermore, the activation timing of these centromere-associated origins was clearly separate from the activation time of non-centromeric origins (Figure 1B and 1C and Figure S1). The colocalization of replication origins with centromere position is further supported by several lines of evidence. First, two-dimensional DNA gels detected a replication bubble structure, indicative of an active origin, at the chromosome 5 centromere DNA region (CEN5) (Figure 1D). Second, genome-wide chromatin immunoprecipitation (ChIP-chip) experiments of the Origin Recognition Complex (ORC), that is essential for replication initiation, revealed the presence of ORC within the centromere regions of each of the eight chromosomes (Figure 1E and Figure S2). These ORC binding sites were among the strongest in the genome (Figure S2 and data not shown). Third, ChIP followed by quantitative PCR of CEN5 detected ORC localization within CEN5 in correlation with the position of Cse4 (Figure 1F; [7]). Thus, centromeres in C. albicans contain replication origins that are the first to replicate on each chromosome.
The distinct replication timing of the centromeres is manifested in the distribution of origin activation times in the C. albicans genome: non-centromeric origins are largely absent from the beginning of S-phase and are otherwise activated throughout S-phase (Figure 1G). In contrast, in S. cerevisiae (that has genetically-inherited point centromeres), most origins are activated in the beginning of S-phase, at a time that is not distinct from that of centromere-proximal origins (Figure 1H; [23]). Thus, C. albicans achieves genomic replication with significantly fewer early-activated origins relative to S. cerevisiae. However, the paucity of early origins does not result in a longer S-phase in C. albicans compared to S. cerevisiae. This could be explained by a difference in the distribution of active origins with different activation timing along the chromosome: in S. cerevisiae, early origins tend to cluster in close proximity, ultimately resulting in ∼2-fold more active origins per DNA length than in the C. albicans genome. In C. albicans, the spacing of early and late origins is more evenly interspersed (Figure S3). This more regular spacing of origins along the chromosome is presumably more efficient, enabling utilization of fewer origins in C. albicans.
Previous work described a replication timing position effect in S. cerevisiae [23], in which centromeres are located within a region of several tens of kilobases of early replicating DNA (and see below). A similar, albeit significantly stronger effect is evident in C. albicans (Figure 1I) where a region of up to ∼100 kb surrounding the centromere replicates earlier than the genomic average, further emphasizing the distinctiveness of centromere replication timing in C. albicans.
The colocalization of centromeres and early replication origins could be caused by replication origins recruiting centromeric determinants. Alternatively, centromeres could recruit replication determinants. To address the direction of dependency between centromeres and replication origins, we exploited C. albicans strains in which the native chromosome 5 centromere had been deleted and a heritable neocentromere had formed at a different novel locus on the chromosome in each strain [7]. These strains provide a unique opportunity to address the order of this dependency. In four neocentromere strains, the centromere shifted to non-telomeric loci that were not ORC-binding sites in the wild-type progenitor strain (Figure 2A and 2C). This indicates that neocentromeres do not form at pre-existing replication origins.
We then assayed replication timing in a strain that acquired a stable homozygous neocentromere following deletion of CEN5 DNA and homozygosis of the entire chromosome (neoCEN4, strain YJB9929s; see Materials and Methods; [7]). Origin activity at the CEN5 region was completely lost, indicating that origin determinants had resided within the deleted centromeric DNA (Figure 2B). Strikingly, neocentromere formation was associated with the appearance, in the neoCEN4 region, of a new replication origin that was the first to replicate in the chromosome (Figure 2B). Consistent with the idea that a de novo origin formed at neoCEN4, a new ORC binding site in the neoCEN4 region was readily evident by ChIP-PCR only in the strain that had formed a neoCEN and not in its wild-type parental strain (Figure 2C). Of note, the replication timing of origins flanking the centromere and neocentromere was also altered, possibly reflecting a broad position effect exerted by centromeres on the replication timing of flanking origins of replication. This resembles the position effect that mediates generally earlier replication timing within tens of kilobases surrounding conventional centromeres (Figure 1I). The profiles of chromosomes 2 and 3, which would not be expected to be altered by deletion of CEN5, were indeed similar between the neocentromeric and wild type strains (correlation between the profiles r = 0.92−0.95 versus 0.36 for chromosome 5).
To further test the hypothesis that centromeric activity can recruit replication determinants, we used a conditional promoter to manipulate the level of Cse4 expression and assayed the effect of Cse4 levels on the level of ORC-binding at centromeres. Consistent with the idea that centromeres recruit origins of replication, we found that over-expression of Cse4 resulted in increased levels of ORC at CEN5, while repression of Cse4 expression resulted in reduced levels of ORC (Figure 2D). Taken together, we conclude that centromeric determinants confer both the presence of replication origins and a distinct, early replication timing to the loci at which they reside.
As in higher eukaryotes, C. albicans has regional centromeres with different primary sequences at each chromosome [18], [20]. Furthermore, centromeric DNA sequences are highly divergent between strains and related species, mutating at rates higher than those at intergenic regions and synonymous sites [21]. Despite this lack of homology, sequence analysis revealed a common feature among C. albicans centromeres: all of them have a sequence pattern that has been detected previously at replication origin sites of hundreds of bacterial, archeal, viral and organellar genomes that have a single replication origin in their genome [24]. This pattern, indicative of constitutive replication origin activity, is an asymmetric GC skew, in which G nucleotides are more abundant than C nucleotides on one side of the centromeric replication origins, and G nucleotides are less abundant than C nucleotides on the same strand on the other side of the centromere (Figure 3A, 3B, 3D). For A versus T nucleotides, a similar skew pattern was observed, in the opposite direction relative to the GC skew, with a lower magnitude, and with essentially the same position of switch in skew direction. A similar AT skew pattern is present in many of the species with single-origin genomes.
Such skew patterns are inferred to be a consequence of mutations that accumulate in a strand-specific manner: specific substitutions occur at different rates on the leading and lagging strands and the leading and lagging strands switch identities at the point of replication initiation (Figure 3A). Thus, a nucleotide skew pattern implies the presence of a replication origin at the point of symmetry switch [24]. This is consistent with our results detecting the presence of replication origins within centromeric regions. In particular, strong ORC-binding sites reside in very close proximity to the points of skew sign switch (Figure 1E and Figure S2)).
Moreover, skew patterns accumulate over evolutionary time scales, only in cases where a replication origin is consistently active for many generations; hence skews are evident in most genomes with single origins. Conversely, and despite the observation that replication fork asymmetry also causes a strand bias in mutagenesis experiments [25] in S. cerevisiae, skew signals are rarely detected in eukaryotic genomes, presumably because eukaryotic replication origins are not constitutively active at specific loci over long time periods.
Thus, the identification of asymmetrical skew patterns at centromeric replication origins of C. albicans suggests that these origins have been active in most, or all, cell cycles for many generations. Indeed, skew patterns are conserved among C. albicans strains that represent divergence times of 1-3 million years (Figure 3C; [20]). Furthermore, skew patterns are virtually identical at all centromeres in C. albicans and C. dubliniensis (Figure 3D and Figure S9), which diverged from each other ∼20 million years ago and share complete conservation of centromere synteny, yet no conservation of centromere primary DNA sequence [21]. The conservation of skew patterns at centromeres in both organisms suggests that DNA replication activity is intimately associated with the mechanism that has ensured the epigenetic inheritance of centromere synteny.
In scanning the entire C. albicans genome sequence, we found that the strongest skew patterns were at centromeres, but that skew patterns were also identifiable at many telomeres (Figures S4, S5, S6, S7). No skew patterns different than background were seen in any of the loci where neocentromeres have been observed to form (data not shown), indicating that skewed DNA is not an attractor for centromeric proteins. Rather, the lack of skew patterns at neocentromeric loci is consistent with the idea that skews only arise following constitutive origin activity at centromeres over evolutionary time scales.
A DNA sequence pattern characterizing epigenetically-inherited centromeres provides a potential tool for predicting centromere locations in other species. Indeed, we found one distinctive skew pattern per chromosome, providing a prediction of centromere (and origin) locations, in two additional members of the Candida clade, L. elongisporus and C. lusitiniae, as well as in the more distantly related yeast, Yarrowia lipolytica (Figures S8, S9; Table S1). In Y. lipolytica, our approach re-identified the five known centromere locations (and predicts the centromere location of the sixth chromosome) in this species. Remarkably, Y. lipolytica centromeres were identified by searching for chromosomal sequences that function as DNA replication origins and subsequent work showed that, for Y. lipolytica plasmids, replication origin and centromere activity are inter-dependent [26], [27].
S. pombe centromeres also include several replication origins that fire very early in S-phase despite being embedded within heterochromatin [28]. We re-analyzed S. pombe replication timing microarray data (Materials and Methods), and found that CEN1, the only centromere that could be probed at sufficient resolution, is the first locus to replicate on chromosome 1 (Figure 4). Notably, no skew patterns were observed at S. pombe centromeric regions (data not shown). This is likely due to the presence of more than one origin within these regions. We suggest that centromeric skew patterns may be unique to species with smaller regional centromeres. In contrast, in a clade containing the Saccharomyces species (see Figure S9A), genetic, point-centromere positions have become fixed in DNA sequence [29] and do not coincide directly with replication origins. Nonetheless, generally early DNA replication surrounding centromeres (Figure 1H and 1I) could be an evolutionary remnant of replication-timing-dependent, epigenetically-inherited centromeres. Taken together, the association between centromeres and replication origins is conserved at least across a wide range of yeast phylogeny (Figure S9A). Since neither DNA sequence nor sufficient replication timing data are available for centromeric regions of higher model organisms, it remains to be determined whether this association is common among other eukaryotic species.
This study provides the first comprehensive analysis of DNA replication of epigenetically-inherited centromeres. Our results show that centromeres replicate at a distinct time from the rest of the chromosome, centromeric determinants can change the replication time of the loci at which they reside, and the DNA replication properties of centromeres are linked to their epigenetic inheritance over evolutionary time scales. Based on this, we propose a self-reinforcing, positive-feedback loop model, in which centromeric determinants affect DNA replication timing and in turn, a distinct replication time facilitates the recruitment of centromeric determinants to that specific locus (Figure 5). Specifically, centromere site specification by Cse4 nucleosomes recruits factors that mediate replication initiation at a distinct time at the very beginning of S-phase (Figure 2). Furthermore, we speculate that early, distinctive replication timing, in turn enables subsequent deposition of centromere-specific proteins such as Cse4 nucleosomes at the region that is first to replicate, for instance due to elevated levels of these proteins in the very beginning of S-phase [11]–[14]. In particular, it has previously been shown in S. pombe that expression of Cse4 occurs from late-M to G1/S phase, and precedes maximal expression of histone H3 in S-phase [11]. Our calculations suggest that centromeric chromatin is replicated very rapidly- within the first <20 seconds of S-phase in C. albicans (Materials and Methods). Further studies will be required to link the distinct timing of CENP-A deposition and the time of centromere replication.
This self-reinforcing association between centromere determinants and early replication timing does not require a particular underlying DNA sequence, establishing it as an epigenetic system. This epigenetic system determines centromere position, as well as replication timing and the constitutive nature of centromeric replication origins, which our results show for the first time can be determined epigenetically. Early and constitutive replication are directly related to each other, as a locus that replicates first will never be passively replicated and hence will be active in every cell cycle. The combination of constitutive origin activity and the relaxation of sequence constraints that stems from the epigenetic nature of this system, enables the underlying DNA to mutate at elevated rates, explaining previous observations that centromeric DNA mutates at high rates while centromere position remains unaltered [21], [30]. We now show that centromeric DNA in several yeast species mutates with a specific sequence bias due to the influence of a centromere-associated replication origin.
We describe a novel type of epigenetic inheritance mechanism that is directly related to the fundamental mechanism of genetic inheritance in that both depend on DNA replication. Centromeres are maintained at specific chromosomal positions in a highly robust manner that is crucial for ensuring high-fidelity chromosome segregation; thus, this replication-timing-dependent epigenetic inheritance mechanism may confer a level of stability unprecedented for an epigenetic system. Finally, another attractive aspect of this model is that, instead of requiring a specialized mechanism for inheritance of centromeric nucleosomes outside of S-phase [11], [17], it directly connects centromeric nucleosome inheritance with DNA replication, similar to the inheritance of canonical nucleosomes that occurs following passage of the replication fork.
Strains SC5314 (wild type; diploid) and YJB9929s [7] (homozygosed for chromosome 5 subsequent to neocentromere formation) were grown in YPAD media at 30°C. Strain CAKS3b in which Cse4 expression is under the control of the glucose-repressible PCK1 promoter [31] was grown in YPA-succinate to induce Cse4 overexpression and diluted into fresh YPA-succinate or into YPAD to repress expression and grown for 6 hours at 30°C before harvesting cells.
FACS and replication timing microarray experiments and analysis were as previously described [22]. Agilent arrays (2×105 format) were custom designed with probes spaced every 140 bp, on average, across the entire genome (assembly21). The experiment was repeated four times with correlation and autocorrelation values comparable to those previously obtained [22] (data not shown), validating the high quality of the data. Strain YJB9929s was repeated twice with a microarray that included the same probes for chromosomes 2, 3 and 5 only. Microarray data have been submitted to GEO (http://www.ncbi.nlm.nih.gov/geo/) with accession number GSE17963.
The experimental repeats were weighted-averaged using Tukey's biweight method [22] and the averaged data was smoothed using a smoothing spline (as implicated in the Matlab function Csaps), which optimizes the degree of curvature versus deviance from the data, with the parameter chosen determining the weight of these two criteria. The parameters chosen for the different chromosomes were as follows (numbers are the -log10 of the parameters provided to Csaps): ChrR: 16; Chr1,2: 15.75; Chr3: 15.25; Chr4; 15; Chr5,6,7: 14.75 (generally a function of chromosome length). This choice of parameters maximized the similarity to the S. cerevisiae data with respect to the significance of frequency components retained (data not shown), enabling valid comparison of the number of replication origins between S. cerevisiae and C. albicans (Figure S3).
The rDNA and chromosome 5 Major Repeat Sequence (MRS; which is larger than 50 Kb on chromosome 5) loci were treated as data gaps; the corresponding chromosomes were smoothed separately on each side of these gaps (with the same parameter). Scaling the entire genome, rather than individual chromosomes, to 0–100, had little if any effect on the data.
2D gels were performed as previously described [32]. Chromatin immunoprecipitation was as previously described [7] with polyclonal antibodies against the entire ORC complex [33] kindly provided by Stephen P. Bell. PCR was performed with 33 primer pairs interrogating the CEN5 region or 24 primer pairs for the neocentromere region; data was averaged over two experimental repetitions and three consecutive primer pairs. RT-PCR was performed in duplicates using the LightCycler 480 system according to the manufacturer's instructions. ChIP-chip was performed in nine biological repeats according to Agilent protocols and will be described in detail elsewhere together with a complete list of replication origin sites (A.K., H-J.T, L.B. and J.B, manuscript in preparation). Results remained unaltered when normalizing to probe GC content.
Datasets used were: a 65, 75, 85, 95 minutes S-phase time course experiment [34]; a replication timing in HU experiment performed in three repeats[34]; a single experiment using ssDNA mapping [35]; and a two time point (2 h, 4 h) S-phase time course experiment [36]. Chromosomes 2 and 3 had probe gaps of ∼50–80 Kb in the centromere region and thus were not amenable to centromere replication timing analysis. Data from each dataset was weighted-averaged over three consecutive probes and all the time points/repetitions of that experiment and smoothed using spline parameters (as above) of 17 for all experiments and 16 for the HU experiment (use of a parameter of 17 did not alter the result for the centromere). This data analysis approach effectively removes outlier data points that prevented the previous identification of CEN1 as the first locus to replicate on chromosome 1.
To avoid influences of the data smoothing on replication fork progression near origins [22], we linearly interpolated the replication profiles from peaks to valleys. Replication timing was converted to minutes by multiplying replication time in percent by the total duration of S-phase in minutes (Figure 1). The latest replication timing of the loci 3 Kb from each side of the centromere-proximal origin on each chromosome (excluding chromosome 6 as indicated above) was the time to replicate each centromere.
Genome sequences were obtained from the following sources: Candida albicans (assembly 21): from CGD (Candida Genome Database; http://www.candidagenome.org/); Saccharomyces cerevisiae: from SGD (Saccharomyces Genome Database; http://www.yeastgenome.org/); Candida dubliniensis, Schizosaccharomyces pombe: from The Wellcome Trust Sanger Institute (http://www.sanger.ac.uk/sequencing/Candida/dubliniensis/), with gene annotation from GeneDB (http://www.genedb.org/genedb/); Candida albicans strain WO-1, Candida guilliermondii, Candida lusitaniae, Candida parapsilosis, Candida tropicalis, Debaryomyces hansenii, Lodderomyces elongisporus: from the Broad Institute (http://www.broadinstitute.org/annotation/genome/candida_group/MultiHomehtml); Pichia stipitis: from JGI (http://genomejgi-psf.org/Picst3/Picst3homehtml); Yarrowia lipolytica: from Genolevures (http://www.genolevures.org/indexhtml) (with centromere positions from [37]).
GC skew was calculated as (G−C)/(G+C), AT skew as (A−T)/(A+T), and each was smoothed using a smoothing spline with parameters equivalent to smoothing by sliding windows of ∼1.5 Kb. For identifying strong skew patterns in the genome (see Figure S5), we categorized intergenic regions by the combination of length, GC content, and either of several parameters of: GC skew, AT skew, GC content, proximity of the latter to each other (of zero-intersection points for skews and extrema for GC content), and combinations thereof. We then searched for a discrete cluster in any of these parameters.
|
10.1371/journal.pcbi.1002157 | Integrating Quantitative Knowledge into a Qualitative Gene Regulatory Network | Despite recent improvements in molecular techniques, biological knowledge remains incomplete. Any theorizing about living systems is therefore necessarily based on the use of heterogeneous and partial information. Much current research has focused successfully on the qualitative behaviors of macromolecular networks. Nonetheless, it is not capable of taking into account available quantitative information such as time-series protein concentration variations. The present work proposes a probabilistic modeling framework that integrates both kinds of information. Average case analysis methods are used in combination with Markov chains to link qualitative information about transcriptional regulations to quantitative information about protein concentrations. The approach is illustrated by modeling the carbon starvation response in Escherichia coli. It accurately predicts the quantitative time-series evolution of several protein concentrations using only knowledge of discrete gene interactions and a small number of quantitative observations on a single protein concentration. From this, the modeling technique also derives a ranking of interactions with respect to their importance during the experiment considered. Such a classification is confirmed by the literature. Therefore, our method is principally novel in that it allows (i) a hybrid model that integrates both qualitative discrete model and quantities to be built, even using a small amount of quantitative information, (ii) new quantitative predictions to be derived, (iii) the robustness and relevance of interactions with respect to phenotypic criteria to be precisely quantified, and (iv) the key features of the model to be extracted that can be used as a guidance to design future experiments.
| Understanding the response of a biological system to a stress is of great interest in biology. This issue is usually tackled by integrating information arising from different experiments into mathematical models. In particular, continuous models take quantitative information into account after a parameter estimation step whereas much recent research has focused on the qualitative behaviors of macromolecular networks. However, both modeling approaches fail to handle the true nature of biological information, including heterogeneity, incompleteness and multi-scale features, as emphasized by recent advances in molecular techniques. The principle novelty of our method lies in the use of probabilities and average-case analysis to overcome this weakness and to fill the gap between qualitative and quantitative models. Our framework is applied to study the response of Escherichia coli to a carbon starvation stress. We combine a small amount of quantitative information on protein concentrations with a qualitative model of transcriptional regulations. We derive quantitative predictions about proteins, quantify the robustness and relevance of transcriptional interactions, and automatically extract the key features of the model. The main biological novelty is therefore the presentation of new knowledge derived from the combination of quantitative and qualitative multi-scale information in a single approach.
| There have been a number of success stories in macromolecular network modeling during the last decade. Special attention has been paid to dynamical modeling approaches. Among a broad spectrum of strategies, qualitative models and their associated methods have played a central role, allowing modelers to investigate the full space of possible discrete behaviors of several regulatory networks. To that end, a variety of methods for qualitative modeling, analysis and simulation of genetic regulatory networks (GRN) have been proposed since the seminal works of Kauffman [1] and Thomas [2], [3] (see [4] for a review). As they rely on discrete representations of both time and variables, these methods share two main advantages: first, the space of possible states is finite (although possibly large), making it possible to hypothesize about the dynamics of biological regulatory systems despite the lack of kinetic information at transcriptional level. Second, regulatory networks can be built from local experimental observations or knowledge-based information (gene-gene or gene-protein interactions).
Although these approaches provide high-level insights into the functioning of gene networks, they often do not accurately reflect the real dynamics of GRN. Indeed, transitions between states in a GRN may exhibit a stochastic component as observed in [5]. This stochastic signal is closely related to population average behaviors [6]. Consequently, the dynamics of GRNs have a stochastic component which is difficult to observe in real time and to capture in discrete models. This has emphasized the need for probabilistic models and methods for analyzing and simulating GRN. Such probabilistic representations of gene networks are now widespread to complement discrete approaches. The Probabilistic Boolean Network (PBN) approach [7], [8] is one of these. Due to its flexibility and the fact that it can be inferred directly from data, it has been extensively studied over the last decade. In [9], finite state Markov chains are also proven to be useful in dealing with microarray data. It was established that the automatically reconstructed Markov chain gave rise to steady state distributions in accordance with some phenotypic biological observations. This suggests that Markov chain models are capable of mimicking biological behavior. More generally, Markov chain models are usually applied in the following way. First, a model that fits a given set of data is inferred [10], [11]. Then, steady state distributions are computed, giving access to biological information, as they reflect some expected phenotypes [8], [12]. In a final step, important product nodes are exhibited, as they control the steady-state distribution and the phenotype [5], [13], [14]. This latter task gives insights useful in designing new biological experiments, allowing both a better validation of the model and suggesting some therapeutic targets. Although those approaches are very efficient, they mainly rely on the quality of the network reconstruction process, that yields a two sides issue: inferring the “structure” of the gene regulatory network and computing transition probabilities that are consistent with the available data. In concrete terms, the lack of accurate observation datasets on the result of transition in a GRN usually makes the inference of the structure more accurate than the computation of the probabilities [5].
In a quite complementary way, [15], [16] have proven that adding a probabilistic aspect to already qualitatively validated discrete models may help in determining parameters of the qualitative model. To do so, the authors add a probabilistic dimension to a discrete piecewise affine model. They introduce unknown transition probabilities between two states as the ratio of volumes defined by the qualitative parameters of the system. The main novelty of their approach is that they compute the whole set of transition probability matrices leading to given qualitative attractors of the system, instead of selecting a precise matrix as the above-mentioned approach does. This approach allows them to exhibit relations between transition probabilities and important coefficients of the system such as synthesis rates. However, as they use an analytic description of the set of accurate probability matrices, their method is limited to small networks composed of two or three genes.
In the present work, we advance the idea of studying discrete knowledge-based transcriptional “intracellular” regulatory information given by qualitative models within a global probabilistic approach. The main novelty of our approach is that we compute the full set of probability transition matrices that correspond to quantitative “population scale” observations provided by protein time-series measurements. We rely on methods inspired by average-case analysis of algorithms theory [17], [18], making use of Markov chains coupled with transition costs to study statistical properties of pattern matching issues. We design a probabilistic framework allowing population scale observations to be integrated into a qualitative gene expression network assumed to be shared by several individual cells. Our approach should therefore be considered as a bridge between purely discrete modeling approaches and probabilistic simulations. We introduce three main novel features: first, we rely on a strong asymptotic property of Markov chains to fully describe the set of all possible weighted probabilistic networks matching with protein time-series observations. Second, we overcome computational problems as we drastically reduce the size of the model by focusing on slope changes (switch from a variable increase to a variable decrease, for instance) instead of changes in product levels. Third, we develop numerical methods to incorporate a set of suitable Markov chains – all those matching the numerical observations – rather than a single Markov chain that cannot be uniquely determined from the few quantitative observations we have at hand. These three novelties allow us to increase the robustness of our approach while reducing the set of data required to perform the analysis. Concretely, our approach involves first computing a discrete (non-deterministic) description of possible succession of slope variations. This can be deduced from knowledge-based transcriptional information, i.e., either a logical graph or a qualitative event succession like those observed in novel generations of microarrays [19]. This provides us with a graph of transcriptional event transitions. The transcriptional events, arising on the scale of an individual cell, affect the protein concentrations, observed on a population scale. These two scales are related by adding an impact cost for each transition over a given protein concentration. This cost is easily deduced by fixing an arbitrary “natural” degradation rate and by applying an equilibrium principle as follows. Intuitively, in the absence of any information – when all the transition probabilities are chosen to be uniform – the expected protein concentrations will be constant. The next step consists of numerically determining the set of transition probability matrices that fit a global quantitative observed outcome. As an example, we expect the model to fit the time-series quantitative observations of the mean concentration of a single protein over a cell population - in this paper we focused on carbon starvation response in Escherichia coli. We have combined theoretical properties of Markov chains - inspired by symbolic dynamics - with reverse-engineering methods (local inference methods) to describe the full space of weighted Markov chains having the appropriate topological structure and whose global mean outcome fits the time-series curve. Then we investigate the geometric structure of the space of Markov chains to derive biological properties of the system: we derive a ranking of gene interactions with respect to their importance in achieving the considered protein variations. Such a classification is confirmed by the literature. We also accurately predict the quantitative time-series evolution of several non-observed population-cell protein concentrations using only knowledge of discrete gene interactions and very few quantitative observations on a single protein concentration. According to our modeling framework, variations in protein quantities appear to be driven by the dynamical behaviors, qualitatively described, that occur underneath at the gene regulatory scale.
As a major modeling contribution, and in the light of the above assumptions, this paper establishes a relationship between the concentration time series ( i.e., quantitative knowledge) and the qualitative behaviors of the biological system, as modeled by genetic regulatory networks. To that end, two matrices are considered (see Figure 1). Note herein that an exhaustive illustration of following features is proposed in the end of the Method section. The first matrix describes an event transition Markov chain which constitutes the core of the model. It depicts the probabilities (latent variables of the model) that the system will switch from one qualitative “basic behavior” to another, where a qualitative basic behavior means a constant slope for the variation of a product. The structure of the matrix is determined by the current extent of our knowledge of what regulates the system. Its numerical coefficients stand for the mean ratio of trajectories of the system that may cross a given transition. Our reverse engineering method aims at computing these numerical non-zero coefficients. As a companion matrix to this event description, a family of impact matrices is built for each protein involved in the system. Given a protein , the corresponding impact matrix will describe the global outcome of each transition between two events – corresponding to an arrow of the Markov chain – over the concentration of the protein . By way of example, if we assume that the system goes through a transition that activates the mRNA production of a gene , the effect (or “impact”) of this event may be modeled by an increase in the production rate of the protein encoded by , say 20%. Additionally, the effect of this event on all other proteins in the system may be modeled by a decrease in the production rate, a free parameter that we fix to 5%, since they undergo a natural degradation process and are not affected by the event transition. As detailed hereafter, the exact rates that are used are computed so that active and passive degradation have the same average impact during a random process. With these two matrices at hand, average-case analysis properties of Markov chains reveal a relationship between the event transition matrix, the impact matrices and the quantitative evolution of a protein concentration under given stimuli, allowing to establish some relations between observable variables of the model (the observed growth ratio of given proteins) and the latent variables of the model. Roughly, the time-series concentrations of a given protein make it possible to recover the main eigenvalue of the event transition matrix, which can be reformulated to infer times-series concentrations of other proteins, as well as global properties of the system.
A Markov chain is a random process for which the next state depends on the current state only. It is described by a graph over the set of nodes , and edges labeled with probabilities in . Likewise, the random process can be described by a transition matrix . The Markov chain is described as minimal when this matrix is aperiodic and irreducible meaning that for sufficiently large and all vertices , there exists an -length cycle including . A stationary state of the Markov chain represents a numerical distribution of the nodes that does not evolve anymore, which corresponds to the eigenvector of the matrix .
The main goal is to estimate the quantitative asymptotic impact of the Markov chain on a biological product quantity or a generic yield. Biologically, such a quantity is any of the phenotypic measurements that is impacted by the gene regulatory network, i.e., any experimental bio-product concentration that might be inferred from either a cell growth rate or a protein concentration encoded by a gene that belongs to the system. To this end, an impact matrix is linked to the transition matrix of the Markov chain. The impact matrix is the same size as . Zero-coefficients in yield zero-coefficients in the impact matrix. Coefficients of the impact matrix are positive real values that describe the estimated cost of a transition on the change in the phenotypic quantity.
Impact matrices simulate the effect of a Markov process over the global quantity as follows. Let , be two nodes of the Markov chain connected by an edge . Let denotes the probability of this transition and its impact. The elementary cost of the transition over the quantity is defined as . The induced elementary cost matrix is denoted by . The quantity is then said to evolve following a multiplicative accumulation rule from an initial distribution . Its mean value at time – that is, after iterations of the Markov process – ( i.e., the average of the costs of all trajectories of length ) is strongly related to powers of elementary cost matrix, that is . In other words, to compute the mean value of the quantity at step , the elementary cost is multiplied along all paths of length – therefore introducing . Each path is weighted with the probability of starting from its initial node – information given by . The final impact is given by the sum of all these quantities – therefore multiplying by . In particular, as detailed below, such a multiplicative accumulation rule is useful to model the burst effect of a gene regulatory network on a metabolic scale, in which a single mRNA stochastically transcribed produces a burst of protein copy numbers [20]–[23].
When a Markov chain is fully determined and when an impact matrix is given, simple linear algebraic computations allow to calculate the growth rate of the corresponding quantity. The added value of a multiplicative law over a Markov chain relies on its asymptotic behavior, that is proved to be exponential, as stated in Theorem 1. More precisely, a multiplicative accumulation rule follows an explicit log-normal law with explicit mean, variances and estimation of error terms. All these characteristics, such as the growth rate of the exponential, are related to dominant eigenvalues of the elementary cost impact matrix . It should be noted that when the Markov chain reaches a stationary state, the accumulation law itself enters a permanent regime, where its exponential rate is fixed. The error term is also exponential, but with a much smaller growth rate, ensuring that the stationary state of the Markov chain is quickly reached.
(Average case analysis theory for accumulation rules) Let be a minimal Markov chain with transition matrix . A multiplicative accumulation rule with impact matrix asymptotically satisfies a normal law with mean and variancewhere is the dominant eigenvalue of the elementary cost matrix . The other quantities express by means of a generation of the elementary cost matrix, defined by . More precisely, express by means of the dominant eigenvalue of , and are constants corresponding to the dominant eigenvectors of and . There exists such that the error terms and verify and .
Here, the minimality assumption restricts applications to a biological process such that its underlying Markov chain is aperiodic and irreducible; and (ii) for every considered cost matrix, there exists at most one aperiodic trajectory (meaning that the cost evolution is aperiodic through times for this trajectory). Note that in the present work, these assumptions are those that will most restrict the biological referential. For instance, biological systems that display oscillatory behavior are outside the natural range of the approach. Nonetheless, one may overcome this weakness by modeling an input with oscillatory behavior and modeling the steps of the dynamics with independent Markov chains. This modeling device is particularly useful when one aims at modeling the circadian system. For a better illustration, please see below how to build such a Markov chain that describes the behaviors of a gene regulatory network.
Given a set of impact rules and assuming that they all follow accumulation rules, optimization techniques were used to infer a Markov chain fitting all available experimental results – the growth rate of several biological quantities. The identification process was divided into two optimization problems. First, in the exact case, a Markov chain is computed which minimizes the euclidean distance between the growth rates and – see Theorem 1 above – of every impact rule associated with the Markov chain and the objective numerical values provided by the experimental results at hand. Local search algorithms are well suited to such an inference task (see [24] for a review). Here, it is necessary to develop an ad-hoc local search algorithm capable of handling eigenvalues that have only an implicit definition.
In order to take experimental errors into account, we considered a second optimization problem, in which the objective values were defined by an interval of validity. Our goal was to infer a Markov chain such that the growth rate of every impact rule belongs to its objective numerical interval, allowing some sets of valid Markov chains to be defined. These sets were approximated by using a polyhedra, defined as follows. First the local search algorithm was used to find a Markov chain whose growth rates were close to the middle of every objective intervals. This Markov chain defines a point, hereafter called the source point in the sequel, inside the solution set. Some points on the boundary of the solution set were then identified by setting a random direction and using a dichotomy method to find the intersection between the boundary and the line, starting from the source point with the expected random direction. As shown in the results section, the volume provides particularly meaningful information. In both cases, sensitivity analysis was performed by considering the following definition. The function was introduced, standing for the Euclidean distance between the growth rate of all impact rules and their objective numerical values. The sensitivity of a transition is then defined by the modification, in percent, when is modified by 1%. Note that it is closely related to the partial derivative according to variable of the function . The higher is the sensitivity of a transition, the more sensitive is the overall score to small variations of this variable.
The previous theoretical framework can easily be adapted to the biological regulatory networks that display discrete dynamics [25]. Products of the system are gathered in a set and a relevant Markov chain summarizes the dynamics of the system. In order to handle computational issues of reverse engineering, the focus is on shapes of trajectories instead of graph states, formalized as follows.
The main component of the modeling operation are transcriptomic events, i.e., elements of . They describe the possible slopes in the variation of a bioproduct during a time unit (i.e. increasing or decreasing). For instance, , also denoted by , stands for the increase in the transcriptional activity, or mRNA production, of the gene . The two events occurring over a product are denoted by and . It is sometimes useful to add some supplementary biological events such as a complex formation, when the information is available. This increases the accuracy of the model. The Event Transition Graph (ETG) encodes the possible successions of events. Its nodes are given by the set of events. An event targets if, in at least one trajectory of the system, varies with the slope and then 's slope changes to the sign . This graph may be derived easily from a state transition graph such as those produced by logical asynchronous multivalued Thomas mode piecewise linear models [26].
An Event transition Markov chain is an event transition graph endowed with a matrix probability . Biologically, considering a Markov chain means considering an average behavior of the system over a set of different cells. Since the focus is on events only (i.e. successions of changes in the slope variations of products) instead of states, the stationary states of the Markov chain correspond to cell populations where the proportion of cells with increasing/decreasing transcripts is fixed. Therefore, the stationary states of Markov chains do not correspond to stationary states of the biological system (where all transcripts have a stable concentration). In order to avoid misunderstandings, a stationary state of an event transition Markov chain is called a permanent regime.
The Initial state of the Event transition Markov chain depends on the biological process that is studied. Assuming that the cells within a population are not synchronized suggests that the initial distribution of events in the system is uniform. If the cells are forced to be synchronized at an early stage of the experiments, a dedicated initial state describing the forced condition must be taken into account.
It was pointed out that the evolution of one – or several – protein concentrations resumes a multiplicative phenotypic impact of the gene regulatory network [21], [23]. The multiplicative assumption was considered as relevant since the protein concentrations in a single cell follow standard evolution laws which are of exponential nature, similarly to the behaviors of systems governed by multiplicative laws [27]. Let be a gene in the system at hand and its encoded protein. The impact matrix describes the impact of the event transition Markov chain on the protein production. To define this matrix, an active impact scale and a passive impact scale must be introduced. If a given transition impacts a given gene via its mRNA production, we assume that its encoded protein production increases or decreases by the scale . Otherwise the protein rate is assumed to decrease via its natural degradation by the scale . Formally, let be an edge in the Markov chain ( can be any product and is either or ). Reaching state means that the activity of gene changes leading to an active production or degradation of its associated protein . During all other transitions , where does not encode the protein , the system undergoes a natural degradation of protein . The production and degradation rate values are chosen as follows. The passive effect is set as equal to ( i.e., a natural degradation of ). The active degradation coefficient is defined according to the following equilibrium rule. Let (resp. ) be the set of all events associated to an active degradation (resp. production) of the given protein. We first fix all the transitions to be uniform ( i.e., all the probabilities of leaving a given state are equal), and denotes by the steady-state distribution of the associated Markov chain. Protein concentration is stable ifwhere and . This defines a degree two equation. Simple arguments prove that this equation has only one solution smaller than 1 that is assigned to . The active production coefficient is then defined as , the inverse of the active degradation coefficient. Eventually, the impact matrix associated to the protein is fulfilled thanks to the passive effect rate and the passive and active degradation rates.
As the approach is dedicated to prokaryotic systems, a linear relationship between gene activities and their protein concentrations is assumed. This induced a standard evolution law to describe the quantitative evolution of the protein concentrations in the system in accordance with the qualitative events as described by the event transition Markov. More precisely, it was assumed that, as with other modeling studies [23], [27], a protein concentration evolves according to a succession of exponential laws , with . The cutting points are obtained using the available experimental data. The meaning of this succession is that the protein concentration at time is if . Then, for each , , and expresses byIt can be noted here that the concentration of a protein that is only degraded tends to , which is its basal concentration. Assuming it to be null leads to simpler formulas for and .
According to the hypotheses discussed below, we assume that the protein concentration follows a multiplicative accumulation rule in each time interval . Let be the mean duration of a transition. In the permanent regime of , which is reached very quickly, the relation holds. According to Theorem 1, this equation implies that the product is nothing but the dominant eigenvalue of the elementary cost matrix of . Additionally, introduced below equals the constant introduced in Theorem 1.
Taking all into account, the growth rates and required to apply our reverse-engineering methods described below, can be calculated from the protein concentration shape as soon as the mean duration time of a translation has been estimated. To that end, it is assumed that the duration is independent from the studied dynamics, allowing it to be computed from experimental knowledge on passive degradation. We introduce the shortest half-life of amino-acids of the protein of interest – usually available in the literature. According to the N-end rule, as depicted in [28], fixing a passive degradation rate of entails that , which allows an explicit computation of and completes the inference of growth rates.
For the sake of clarity, we propose to illustrate now the modeling method when applied on a simplistic Event Transition Graph (core model). It is composed of two genes that monitor four events as depicted in Figure 2. The graph is also depicted using a transition matrix in which one adds two unknowns (latent variables) for describing a Markov chain: and . To solve the problem in a biological context, one then considers the two following complementary informations:
These informations are then used to infer and and relative probabilities. The inference task is performed by an adhoc matlab script (The complete package and its corresponding tutorial are available in http://pogg.genouest.org). As a general result, several combinations of probabilities satisfy the given constraints. They are depicted in Figure 3. Emphasizing a unique set of probabilities is therefore not possible. Unlike other Markov-like techniques, the Event Transition Markov chain models the impact of the Markov chain behaviors over the production of each protein of the system. We are thus able, for each combination of probabilities that satisfies the constraints, to estimate the protein growth rates in the permanent regime. Indeed, one can describe the distribution of Y protein growth rates for 10,000 probability combinations that satisfy the constraints (Figure 4(A)). This distribution is obviously sensitive to the probabilities. For illustration, the distribution of the protein Y growth rate for 10 000 probability combinations picked randomly is different, as attested when one depicts the difference of random and constrained distributions of Y protein growth rates in Figure 4(B), illustrating the close relations between protein X and Y concentration evolutions. Computing the distribution is not an easy task when one considers more than 3 genes or 6 events. In practice, we then overcome this problem by estimating the mean of each growth rate ( i.e., (prediction) in the case of the Y protein growth rate as presented above), instead of each growth rate distribution. This provides some accurate predictions of protein concentration evolutions.
To illustrate the accuracy of the use of Event Transition Markov chains in a biological context, we propose now to focus the Event Transition Markov chain approach on predicting the behavior of protein concentrations during a period of bacterial stress. D. Ropers and collaborators model the growth phase transition of Escherichia coli after a period of nutritional stress [29]. In particular, their model shows the move from an exponential growth state to stationary growth during a carbon starvation stage. This elegant “switch” is evidenced at gene regulatory level with implications at phenotypic level. This model is based on the qualitative results available in both the literature and gene regulatory experiments as performed by the authors (see Figure 5). Furthermore, the proteins encoded by the genes that interact within the model have been well researched by independent studies [30], [31]. This provides partial quantitative information that may be introduced into the qualitative model.
The original model [29] is given as a system of piecewise affine differential equations. It contains 6 genes and 37 constraints over inequalities and thresholds. This yields a state transition graph of 912 qualitative domains. The corresponding Event Transition Graph was automatically computed by applying the definition introduced in the method section and detailed in Supplementary Text S2. The resulting graph, composed of 22 edges and 11 nodes, is depicted in Figure 6. Note that for the sake of clarity, we manually introduced a component named “complex” that summarizes the effect of cAMP metabolite as depicted in [32]. This node, in accordance to the original model [29], stands for a complexation of the Crp and Cya proteins and the carbon starvation signal. Following our formalization, this component is thus a natural product of , and the signal component. Although the event transition graph roughly summarizes the behaviors of the original qualitative model, it still highlights the major biological properties by its reading. For illustration, the repression of the crp gene by the Fis protein [33] is depicted by an active effect of on . However, the information about crp controlled by two distinct promoters is lost.
As detailed above in the method section, we computed the impact matrices based on bacterial protein production growth rates. This setting appears to be suitable since E. coli can be viewed as a multi-scale system. Indeed, the change in protein concentration shall be considered as a protein scale amplification of events that occurs at the transcriptomic scale that are depicted as protein burst by experiments [20]–[22]. By way of illustration and following the equilibrium rule defined above, in the impact matrix over the Fis protein, the concentration of Fis, denoted by , undergoes a increase for each transition targeting . It suffers from a decrease for all transitions targeting . Finally, it goes through a decrease for all other transitions, reflecting a natural degradation for Fis (see Supplementary Text S2 for a complete description of the impact matrix). This depicts the Event Transition Markov chain.
We used quantitative information about changes in Fis protein concentration to reverse-engineer the transition matrix. Experimental evidence [30] shows that the Fis concentration multiplies by 10 in 80 minutes, during the stationary growth phase (i.e. carbon starvation conditions) and then decreases in the exponential phase (see Figure 7 and Supplementary Text S2 for details). Therefore, the protein concentration curve was approximated by two successive steps (stationary phase, from with until with ) and (exponential phase, from with until with ). The shortest half-life of amino-acids of the protein of interest is estimated as by the literature [28], leading to a mean transition duration of . Applying our inference growth rate procedure – see method section – resulted in the computation of the growth rates for both the accumulation rules corresponding to the stationary phase (, , i.e., ) and the exponential phase (, , i.e., ). Then, the reverse-engineering approach using , , , (see Method section) produced a probability transition matrix that fits the protein growth rates in both stationary and exponential growth phases. By repeating several times this procedure, one obtains a sampling of the set of all probability matrices that fits the given experimental protein growth rates.
Using the transition matrix of the Event Transition Markov chain, we perform several simulations on protein concentrations, as impacted by the gene regulation network. First, the transition matrix was coupled with impact matrices on proteins Fis and Cya to simulate their permanent regimes during the stationary phase. Then, after minutes, it is assumed that the exponential phase is initiated, inducing a change in the structure of the gene regulatory network. This change takes place by adding 2 transitions from the “signal” box on the Event transition Markov Chain which activates and the “complex” compound. Because of the given initial conditions during the exponential growth phase, these transitions were neglected, but not in stationary phase conditions. Then, based on the same matrices (impact and probability transition), new simulations are performed on the evolution of Fis and Cya concentrations. Figure 7 depicts the predicted variations of the Cya and Fis proteins during both phases.
Compared to the available independent experimental results [30], [31], the simulations and experiments are overall significantly similar according to a Pearson correlation test. The transition matrix allows us to compute the quantitative behavior of Cya in both stationary and exponential phases. Based on sparse information about Fis only, the predicted Cya behavior is consistent with the experimentally observed behavior (, p-value) [31], which is a quantitative validation of our model. Notice herein that we also predict the complete time series of Fis (, p-value = ), which confirms the exponential growth rate assumption. As a complementary result, the system remains for only a short time in the transient regime ( i.e., the error made herein when one computes the mean is significantly lower than 1% after 7 minutes, or 20 iterations of the Markov Chain), which backs up our assumption of studying this microbial system in permanent regime in both growth conditions. This confirms the usefulness of our modeling approach for this specific biological system.
In addition to the prediction feature, properties of the Markov chain provide insights into biological system behavior. According to the inference process, the proteins Cya and Crp have the same predicted behavior, as a posteriori confirmed by [34]. Furthermore, the sensitivities associated with the transitions of the Markov chain also represent an appreciation of the impact of a given biological compound. In particular, this demonstrates that, in stationary growth phase, transition is highly constrained. Interestingly, this transition implicitly monitors the CAMP-CRP complex that controls the metabolism of alternative carbon sources [33]. It is closely related to ability to the bacterial system to switch between both growth phases in function of the carbon starvation. Furthermore, Schneider and co-workers [35] suggest that fis is involved in a fine tuning of the homeostatic control of DNA supercoiling. A small change in the supercoiling drastically affects the expression of the gene fis, which is in total agreement with the constraints extracted from the Event Transition Markov chain. We performed a similar analysis over the whole system ( i.e., in both stationary and exponential growth conditions). The most sensitive transitions are reported in Table 1, in which we detail the biological meanings of such interactions. Not surprisingly, fis regulation is one of the corner stone genes of the system, but it might be a natural consequence of the inferring process in our modeling approach. However, with no specific transition matrix inference, gyrAB also emerges as one of the most, if not the most, important gene of the microbial system. Implicitly, this confirms the usefulness of the DNA topology for E. coli under carbon starvation conditions.
Our purpose was to illustrate the strength of coupling Markov models together with accumulation rules to study the dynamics of a gene regulatory network, by focusing on its effects at a larger scale – the quantitative protein scale. We assumed that the production of a protein by a gene that belong to a regulatory network, follows a multiplicative accumulation rule. This implies that a permanent distribution of the protein system will be reached in a very short time. In such a regime, each protein concentration follows an exponential dynamic. The permanent regime may be modified by external events, inducing a short transition to another permanent regime. This paper details why observing such a permanent distribution – possibly several – at the protein level allows us to recover the main probabilistic law that governs the gene regulatory network. The law is thus described by a Markov chain over the succession of transitions at the transcriptomic scale. Very general properties of this Markov chain – average case analysis (see Theorem 1) – allow us to infer the Markov chain from a variety of heterogeneous information, such as qualitative behaviors based on existing models and partial quantitative data. We proposed an efficient algorithm based on this average case analysis to infer the Markov chain. In this method, it must be emphasized that the fundamental interest is to focus on transitions between biological events (slope variations of products during a time unit) instead of state variation as proposed by other state-of-the-art methods. Indeed, this abstraction of the system is required to reduce the size of the Markov chain in order to achieve the inference process.
Having determined this Markov chain allows us to study the main asymptotic properties of the dynamic system: identifying the main transitions implied in the permanent regime and sorting the relevance of transition patterns. All these predictions may be quite easily checked with additional experimentation. Conversely, experimentation allows refinement of the Markov chain inference process. Taken together, mixing the properties of a Markov chain with accumulation rules, provides a tool to determine the quantitative and asymptotic properties of a dynamic system.
For illustration and validation purposes, we computed a Markov chain for the event transitions of the Escherichia coli system in the carbon starvation. The computations were performed by using a gene regulatory network of this process and quantitative data about protein Fis production during the stationary phase. Our predictions of the behavior of Fis during the exponential phase and of Cya protein changes were confirmed by independent experimental observations, which emphasizes the ability of our approach to spread partial quantitative information through an Event Markov chain built from qualitative models. Moreover, our results produce various emerging properties such as (i) the sensitivity of a specific transition within the Markov chain or (ii) the quantitative prediction of gene products that are not directly optimized during the simulation. All these features reinforce our interpretation of the global quantitative behaviors of the natural system as modeled.
From a technical viewpoint, the main interest of this approach is as follows: it is not necessary to build quantitative differential dynamic systems that need accurate and complex parameter estimations. Our method uses the results of several available observations to recover the main characteristics of the dynamics (its exponential ratio) and to export several dynamic and biological features. Such probabilistic-like reasoning shall be considered as complementary to formal verification techniques used for validating the qualitative properties of a system [29].
Other recent methods also use probabilistic techniques for studying gene regulatory networks [7], [9], [36]. However, their main purpose is to embed a deterministic model with probabilities. Their main analyses therefore focus on estimating impacts of variation. Probability matrices are computed to represent experiments accurately. Finally, transition probability matrices are used to compute permanent distributions. We argue that our approach is complementary since our average case analysis theory allows us to emphasize emerging properties of the system. Relations between the two scales of observations allow us to exhibit constraints between the gene regulatory network and protein observations. Eventually, this process elucidates transition probabilities that did not come to light with other available methods.
A weakness of our approach relies on the fact that the Markov Chain inference process is based on knowledge of a full qualitative gene regulatory network [4]. This shortens the range of application of our method since, nowadays, relatively few biological systems are described at this level of abstraction. However, this flaw will be moderated by the fact that the gene regulatory network is used only in order to build a global frame of the event transition Markov chain, which is much more abstracted and smaller that the gene regulatory dynamics description. It is reinforced by our main approach which is to build the Markov chain automatically from biological assumptions – either from the literature or experiments such as microarrays.
Another weakness lies in the assumption of a linear relationship between gene activity and the production of the corresponding protein (relevant for a microbial system only). To avoid such a restriction, one must build novel accumulation rules based on other biological abstractions – metabolic and environmental phenotypes are the most natural candidates here. Extending the construction of event transition Markov chain to the models containing reactions instead of qualitative regulations – for instance, signaling networks – is also under study to extend the range of application of our approach. A final range of future works relies on extracting more precise properties from the Markov chain description of a given dynamic system. Such studies shall initially focus on the interpretation of the concentration joint law, standing as a correlation coefficient between time-series observations. They will also investigate the use of these Markov chains to isolate experimental noise from the noise inherent to the chaotic properties of the system. This would provide an estimation of measurement qualities. Finally, average case analysis can be performed on a class of probabilistic models that is much larger than Markov chains. This would allow us to deal with Markov chains that may handle slight variations over the course of times, eventually studying the adaptation of the model behaviors under given environmental variations.
|
10.1371/journal.ppat.1007283 | Prions activate a p38 MAPK synaptotoxic signaling pathway | Synaptic degeneration is one of the earliest pathological correlates of prion disease, and it is a major determinant of the progression of clinical symptoms. However, the cellular and molecular mechanisms underlying prion synaptotoxicity are poorly understood. Previously, we described an experimental system in which treatment of cultured hippocampal neurons with purified PrPSc, the infectious form of the prion protein, induces rapid retraction of dendritic spines, an effect that is entirely dependent on expression of endogenous PrPC by the target neurons. Here, we use this system to dissect pharmacologically the underlying cellular and molecular mechanisms. We show that PrPSc initiates a stepwise synaptotoxic signaling cascade that includes activation of NMDA receptors, calcium influx, stimulation of p38 MAPK and several downstream kinases, and collapse of the actin cytoskeleton within dendritic spines. Synaptic degeneration is restricted to excitatory synapses, spares presynaptic structures, and results in decrements in functional synaptic transmission. Pharmacological inhibition of any one of the steps in the signaling cascade, as well as expression of a dominant-negative form of p38 MAPK, block PrPSc-induced spine degeneration. Moreover, p38 MAPK inhibitors actually reverse the degenerative process after it has already begun. We also show that, while PrPC mediates the synaptotoxic effects of both PrPSc and the Alzheimer’s Aβ peptide in this system, the two species activate distinct signaling pathways. Taken together, our results provide powerful insights into the biology of prion neurotoxicity, they identify new, druggable therapeutic targets, and they allow comparison of prion synaptotoxic pathways with those involved in other neurodegenerative diseases.
| Prion diseases are a group of fatal neurodegenerative disorders that includes Creutzfeldt-Jakob disease and kuru in humans, and bovine spongiform encephalopathy in cattle. The infectious agent, or prion, that transmits these diseases is a naked protein molecule, the prion protein (PrP), which is an altered form of a normal, cellular protein. Although a great deal is known about how prions propagate themselves and transmit infection, the process by which they actually cause neurons to degenerate has remained mysterious. Here, we have used a specialized neuronal culture system to dissect the cellular and molecular mechanisms by which prions damage synapses, the structures that connect nerve cells and that play a crucial role in learning, memory, and neurological disease. Our results define a stepwise molecular pathway underlying prion synaptic toxicity, which involves activation of glutamate neurotransmitter receptors, influx of calcium ions into the neuron, and stimulation of specific mitogen-activated protein kinases, which attach phosphate groups to proteins to regulate their activity. We demonstrate that specific drugs, as well as a dominant-negative kinase mutant, block these steps and thereby prevent the synaptic degeneration produced by prions. Our results provide new insights into the pathogenesis of prion diseases, they uncover new drug targets for treating these diseases, and they allow us to compare prion diseases to other, more common neurodegenerative disorders like Alzheimer’s disease.
| Prion diseases are a group of fatal, infectious neurodegenerative diseases affecting humans and animals. The infectious agent, or prion, is composed of PrPSc, a conformationally altered form of a normal, cell-surface glycoprotein designated PrPC. Prions propagate themselves by a highly specific templating process in which PrPSc molecules impose their unique, β-sheet-rich conformations on endogenous PrPC substrate molecules [1–4]. Consistent with this model, PrP knockout mice, in which PrPC expression is absent, are completely resistant to prion infection [5, 6]. Moreover, these mice do not display symptoms of a prion disease [7], indicating that the disease phenotype is due primarily to a gain-of-function attributable to PrPSc or a related toxic species, rather than to a loss of the normal function of PrPC. Therefore, it is important to understand the molecular mechanism of PrPSc neurotoxicity.
Strikingly, the question of prion neurotoxicity has received relatively little attention in the field, in comparison to the extensive body of research that has been published on prion infectivity, propagation, and transmission. An important clue to the underlying mechanism is the observation that neurons that do not express endogenous PrPC are relatively resistant to the toxic effects of exogenously supplied PrPSc [8, 9]. This result suggests that a critical neurotoxic signal is generated as part of the process by which endogenous cell-surface PrPC is converted into PrPSc, and in the absence of PrPC, this signal is not produced. Consistent with a role for PrPC as a neurotoxic mediator, there is evidence that prion disease progression in mice is characterized by two, mechanistically distinct phases: rapid accumulation of PrPSc and infectivity, followed by slower development of neuropathology and clinical symptoms over a time course that is inversely related to expression levels of PrPC [10, 11]. Although there are a number of studies suggesting signal-transducing activities for cell-surface PrPC [reviewed by 12], the pathways by which its interaction with PrPSc produces neurotoxic signals remain mysterious.
Synaptic loss is a common theme in many neurodegenerative disorders [13, 14]. In prion diseases, neuropathological and in vivo imaging studies in infected mice suggest that synaptic degeneration begins very early in the disease process, predating other pathological changes, and eventually contributing to the development of clinical symptoms [15–22]. However, there is very little mechanistic understanding of this process, due largely to the absence of suitable cell culture models amenable to experimental manipulation.
To address this gap, we previously established a novel neuronal culture model, using which we showed that PrPSc induces rapid retraction of spines on the dendrites of hippocampal neurons [23]. Importantly, this effect is entirely dependent on expression of endogenous PrPC by the neurons, consistent with the previously demonstrated role of PrPC as an essential transducer of PrPSc toxicity. Dendritic spines are the contact sites for most excitatory synapses in the brain, and they undergo constant morphological remodeling during development, learning, and memory formation [24, 25]. Therefore, spines are an important locus for the pathogenesis of neurological diseases, particularly those involving symptoms of dementia.
Here, we have used cultured hippocampal neurons to dissect, using specific pharmacological inhibitors as well a dominant-negative kinase mutant, the mechanism of PrPSc-induced synaptotoxicity. Our data establish a synaptotoxic signaling pathway involving, in stepwise sequence, activation of NMDA and AMPA receptors, calcium influx, stimulation of p38 mitogen-activated protein kinase (MAPK), and depolymerization of actin filaments in dendritic spines. Blocking any one of these steps prevented dendritic spine retraction in response to PrPSc, and could, in some cases, even restore normal morphology to spines that had already degenerated. Taken together, our results provide powerful insights into the biology of prion neurotoxicity, they identify new, druggable therapeutic targets, and they allow comparison of the synaptotoxic pathways underlying prion diseases with those responsible for other neurodegenerative disorders like Alzheimer’s disease.
Previously, we showed that treatment of cultured hippocampal neurons for 24 hrs with purified PrPSc, but not control preparations, induced a dramatic retraction of dendritic spines, an effect that was entirely dependent on expression of endogenous PrPC by the neurons [23].
Before embarking on pharmacological studies, we undertook several experiments to characterize the synaptotoxic effects of PrPSc in neuronal culture, and to gain further insight into the underlying cellular mechanisms. First, we wondered whether the dramatic changes in dendritic spine morphology caused by PrPSc were accompanied by alterations in the efficiency of synaptic transmission. To test this possibility, we used patch clamp recording to measure the amplitude and frequency of miniature excitatory postsynaptic currents (mEPSCs) in hippocampal neurons treated with PrPSc. mEPSCs, which are recorded in the presence of TTX to block action potentials and picrotoxin to block GABA-evoked inhibitory currents, are a measure of spontaneous synaptic currents evoked by glutamate. We found that treatment of hippocampal neurons with purified PrPSc, but not with mock-purified material from uninfected brains, caused a marked reduction in mEPSC frequency, and a less pronounced but statistically significant decrease in mEPSC amplitude (Fig 1A–1C). These effects were not observed in neurons derived from PrP knockout mice (Prn-p0/0), demonstrating that the functional as well as the morphological effects of PrPSc on synapses are entirely PrPC-dependent (Fig 1D–1F). Of note, PrPSc did not significantly affect the frequency or amplitude of miniature inhibitory postsynaptic currents (mIPSCs) (Fig 1G–1I). This result indicates that PrPSc is strikingly selective in its effects, targeting primarily excitatory and not inhibitory synapses.
For the electrophysiology experiments shown in Fig 1, we used a culture system that was slightly different from the one used to monitor dendritic spine retraction. In this system, neurons were grown at a higher density on the same substrate with supporting astrocytes, an arrangement which improves neuronal integrity during patch clamp recording. We confirmed that, when recordings were done on neurons cultured at low density over an astrocyte feeder layer, PrPSc also caused a reduction in mEPSC amplitude and frequency (S1 Fig).
The reduction in mEPSC amplitude and frequency caused by PrPSc could be due to effects on either presynaptic processes (e.g. synaptic release) or postsynaptic characteristics (e.g., number and distribution of active zones). To distinguish presynaptic from postsynaptic effects, we immunostained neurons with antibodies to GluR1, an AMPA receptor subunit present in the postsynaptic density, and synaptophysin, a presynaptic marker. When neurons were treated with purified PrPSc, but not mock-purified material, there was a dramatic loss of GluR1 staining, consistent with the retraction of dendritic spines revealed by fluorescent phalloidin staining (Fig 2A–2J and 2U). In contrast, there was no statistically significant loss of synaptophysin staining after PrPSc treatment (Fig 2K–2T and 2U). These data demonstrate that PrPSc exerts a highly selective effect on postsynaptic elements, with no detectable effect on presynaptic structures, even in the face of massive morphological changes in dendritic spines.
We also found that PrPSc treatment did not change the number of inhibitory synapses, as shown by staining for gephyrin (a postsynaptic anchoring component for glycine and GABAA receptors) (S2 Fig). This result is consistent with the inhibitory effect of PrPSc on mEPSCs and not mIPSCs (Fig 1). We conclude that PrPSc primarily targets the postsynaptic elements of excitatory synapses in the neuronal culture systems we are using.
Glutamate receptor-dependent excitotoxicity contributes to the pathogenesis of many neurodegenerative diseases [26–30]. The selective effect of PrPSc on excitatory synapses suggested the possibility that ionotropic glutamate receptors might play a role in the degeneration of dendritic spines seen in this system. To test this possibility, we treated hippocampal cultures with purified PrPSc, or with mock-purified material form uninfected brains, in the presence or absence of NMDA or AMPA receptor blockers (see Table 1 for a list of pharmacological inhibitors used in this study). We observed that the competitive AMPA receptor antagonist CNQX, as well as the uncompetitive NMDA channel blocker memantine, prevented PrPSc-induced retraction of dendritic spines (Fig 3). Spine number in PrPSc-exposed cultures in the presence of these drugs was not statistically different from cultures exposed to mock material from uninfected brains.
Glutamate receptor-mediated excitotoxicity is often accompanied by an influx of Ca2+ ions via NMDA receptors. To test whether this mechanism was operative during PrPSc synaptotoxicity, we used the calcium-sensitive dye Fluo-3 to image intracellular calcium levels in neurons treated with PrPSc. We found that purified PrPSc, but not mock-purified material, caused a significant increase in intracellular calcium (Fig 4). The effect of PrPSc on Ca2+ levels was absent in neurons derived from Prn-p0/0 mice lacking PrPC, and was completely blocked by memantine. Since voltage-gated calcium channels (VGCCs) are also major mediators of calcium influx into dendritic spines, we tested the effect of inhibitors of the major classes of VGCCs (R, T, N, P/Q, and L type). None of these inhibitors had a statistically significant effect on PrPSc-induced reduction in dendritic spine numbers, with the exception of lomerizine (an L-type VGCC inhibitor), which had a partial protective effect (S3 Fig).
We conclude from these data that activation of NMDA and AMPA receptors plays an essential role in dendritic spine retraction induced by PrPSc, and that the effect of PrPSc is accompanied by Ca2+ influx, primarily via NMDA receptors.
Mitogen-activated protein kinases (MAPKs) are important signal transducers downstream of many kinds of intracellular and extracellular stimuli [31], including stressful stimuli like excitotoxicity [32]. In mammals, the MAPKs are grouped into three main families, ERKs (extracellular-signal-regulated kinases), JNKs (Jun amino-terminal kinases), and p38/SAPKs (stress-activated protein kinases) [31]. We tested whether any of these MAPK families are involved in PrPSc synaptotoxicity. We observed that a p38 MAPK inhibitor (SB239063), which targets all four isoforms (α, β, γ, δ), effectively prevented spine retraction caused by PrPSc, while pan-isoform inhibitors of ERK and JNK were without effect (Fig 5). The inhibitors alone had no effect on spine number. In parallel with its ability to block the effects of PrPSc on spine morphology, we found that the p38 MAPK inhibitor prevented the PrPSc-induced reduction in mEPSC frequency and amplitude (Fig 6). Taken together, these data indicate that p38 MAPK plays an essential role in mediating the toxic effects of PrPSc on synaptic structure and function in this system.
Mammalian p38 MAPK has four isoforms (α, β, γ, and δ) [33]. p38γ is most highly expressed in skeletal muscle, and p38δ in testis, pancreas, kidney and small intestine [34]. Thus, p38α and p38β are the isoforms most likely to be involved in neuronal signaling. The α and β isoforms are 75% identical, and both are inhibited by the compound SB239063 used in Figs 5 and 6. To determine which of the two p38 isoforms is involved in PrPSc synaptotoxicity, we utilized the p38α -specific inhibitor VX745. We found that VX745 completely blocked the effects PrPSc on dendritic spine number and mEPSC properties, suggesting that p38α is one of the isoforms involved (S4 Fig). The absence of a p38β-specific inhibitor precluded specific testing of this isoform.
Because abnormalities in dendritic spines and synaptic transmission are early effects of PrPSc, which occur well before large-scale changes in neuronal morphology or loss of neuronal viability, we wondered whether the effects of PrPSc might be reversible by treatment with a p38 MAPK inhibitor. To test this possibility, we treated neurons with PrPSc for 24 hrs, at which point most of the dendritic spines were retracted (Fig 7A). We then exposed the neurons to PrPSc for additional 24 hrs in the presence of a p38 MAPK inhibitor (SB239063) or vehicle control, after which cultures were fixed and assessed for dendritic spine morphology with fluorescent phalloidin. Amazingly, we found that the p38 MAPK inhibitor was able to reverse the dendritic spine retraction that had accrued during the first 24 hrs of PrPSc treatment (Fig 7C), compared to the cultures treated with vehicle (Fig 7B). Quantitation of spine number under the three conditions is shown in Fig 7E. These data indicate that the extensive dendritic spine abnormalities induced by PrPSc are reversible by p38 MAPK inhibition, at least within a 48 hr time window.
To confirm the results obtained with pharmacological inhibition of p38 MAPK, we employed a genetic method to suppress signaling through this pathway, which makes use of a dominant-negative form of p38α MAPK (T180A/Y182F, referred to as p38AF). This double-mutation in the activation loop of the kinase prevents phosphorylation by upstream kinases, and has a dominant-negative effect on the activity of co-expressed wild-type p38, thereby significantly attenuating signaling [35]. We prepared hippocampal neurons from mice that were heterozygous for the p38AF allele. This method of reducing p38 signaling avoids the embryonic lethal phenotype that results from complete germline inactivation of the p38 MAPK gene [36]. We found that neurons prepared from p38AF mice were morphologically comparable to WT neurons, but were almost completely resistant to the dendritic spine retraction effect of PrPSc (Fig 8). This experiment, which depends on constitutive down-regulation of p38 signaling, complements the previous experiments, which involved acute, pharmacological inhibition of p38 MAPK activity.
To provide biochemical evidence for activation of the p38 MAPK pathway in response to PrPSc, we utilized an immunocytochemical approach in order to detect localized changes in p38 phosphorylation. We doubly stained PrPSc-treated cultures with antibodies to phospho-p38 and total p38, and then imaged the ratio of the two signals in the region of dendritic spines, as marked by phalloidin staining. We found that the amount of phosphorylated p38 in dendritic spines was increased after 1 hr of PrPSc treatment, and remained elevated after 24 hrs in the region of collapsed spines (Fig 9).
Known downstream targets of p38 MAPK include MAPK-activated protein kinase (MAPKAP or MK), MSK and MNK [reviewed in 37]. To determine which of these kinases is downstream of p38 MAPK in the PrPSc synaptotoxic pathway, neurons were treated with PrPSc in the presence or absence of specific inhibitors. We found that a pan MK inhibitor (CAS1186648), which targets all three isoforms (2, 3, and 5), effectively prevented PrPSc-induced retraction of dendritic spines, while pan-isoform inhibitors of MSK and MNK had no significant effect (Fig 10). None of the three inhibitors alone had a significant effect on spine number. We conclude from these results that MK2, 3, or 5 are potential targets of p38 MAPK in the pathway leading to dendrite retraction caused by PrPSc. Of these, MK2 and MK3 are the most likely isoforms to be involved, since MK5 is expressed primarily in the heart, while MK2 and MK3 are ubiquitously expressed [38, 39]. MK2/3 are both substrates for the α isoform of p38 MAPK.
One possible mechanism by which inhibitors of MAPK pathways could prevent PrPSc-induced synaptotoxicity would be by inhibiting conversion of endogenous neuronal PrPC to PrPSc, on the assumption that newly formed PrPSc is the primary trigger for synaptic degeneration. To address this possibility, we tested whether these inhibitors affected PrPSc formation in scrapie-infected N2a (ScN2a) cells. We found that neither the p38 MAPK inhibitor SB239063, nor the MK2/3/5 inhibitor CAS1186648, had a significant effect on the levels of protease-resistant PrPSc in ScN2a cells after 7 days of treatment (S5 Fig). As a positive control, the known anti-prion agent Congo red [40, 41], dramatically reduced PrPSc levels under the same conditions. None of the compounds had a significant cytotoxic effect, as indicated by measurement of the total amount of cellular protein. These results support the notion that the p38 MAPK and MK inhibitors block PrPSc synaptotoxicity by interfering with signaling pathways linked to synaptic integrity, rather than by reducing the amount of the toxic agent (PrPSc).
Actin is abundant in dendritic spines and has been shown to regulate spine morphology [42]. In addition, there is evidence that most signaling pathways linking synaptic activity to spine morphology influence local actin dynamics [43, 44]. To address the role of actin in PrPSc-induced alterations in dendritic spine morphology, we used SiR-actin, a fluorogenic, cell-permeable peptide derived from jasplakinolide that both stabilizes and labels F-actin [45]. Neurons were treated with SiR-actin alone or in combination with PrPSc, after which spines were visualized using SiR-actin fluorescence. For comparison, we exposed cultures to taxol, a microtubule-stabilizing agent, with or without PrPSc, after which spines were visualized using fluorescent phalloidin staining. We found that SiR-actin, but not taxol, prevented PrPSc-induced spine retraction (Fig 11A–11E).
In a second experiment, we treated cultures with SiR-actin in the presence of PrPSc or mock-purified material, and then stained them with an antibody to the AMPA receptor subunit GluR1. We observed that SiR-actin prevented loss of GluR1 staining in response to PrPSc (Fig 11F–11H), parallel to the effect of this compound on spine retraction monitored using SiR-actin fluorescence (Fig 11A, 11B and 11E).
Consistent with these microscopic imaging results, electrophysiological recordings showed that SiR-actin prevented the decreases in mEPSC amplitude and frequency caused by PrPSc treatment (Fig 11I–11K).
Taken together, these data demonstrate that actin dynamics play an important role in the morphological changes in dendritic spines induced by PrPSc, and that stabilizing actin filaments can prevent these changes.
Activation of the UPR has been suggested to play a role in the pathogenesis of prion diseases, and blocking this pathway has been shown to produce a therapeutic benefit [46–48]. To address whether the UPR plays a role in PrPSc-induced synaptotoxicity in our system, we treated cultured neurons with an inhibitor of PERK kinase (GSK2606414) or an activator of eIF2B (ISRIB [49, 50]), both of which reverse the eIF2α-mediated translational repression that occurs during the UPR. Neither compound had a significant effect on spine retraction induced by PrPSc (S6 Fig), suggesting that the UPR, in particular the translational repression arm mediated by PERK and eIF2α, does not play a role in PrPSc synaptotoxicity in this system.
It has been proposed that PrPC is a cell-surface receptor for Aβ oligomers, which mediates some of the neurotoxic effects of these assemblies, including loss of dendritic spines [51–55]. There is evidence that binding of Aβ oligomers to PrPC triggers a signaling pathway involving mGluR5 and Fyn kinase, and that preventing activation of these molecules using specific inhibitors prevents Aβ neurotoxicity and ameliorates neurological symptoms in mice [56–59]. Since the synaptotoxic effects of both PrPSc and Aβ are dependent on the expression of PrPC as a cell surface receptor in target neurons, we asked whether both toxic aggregates activated the same cellular pathways downstream of PrPC.
To investigate this question, we treated neurons with either Aβ or PrPSc in the presence or absence of the mGluR5 inhibitor, MPEP, or the p38 MAPK inhibitor, SB239063. Consistent with previously published studies [57], Aβ oligomers alone caused a significant reduction in dendritic spine number (Fig 12A and 12B), an effect that was dependent on expression of PrPC (S7 Fig). MPEP completely blocked this effect (Fig 12C), also in agreement with published results [56]. In contrast, MPEP had no influence on PrPSc-induced retraction of dendritic spines (Fig 12E and 12F). Moreover, the p38 MAPK inhibitor, which completely blocked PrPSc synaptotoxicity (Fig 5), had no significant effect on Aβ oligomer-induced dendritic spine loss (Fig 12D). Taken together, these data suggest that Aβ oligomers and PrPSc trigger different neurotoxic signaling pathways downstream of a common cell-surface receptor, PrPC.
Although the molecular templating process underlying the infectivity of prions is now well understood, the mechanisms by which prions cause neurodegeneration, in particular, damage to synapses, remain poorly understood. In a previously published study [23], we established a neuronal culture system that recapitulates one of the earliest events in prion synaptotoxicity, PrPSc-induced retraction of dendritic spines. In the present work, we have exploited the simplicity of this system to dissect the cellular pathways underlying the toxic effects of PrPSc on synapses. Our results uncover a multi-step signaling cascade that begins with binding of PrPSc to PrPC on the cell surface, and is followed by activation of NMDA and AMPA receptors, calcium influx, stimulation of the stress-inducible MAPK, p38, and finally collapse of the actin cytoskeleton, retraction of dendritic spines, and a decrease in excitatory neurotransmission (Fig 13). This work provides new insights into the mechanisms of synaptic degeneration in prion diseases, it identifies novel molecular targets for treatment of these disorders, and it allows comparison with pathologic mechanisms operative in other neurodegenerative disorders such as Alzheimer’s disease.
While our previous study focused on retraction of dendritic spines as the main morphological consequence of PrPSc toxicity [23], the results presented here provide a more complete picture of the underling cellular events. We have shown that PrPSc exhibits striking functional and morphological specificity in its synaptotoxic effects: it damages excitatory and not inhibitory synapses, and it targets postsynaptic and not presynaptic sites. This selectivity may reflect the presence on dendritic spines of glutamate receptors capable of initiating excitotoxic processes (see below), or the presence of a higher concentration of specific intracellular signaling molecules in these locations. We have also shown that spine loss is accompanied by a reduction in postsynaptic AMPA receptors, and by a decrement in mEPSC frequency and amplitude, suggesting that functional changes in synaptic transmission are early indicators of prion neurotoxicity. Although many studies have emphasized neuronal death as an important feature of prion and other neurodegenerative diseases, our results emphasize the importance of looking specifically at synaptic dysfunction and loss in order to understand the earliest events in the pathological process.
Our evidence suggests that synaptotoxic signaling induced by PrPSc is initiated by its interaction with PrPC on the cell surface (Fig 13). All of the synaptotoxic effects of PrPSc observed in our system require expression of endogenous PrPC by the target neurons [this work and 23]. Moreover, neurons from transgenic mice expressing N-terminally deleted forms of PrPC (Δ23–31 and Δ23–111) are resistant to PrPSc toxicity, pinpointing an essential role for these residues in the conversion and/or signal-transduction processes [23]. The requirement for cell-surface PrPC in our system is consistent with the documented role of PrPC in mediating the neurotoxic effects of PrPSc in vivo [8, 9]. We have hypothesized that neurotoxic signals may be generated either by the initial binding of PrPSc to PrPC, or by the subsequent conversion of PrPC into nascent PrPSc on the cell surface [23]. Since both PrPC and PrPSc are attached to the plasma membrane by a GPI anchor, production of intracellular signals would presumably require their interaction with partners that are transmembrane proteins, for example ion channels [see below and 56, 60], adhesion molecules [61], or receptors [62].
We have shown that activation of NMDA and AMPA receptors is required for PrPSc-induced synaptotoxicity in our system, and that specific blockers of these channels prevent dendritic spine retraction. Moreover, PrPSc induces a rapid (within 30 min) increase in intracellular calcium, mediated by NMDA receptors. These results suggest that glutamate receptor activation accompanied by calcium entry is an early step in the signaling cascade initiated by PrPSc, and that glutamate-induced excitotoxicity may play a role in the ensuing synaptic damage. PrPC has been suggested to be a modulator of NMDA receptor function [63] and posttranslational modification [64], and to interact physically with NMDA receptor subunits [65], processes that could be altered by binding of PrPSc to cell-surface PrPC during the initial phase of the conversion process. Also, activation of NMDA-dependent transcriptional pathways is an early event during the pre-clinical phase of prion infection in mice, as determined by microarray profiling [66]. More generally, NMDA receptors play an important role in learning and memory [25], and they are known to be involved in many neurodegenerative diseases in the context of excitotoxic activation [14]. Of note, glutamate stimulation of cultured hippocampal neurons causes dendritic abnormalities (spine retraction) and calcium influx that are similar to those seen in our system with PrPSc treatment [67].
Our results implicate p38 MAPK, specifically the α isoform, in PrPSc-induced synaptotoxicity, based on blockade of dendritic spine retraction and mEPSC reduction by specific inhibitors of this kinase, and by expression of a dominant-negative form of the kinase. Moreover, using immunofluorescence staining, we demonstrated that PrPSc induces rapid (within 1 hr) phosphorylation of p38 MAPK in dendritic spines, a region where this kinase was previously shown to reside [68]. Thus, PrPSc activates a localized, p38-mediated signal transduction cascade that precedes dendritic spine retraction. p38 MAPK is, like its counterpart JNK, a stress-activated protein kinase, and it was originally found to be stimulated by a variety of environmental stresses and cytokines that induce inflammation [31, 37]. It has now been implicated in a wide range of functions, including regulation of the cell cycle, induction of cell death, differentiation, and senescence. In the nervous system, p38 MAPK has been found to play a role in neuronal damage and survival, as well as in synaptic plasticity, and it has been linked to a number of neurodegenerative diseases [32, 69]. Consistent with the results presented here, p38 MAPK, and its downstream substrates MK2/3, have been shown to regulate AMPA receptor trafficking, dendritic spine morphology, and synaptic transmission [70]. We have identified MK2/3 as likely substrates for p38 MAPK in the PrPSc synaptotoxic pathway, based on the use of specific inhibitors, but the identity of the upstream kinases (MAPKKs and MAPKKKs) responsible for p38 MAPK activation remain to be determined. Activation of MAPK pathways typically leads to programmed changes in gene expression [31], and it will be of great interest to characterize these in order to understand PrPSc synaptotoxicity at a genomic level.
Our results demonstrate that re-organization of the actin cytoskeleton within dendritic spines represents a key cellular correlate of PrPSc synaptotoxicity, based on the ability of actin stabilizers (SiR-actin) to block the morphological and electrophysiological consequences of PrPSc treatment. Actin plays an important role in the structure and function of dendritic spines [42]. The equilibrium between G- and F-actin in spines is regulated in an activity-dependent manner, and actin dynamics influences spine development and synaptic plasticity [43, 44]. Postsynaptic receptors, including NMDA and AMPA receptors [71], regulate the actin cytoskeleton via their effect on cytoplasmic actin-binding proteins, either by directly interacting with these proteins or, in the case of NMDA receptors, by increasing intracellular calcium [72, 73]. This mechanism provides a plausible link between PrPSc activation of NMDA receptors, with rapid influx of calcium, and the resulting abnormalities in dendritic spine morphology and function. There is evidence that p38 MAPK can also influence actin cytoskeletal dynamics [74], providing another potential pathway by which PrPSc could alter spine morphology via activation of this kinase.
Several previous studies have implicated specific signal transduction pathways in the pathogenesis of prion diseases, including the UPR [46], oxidative stress [75], MAPKs [76–78], phosphoinositide-dependent kinase-1 (PDK1) [79], metabotropic glutamate receptors [80], NMDA receptors [81], and voltage-gated calcium channels [82]. Given published evidence that the PERK/eIF2α arm of the UPR is activated during prion infection, and that pharmacological blockage of this response has therapeutic benefit [46–48], we tested inhibitors of this pathway in our system. However, we found that these inhibitors had no effect on PrPSc-induced loss of dendritic spines. This discrepancy is likely to reflect fundamental differences between the experimental systems employed. All of the previously quoted studies utilized brain slices or tissue samples from human patients or transgenic mice undergoing neurodegenerative changes after prion infection, making it difficult to isolate the primary synaptotoxic mechanisms engaged by PrPSc. In contrast, we have employed a neuronal culture system that allows us to characterize acute responses to exogenous PrPSc exposure. We have shown that, in this system, treatment with PrPSc causes a detectable increase in intracellular calcium within 30 min, followed by alterations in electrophysiological properties, with complete collapse of dendritic spines by 24 hrs. Moreover, our experiments have demonstrated the involvement of MAPK signaling cascades, which are typically activated within 60 min of exposure to an extracellular stimulus. Our results do not rule out the engagement of additional pathways, which contribute to pathological changes occurring at later times, once PrPSc begins to accumulate.
Our results identify novel molecular targets for therapeutic intervention. Previous efforts to develop anti-prion therapies have focused on strategies for inhibiting formation of PrPSc or enhancing its clearance [83]. In contrast, our studies suggest the possibility of interfering with neurotoxic pathways that lie downstream of PrPSc. We have shown that several kinds of pharmacologic agents prevent PrPSc-induced synaptotoxic effects in our system, including NMDA and AMPA receptor antagonists, and p38 MAPK inhibitors. Importantly, we have found that treatment with a p38 MAPK inhibitor is able to reverse dendritic spine retraction that has already occurred during an initial exposure to PrPSc. This result, which presumably reflects the dynamic nature of dendritic spines, suggests the existence of a therapeutic window for treatment of patients who have already been infected with prions and who might even have sustained a certain level of synaptic damage. The advent of methods for pre-mortem diagnosis of prion diseases [84, 85] makes this type of treatment modality especially compelling. The NMDA receptor antagonist memantine is now a widely used treatment for Alzheimer’s disease [86], and p38 MAP kinase inhibitors have been developed for therapy of inflammatory diseases [87] and CNS disorders, including Alzheimer’s disease [88–90]. It may be feasible to re-purpose these agents for treatment of prion diseases.
It has been proposed that PrPC acts as a cell-surface receptor for Aβ oligomers in Alzheimer’s disease, and that it mediates the neurotoxic effects of these oligomers [51–59]. We have been able to use our neuronal culture system to test whether PrPSc and Aβ oligomers activate a common synaptotoxic pathway upon binding to PrPC. Our results indicate that the two pathways diverge, based on our observation that the synaptotoxic effects of PrPSc are blocked by p38 MAPK inhibitors but not mGluR5 inhibitors, while the reverse is true for the synaptotoxic effects of Aβ oligomers. The precise differences between the PrPSc and Aβ synaptotoxic pathways remain to be determined. Our system could be used to probe the synaptotoxic mechanisms activated by other oligomeric aggregates, including tau and α-synuclein, which are thought to participate in an extracellular transmission phase [91], and which may also utilize PrPC as a cell-surface receptor [92].
Chronic neurodegenerative disorders, like prion, Alzheimer’s and Parkinson’s diseases, are likely to involve multiple pathogenic mechanisms, each of which may be operative at different stages of the disease process. The experimental system described here has allowed us to isolate one very early event (synaptic degeneration) in the pathological cascade, and study its cellular and molecular underpinnings. Future experiments will be aimed at translating these findings into animal models and testing the efficacy of therapeutic interventions directed at specific steps in the signaling pathway we have identified.
Timed-pregnant C57BL/6 mice (referred to as wild-type, WT) were purchased from the Jackson Laboratory (Bar Harbor, ME). Prnp0/0 mice [7] on a C57BL6 background were obtained from the European Mouse Mutant Archive (EMMA; Rome, Italy), and were maintained in a homozygous state by interbreeding.
Mice carrying the p38AF dominant-negative mutation [35] on a C57BL6 background were obtained from the Jackson Laboratory (B6.Cg-Mapk14tm1.1Dvb/J; stock #012736). The mutant allele was maintained in a heterozygous state by breeding with C57BL6 inbred mice. PCR genotyping of tail DNA was performed as per information and protocols are provided by Jackson Laboratory using the following primers: 5’-TAG AGC CAG CCC CAC TTT AGT C-3’ and 5’-GAA GAT GGA TTT TAA GCA TCC GT-3’. The expected PCR products included a 328 bp band representing the dominant-negative allele, and a 195 bp band representing the WT allele.
All procedures involving animals were conducted according to the United States Department of Agriculture Animal Welfare Act and the National Institutes of Health Policy on Humane Care and Use of Laboratory Animals.
Hippocampal neurons were cultured from P0 pups as described [93]. All experiments shown, except those indicated in Figs 1 and 4, were performed with neurons from WT mice. Neurons were seeded at 75 cells/mm2 on poly-L-lysine-treated coverslips, and after several hrs the coverslips were inverted onto an astrocyte feeder layer and maintained in NB/B27 medium until used. The astrocyte feeder layer was generated using murine neural stem cells, as described [94]. Neurons were kept in culture for 18–21 days prior to PrPSc or ADDL treatment.
Hippocampal neurons cultured as described above were treated with purified PrPSc, ADDLs, or control preparations for 24 hrs, followed by fixation in 4% paraformaldehyde and staining with either Alexa 488-phalloidin or rhodamine-phalloidin (ThermoFischer Scientific, Waltham, MA) to visualize dendritic spines, and anti-tubulin antibodies (Sigma-Aldrich, St. Louis, MO) to visualize axons and dendrites. Images were acquired using a Zeiss 880 (Figs 8 and 9) Zeiss 700 (all other figures) confocal microscope with a 63x objective (N.A. = 1.4). The number of dendritic spines was determined using ImageJ software. Briefly, 2–3 isolated dendritic segments were chosen from each image, and the images adjusted using a threshold that had been optimized to include the outline of the spines but not non-specific signals [95]. The number of spines was normalized to the measured length of the dendritic segment to give the number of spines/μm. For each experiment, 15–24 neurons from 3–4 individual experiments were imaged and quantitated.
Immunostaining was performed with the following primary antibodies and corresponding secondary antibodies: anti-gephyrin (Synaptic Systems, Woodland, CA; cat 147011, 1:500); anti-tau (Santa Cruz Biotechnology, Santa Cruz, CA; cat. Sc5587, 1:500); anti-GluR1 (Abcam, Cambridge, MA; cat. Ab31232, 1:500); anti-synaptophysin (Millipore Sigma, St Louis, MO; cat. S5768, 1:500). Quantitation of gephyrin, GluR1, and synaptophysin staining was performed using ImageJ to count the number of fluorescent puncta per μm along isolated dendritic segments (similar to the method described above to quantitate dendritic spine numbers after phalloidin staining).
Hippocampal neurons cultured as described above were washed once with PBS before treatment with Fluo-3 (ThermoFisher Scientific, Waltham, MA) at final concentration of 5 μM. Cells were incubated for 20–30 min at 37°C, followed by a several washes to remove extracellular dye. Neurons were imaged in the green fluorescence channel of an Olympus wide field microscope with 20X objective. Each neuron was imaged for 5 min before either drug or solvent control was added, after which imaging was continued for an additional 30 min. The average signal during the 5 min pre-treatment was considered to represent the baseline calcium level for each cell. The proportion of 30 min recording period during which the calcium signal was ≥ 2X the baseline level was calculated and used as a measure of net calcium accumulation.
Low-density hippocampal neuronal cultures were treated with purified PrPSc or mock preparations for 1 or 24 hrs, followed by fixation in 4% paraformaldehyde containing phosphatase inhibitors (Roche-Sigma, 04 906 845 001) and permeabilization in 1% Triton X-100 for 5 min. Dual immunostaining was performed with anti-phospho-p38 antibody (Thr180/Tyr182) (Cell Signaling, mAb #4511S, 1:100) and anti-total p38 antibody (Abcam, ab31828, 1:300), followed by Alexa 546 and Alexa 633 secondary antibodies, respectively. Cultures were also stained with Alexa 488-phalloidin to visualize dendritic spines. Multi-stack images were acquired using a Zeiss 880 confocal microscope with a 63x objective (N.A. = 1.4). The fluorescence intensities of the phospho-p38 and total p38 signals were quantitated within a region of interest (ROI) using ImageJ software. Images of phalloidin staining were used to determine the location of intact or collapsed dendritic spines. Two or three isolated ROIs on each neuron, coinciding with locations of individual dendritic spines, were captured at 10X zoom, and then the fluorescence intensity and area were determined within the ROI. The fluorescence intensities were normalized to the area of the ROI, and the ratio of phospho-p38/total p38 staining was calculated. For each experiment, 10–12 neurons from 3–4 individual experiments were analyzed.
With the exception of the experiment shown in S1 Fig, hippocampal cultures used for electrophysiological recording were prepared using a procedure that differs from the one used to prepare cultures for dendritic spine imaging. Briefly, hippocampi from newborn pups of the indicated genotypes were dissected and treated with 0.25% trypsin at 37°C for 12 min [96]. Cells were plated at a density of 65,000 cells/cm2 on poly-D-lysine-coated coverslips in DMEM medium with 10% F12 and 10% FBS.
Recordings were made from hippocampal neurons cultured for 18–20 days and treated for 24 hrs with purified PrPSc or control preparations. Whole-cell patch clamp recordings were collected using standard techniques. Pipettes were pulled from borosilicate glass and polished to an open resistance of 2–5 megaohms. Experiments were conducted at room temperature with the following solutions: internal, 140 mM Cs-glucuronate, 5 mM CsCl, 4 mM MgATP, 1 mM Na2GTP, 10 mM EGTA, and 10 mM HEPES (pH 7.4 with CsOH); external, 150 mM NaCl, 4 mM KCl, 2 mM CaCl2, 2 mM MgCl2, 10 mM glucose, and 10 mM HEPES (pH 7.4 with NaOH). Current signals were collected from a Multiclamp 700B amplifier (Molecular Devices, Sunnyvale, CA), digitized with a Digidata 1550A interface (Axon Instruments, Union City, CA), and saved to disc for analysis with PClamp 10 software. Miniature excitatory postsynaptic currents (mEPSCs) were recorded in the presence of TTX (1 μM, Abcam, Cat. # ab120054) and picrotoxin (100 μM, Abcam, Cat. # ab120315). Miniature inhibitory postsynaptic currents (mIPSCs) were recorded in the presence of TTX (1 μM) and CNQX (20 μM, Abcam, Cat. # ab120044). Frequencies and amplitudes of the mEPSCs and mIPSCs were quantitated by Clampfit (Molecular Devices, CA).
Purification was carried out as previously described [23, 97]. In a typical preparation (used for all experiments, except those shown Figs 8 and 9), 18 RML-infected C57BL6 brains were homogenized in 3 ml of 10% sarkosyl in TEND (10 mM Tris-HCl [pH 8], 1 mM EDTA, 130 mM NaCl, and 1 mM dithiothreitol) containing Complete Protease Inhibitor Cocktail (Roche Diagnostics, cat. no. 11836153001) using a glass bead homogenizer. Brain homogenates were incubated on ice for 1 hr and centrifuged at 22,000 x g for 30 min at 4°C. The supernatant was kept on ice, while the pellet was resuspended in 1 ml of 10% sarkosyl in TEND, incubated for 1 hr on ice, and then centrifuged at 22,000 x g for 30 min at 4°C. The pellet was discarded while the supernatants were pooled and centrifuged at 150,000 x g for 2.5 h at 4°C. The new supernatants were discarded, while the pellets were rinsed with 50 ml of 100 mM NaCl, 1% sulfobetaine (SB) 3–14 in TEND plus protease inhibitors, and then pooled by resuspending them in 1 ml of the wash buffer, and centrifuging at 180,000 x g for 2 hr at 20°C. The supernatant was discarded, and the pellet was rinsed with 50 ml of TMS (10 mM Tris-HCl [pH 7.0], 5 mM MgCl2, and 100 mM NaCl) plus protease inhibitors, resuspended in 600 μl of the same buffer containing 100 mg/ml RNase A and incubated for 2 hr at 37°C. The sample was then incubated with 5 mM CaCl2, 20 mg/ml DNase I for 2 hr at 37°C. To stop the enzymatic digestion, EDTA was added to a final concentration of 20 mM, and the sample was mixed with an equal volume of TMS containing 1% SB 3–14. The sample was gently deposited on a 100 μl cushion of 1 M sucrose, 100 mM NaCl, 0.5% SB 3–14, and 10 mM Tris-HCl (pH 7.4), and centrifuged at 180,000 x g for 2 hr at 4°C. The supernatant was discarded and the pellet was rinsed with 50 μl of 0.5% SB 3–14 in PBS, resuspended in 1 ml of the same buffer, subjected to 5 X 5 sec pulses of bath sonication with a Bandelin Sonopuls Ultrasonicator (Amtrex Technologies, Montreal, Canada) at 90% power, and centrifuged at 180,000 x g for 15 min at 4°C. The final supernatant was discarded and the final pellet was resuspended in 900 μl of PBS (50 μl for each starting brain) and sonicated 5 times for 5 sec. Aliquots were stored at -80°C. Mock purifications were also carried out from age-match, uninfected brains. The purified preparations were evaluated by SDS-PAGE followed by silver staining and Western blotting.
For the experiments shown in Figs 8 and 9, PrPSc was purified using the pronase E method [23].
In all experiments, purified PrPSc was added to neuronal cultures at a final concentration of 4.4 μg/ml, identical to what was used in our previous study [23]. An equivalent amount of mock material was used, based on purification from the same proportion of brain tissue.
The uninfected Neuro-2a cells (N2a) were from the ATCC (Cat. #: ATCC CCL-131). A scrapie-susceptible sub-clone (N2a.3) infected with RML prions [98] were plated in 6-well plates. Cells were treated for 3 days with test compounds (p38 MAPK and MK2/3/5 inhibitors), or with DMSO vehicle as a negative control and Congo red as a positive control. Cultures were then split at a 1:5 ratio and treatment with compounds in fresh medium continued for 4 more days. At the end of the 7-day treatment, cells were lysed in 300 μl of lysis buffer (0.5% NP-40, 0.5% deoxycholate, 10mM Tris-HCl, pH8 and 100mM NaCl) and protein concentration was measured using BCA assay (ThermoFisher Scientific, Waltham, MA). Samples were then treated with proteinase K (40 μg/ml) at 37°C for 1 hr. Digestion was stopped by addition of 10x Complete Protease Inhibitor Cocktail (Roche, Indianapolis, IN). Samples were centrifuged at 180,000 x g for 1 hr at 4°C. Pellets were dissolved in 20 μl of 1x Laemmli loading buffer (Bio-Rad, Hercules, CA) and were boiled for 3 min before loading on pre-cast 12% SDS-PAGE Criterion gels (Bio-Rad, Hercules, CA). Western blotting was performed according to standard procedures. PrPSc was detected using the anti-prion antibody D18 [99] and HRP-coupled, anti-human secondary antibody (Jackson ImmnoResearch, West Grove, PA). Quantitation was performed using the ImageJ gel quantification function.
ADDLs were prepared from synthetic Aβ 1–42 peptide as previously described [55, 100]. Aβ peptide (ERI Amyloid Laboratory, Oxford, CT) was dissolved in HFIP at a concentration of 1 mM and sonicated for 10 min in an ice-water bath. The sample was allowed to incubate at room temperature for 1 hr before it was transferred to a centrifuge tube and spun at 15,800 x g for 1 min. The supernatant was transferred into a new glass vial and was dried under a fume hood with nitrogen gas. The film of dried peptide was dissolved in DMSO, and then diluted into Ham’s F12 phenol-red-free medium (Gibco/ThermoFisher scientific, Waltham, MA) to a final concentration of 100 μM. The sample was incubated at room temperature for 16 hrs and was then centrifuged for 15 min at 15,800 x g. The supernatant was aliquoted, flash-frozen in liquid nitrogen, and stored at -80°C.
All procedures involving animals were conducted according to the United States Department of Agriculture Animal Welfare Act and the National Institutes of Health Policy on Humane Care and Use of Laboratory Animals. Ethical approval (AN-14997) was obtained from Boston University medical center institutional animal care and use committee.
|
10.1371/journal.pntd.0000786 | A Nuclear Family A DNA Polymerase from Entamoeba histolytica Bypasses Thymine Glycol | Eukaryotic family A DNA polymerases are involved in mitochondrial DNA replication or translesion DNA synthesis. Here, we present evidence that the sole family A DNA polymerase from the parasite protozoan E. histolytica (EhDNApolA) localizes to the nucleus and that its biochemical properties indicate that this DNA polymerase may be involved in translesion DNA synthesis.
EhDNApolA is the sole family A DNA polymerase in E. histolytica. An in silico analysis places family A DNA polymerases from the genus Entamoeba in a separate branch of a family A DNA polymerases phylogenetic tree. Biochemical studies of a purified recombinant EhDNApolA demonstrated that this polymerase is active in primer elongation, is poorly processive, displays moderate strand displacement, and does not contain 3′–5′ exonuclease or editing activity. Importantly, EhDNApolA bypasses thymine glycol lesions with high fidelity, and confocal microscopy demonstrates that this polymerase is translocated into the nucleus. These data suggest a putative role of EhDNApolA in translesion DNA synthesis in E. histolytica.
This is the first report of the biochemical characterization of a DNA polymerase from E. histolytica. EhDNApolA is a family A DNA polymerase that is grouped into a new subfamily of DNA polymerases with translesion DNA synthesis capabilities similar to DNA polymerases from subfamily ν.
| Genotoxic agents like ultraviolet radiation, alkylating compounds and reactive oxidative species have the potential to originate DNA lesions that are not bypassed by replicative DNA polymerases. Eukaryotic organisms contain a specialized subset of DNA polymerases capable of translesion DNA synthesis. These DNA polymerases belong to DNA polymerases from families A, B, and Y. In this work, we characterized the sole family A DNA polymerase of the parasitic protozoa E. histolytica, EhDNApolA. The biochemical characterization of recombinant EhDNApolA indicates that this protein is an active DNA polymerase able to primer extension and moderate strand displacement. The ability of EhDNApolA to faithfully incorporate dATP opposite thymine glycol, and its nuclear localization indicates that this polymerase may have a role in translesion DNA synthesis. E. histolytica is exposed to oxidative stress during tissue invasion by phagocytes. Understanding DNA metabolism in E. histolytica is important because this parasite has shaped some metabolic pathways by horizontal gene transfer, infects approximately 50 million people annually, and is the second leading cause of death among protozoan diseases.
| DNA replication and translesion DNA synthesis in eukaryotes is accomplished by a battery of DNA polymerases. For instance, the genome of Homo sapiens contains 15 DNA polymerases divided into four families: A, B, X, and Y according to their amino acid sequence homology [1]–[3]. Nuclear replicative DNA polymerases δ and ε · belong to family B, whereas DNA polymerases involved in translesion DNA synthesis are present in all four families.
Entamoeba histolytica is a parasitic protozoa which causes amebic dysentery and liver abscess [4]. In comparison to eukaryotes that contain DNA in organelles like mitochondria or chloroplasts. E. histolytica is an early branching eukaryote in which its mitochondria diverged to form an organelle with no detectable DNA. This organelle is dubbed mitosome [5], [6], and although its function is not definitively established, experimental evidence suggests a role in sulfate activation [7] and oxygen detoxification [8]. Thus, the 24 Mbp genome of E. histolytica is exclusively nuclear and it encodes several putative DNA polymerases (Table S1) [9]. As an eukaryotic organism, the genome of E. histolytica is expected to be replicated by DNA polymerases δ and ε. Although a gene encoding DNA polymerase ε is not present in the current genome annotation of E. histolytica, a gene encoding DNA polymerase δ is present. E. histolytica contains homologs of Rev 1 and Rev 3 proteins, that compose the principal DNA polymerase involved in translesion synthesis of thymine dimers: DNA pol ζ [10], [11]. In addition, the genome of E. histolytica contains five DNA polymerases which share high sequence homology with DNA polymerases from autonomous replicating elements found in other protozoa and with the well-characterized DNA polymerase from bacteriophage φ29[12]. E. histolytica also contain one family A DNA polymerases in its genome. Family A DNA polymerases are modular enzymes consisting of three independent domains: a N-terminal 5′-3′ exonuclease domain, a 3′-5′ exonuclease domain, and a C-terminal polymerase domain [1], [13], [14]. Crystal structures of family A DNA polymerases revealed a modular organization of the polymerase domain and its division into three subdomains: palm, fingers, and thumb, which together form a cleft that binds the primer-template [15]. Family A DNA polymerases contain three conserved motifs: A, B, and C in the polymerization domain [16]. Motifs A and C are located at the palm subdomain and contain two carboxylates involved in the coordination of two metal ions involved in the nucleophilic attack of the incoming deoxynucleotide to the 3′ OH of the primer strand [13]. Motif B is located at the fingers subdomain and is involved in positioning the template strand into the polymerase active site [15]. In eukaryotes, family A polymerases are involved in the replication of mitochondrial and chloroplast genomes [17], [18]. The archetypical DNA polymerase in eukaryotes is DNA polymerase γ, which is the replicative mitochondrial DNA polymerase. Besides DNA polymerase γ, vertebrates contain two other family A DNA polymerases: DNA polymerase ν and DNA polymerase θ. In contrast to DNA polymerase γ, the localization of these polymerases is nuclear. Human DNA polymerases ν and θ are capable of translesion DNA synthesis and they have a role in DNA repair [19]–[24].
In this work, we report the initial characterization of the sole family A DNA polymerase from E. histolytica (EhDNApolA). We propose a role of this DNA polymerase in translesion DNA synthesis of oxidative lesions like 8-oxo guanosine and thymine glycol. These lesions may be generated by the oxidative environment of the colonic tissue and the constant insult of the reactive oxygen species produced by phagocytes during E. histolytica pathogenesis.
To identify putative family A DNA polymerases in E. histolytica, we initially used the amino acid sequence of the Klenow fragment of E. coli (Protein Data Bank accession code: 1KFS) to blast the Pathema database (http://pathema.jcvi.org/Pathema/).
The phylogenetic tree was constructed using the amino acid sequences of family A DNA polymerases of representative mitochondrial DNA polymerases, bacteriophage DNA polymerases, DNA polymerases ν, and bacterial DNA polymerases. The amino acid sequences of these proteins were aligned using the program ClustalW [25]. The catalytic amino acids of motifs A, B, and C, were conserved through the alignment. This sequence alignment was used to construct a dendogram with the Neighbor-Joining method of the Molecular Evolutionary Genetic Analysis (MEGA) software [26]. The robustness of the dendogram was assessed by bootstrap analysis of 1000 replicates.
To build the structural model of EhDNApolA, the amino acid sequence of EhDNApolA was structurally aligned with the amino acid sequence present in the crystal structure of Klenow fragment (Protein Data Bank accession code: 1KFS) [27], using the program Molecular Operating Environment (MOE). As Klenow Fragment contains 605 amino acids and EhDNApolA has 657, the gaps between the two aligned proteins were built according to the peptide library present in the MOE database. Twenty models were generated and each model was minimized using the CHARMM27 force field.
Trophozoites of HM1:IMSS strain were axenically cultured in TYI-S-33 medium supplemented with 15% of bovine serum [28] at 37°C and used in logarithmic growth phase for all experiments.
The open reading frame of EhDNApolA was amplified by PCR from genomic DNA of E. histolytica strain HM1:IMSS. To allow directional cloning, the sense oligonucleotide (5′-ggttgg ggatcc atg gaa aaa aca cca aga aat tct-3′) contained a BamH I restriction site (underlined) and the antisense oligonucleotide (5′-ggttgg aagctt tta att caa gtt gta agg atg aag-3′) contained a Hind III restriction site (underlined). PCR was carried out using 150 ng of genomic DNA, 25 pmol of each oligonucleotide, and 125 µM of each dNTP. The amplified product was simultaneously digested with BamH I and Hind III and ligated into the pCOLD I vector (Takara). The ligation mixtures were transformed into an E. coli DH5α strain. Plasmidic DNA was analyzed using restriction mapping and confirmed by DNA sequencing. Cloning of the open reading frame of EhDNApolA in the pCOLD I vector confers a 6-His tag at the N terminus of the recombinant protein.
Seven Balb/c mice were bled and tested for their response to total protein extracts of E. histolytica. Five mice did not present any response and were inoculated with a peptide corresponding to residues 286 to 297 of the thumb subdomain of EhDNApolA (HKIEMETKKIIG). The mice were immunized with 150 µg of the peptide combined with Freund's adjuvant. Six weekly bursts were applied and the reactivity of each mouse was assessed using recombinant EhDNApolA. After six weeks of immunization, the immune sera was collected, purified, and stored at −20°C.
All animal work was conducted according to the legislation enforced in México (NOM-062-ZOO-1999) and by CINVESTAV's committee for animal care and use. The Mexican legislation is based on the Guide for the Care and Use of Laboratory Animals, NRC.
We tested the antibodies for their response and specificity in total extracts of E. histolytica strain HM1:IMSS and against recombinantly induced EhDNApolA. For Western blot assays, we used total, nuclear, and cytoplasmic extracts from E. histolytica strain HM1:IMSS prepared as previously described [29]. Protein extracts were separated using a 15% SDS-PAGE gel and transferred onto a nitrocellulose membrane. The membranes were incubated with a 1 to 2000 dilution of the purified immune sera and an anti-actin antibody [30] in 1% nonfat dry milk, 0.05% Tween-20 in PBS 7.4 for 2 hours. The reactivity was detected using peroxidase conjugated secondary antibodies (1 to 2000 dilution) with the ECL Plus detection kit (GE Healthcare). As a control, we used antibodies against actin and CBP-B previously characterized.
cDNA was synthesized using 1 µg of total E. histolytica RNA with an oligo(dT) adaptor. The RT-PCR reactions contained 0.5 µl of cDNA and 15 pmol of each specific oligonucleotide combination. The segment corresponding to motif A was amplified using the sense oligonucleotide 5′-agagacttattattacacat3-' and antisense oligonucleotide, 5′-attctttttaagccaatgtgc-3′. Motif C was amplified using the sense oligonucleotide; 5′-ttacattcaagttgggtaggt-3′ and antisense oligonucleotide 5′-aacagtaactacaacaggaac-3′. The actin control was amplified with the sense oligonucleotide 5′-aag ctg cat caa gca gtg aa-3′ and antisense 5′-gga atg atg gtt gga aga gg -3′. RT-PCR products were separated by gel electrophoresis in 1.5% agarose gels, stained with ethidium bromide, and visualized with a standard UV transilluminator.
Semi-quantitative RT-PCR assays were performed using total cellular RNA isolated from Entamoeba histolytica grown in basal culture conditions using SV Total RNA Isolation System (Promega Madison, WI, USA). The amount of total or messenger RNA isolated from the cells was quantified using an ND-1000 spectrophotometer (NanoDrop, Fisher Thermo, Wilmington, DE, USA). cDNA was synthesized using gene-specific primers. 1 µg of total RNA was added to a reaction containing 625 mM EhDNApolA motif A antisense oligonucleotide or actin antisense oligonucleotide, 0.5 mM of the deoxynucleotide triphosphates, 1 unit of RNasin Ribonuclease Inhibitor, 1 ml of ImProm-II™ Reverse Transcriptase (Promega Madison, WI, USA) and RNase-free water to 20 µl. Reactions were incubated at 25°C for 5 min, then at 42° for 60 min followed by 75°C for 15 min, to inactivate the reverse transcriptase. PCR was performed using EhDNApolA or actin specific sprimers to amplify cDNA segments of 168 or 192 bp in length respectively, with the estimated primer melting temperature of 61.5 or 52°C. RT-PCR products were separated by gel electrophoresis in 1% agarose gels, stained with ethidium bromide, and visualized with a standard UV transilluminator.
The pCOLDI-EhDNApolA construct was transformed into an E. coli BL21 DE3-Rosseta II strain. Transformants were inoculated in 100 ml of LB supplemented with 100 µg/ml of ampicilin and 35 µg/ml of chloramphenicol and used to inoculate 2 liters of LB. This culture was grown at 37°C until it reached an OD600 of 0.6. The culture was incubated in ice for 30 minutes and IPTG was added to a final concentration of 0.5 mM. Induction proceeded for 16 hours at 16°C. The cell pellet was harvested by centrifugation at 6,500 rpm. Cell lysis was carried out using a French press in a buffer containing 50 mM potassium phosphate pH 8, 300 mM NaCl, and 1 mM PMSF. The lysate was centrifuged at 17,000 rpm for 30 minutes at 4°C. The soluble fraction was filtrated and the recombinant EhDNApolA was purified using a Ni2+-NTA affinity chromatography in a previously equilibrated Hi-Trap Column (GE Healthcare). The initial wash consisted of 50 ml of lysis buffer supplemented with 35 mM imidazol and the second wash consisted of 100 ml of lysis buffer supplemented with 50 mM imidazol. Protein elution was carried out in lysis buffer supplemented with 500 mM imidazol. The eluate was dialyzed in a buffer containing 50 mM potasium phosphate pH 7.0, 5 mM β-mercaptoethanol (BME), 50 mM NaCl, 2 mM EDTA and 5% glycerol. To further purify EhDNApolA, the dialyzed protein was loaded into a phosphocellulose column and eluted with a NaCl gradient (100 to 1500 mM). EhDNApolA eluted between 600 to 650 mM of NaCl. The collected fractions were dialyzed in 50 mM potasium phosphate pH 7.0, 1 mM β-mercaptoethanol, 150 mM NaCl and 1 mM EDTA and stored at 4°C. Protein samples were run on a 10% SDS-PAGE and stained with Coomassie Brilliant Blue R-250.
The following oligonucleotides were used to generate double stranded polymerization substrates: a) 45mer template (5′-cct tgg cac tag cgc agg gcc agt tag gtg ggc agg tgg gct gcg-3′) b) 24mer primer (5′-cgc agc cca cct gcc cac cta act-3′); c) 18mer primer (5′-cgc agc cca cct gcc cac-3′); and d) 21mer non-template (5′-ggc cct gcg ctagtgccaagg-3′).100 nmols of the 24mer primer or the 18mer primer were 5′end labeled using T4 Kinase with γ-[32]ATP. The probes were purified using the nucleotide removal kit (Qiagen) according to the manufacturer instructions. The polymerization substrates were annealed to a final concentration of 10 nM in 20 mM Tris pH 7.5, 150 mM NaCl.
A radiolabeled DNA substrate consisting of the 45mer template annealed to the 24mer primer was incubated with increasing concentrations of EhDNApolA (from 0 to 180 nM) in a buffer containing 50 mM NaCl, 10 mM Tris-HCl pH 7.5, 2.5 mM MgCl2, 1 mM dithiothreitol (DTT), 1 µg/ml BSA, and 5% glycerol. DNA-protein complexes were resolved through a 6% non-denaturing polyacrylamide gel (PAGE) and electrophoresed at 80 V for 2 h at room temperature in 0.5x TBE buffer. Gels were vacuum-dried and radioactive complexes were detected in a Phosphor Imager apparatus and analyzed using the ImageQuant software (BioRad).
20 µl polymerization reactions consisted of 20 mM Tris-HCl pH 7.5, 2.5 mM MgCl2, 1 mM DTT, 1 µg/ml BSA, 200 fmol primer-template, 60 fmol EhDNApolA. Reactions were stopped with a buffer containing 95% formamide, 1mM EDTA, 0.01% xylene cyanol. Samples were resolved on a 16% polyacrylamide, 8M urea denaturing gels. Quantification of the polymerization products was carried out in a Phosphorimager using ImageQuant software.
Templates containing 8-oxo guanosine and abasic site were purchased from Oligos Etc. Templates containing 5 S-6R thymine glycol, 5R-6S thymine glycol, cis-syn cyclobutane pyrimidine dimer, and 6-4 photo product were synthesized by Professor Shigenori Iwai's group as previously described [31]. A specific 5′ γ-[32]ATP labeled primer was annealed to each template, so the first template base corresponds to each specific lesion. 60, 120 and 240 fmol of EhDNApolA were incubated with 100 fmol of each primer-template at 37°C for 2.5 minutes with 100 µM of each dNTP. Reactions were stopped by adding an equal volume of gel stop/loading buffer. The reactions were run on a 16% denaturing 8 M urea polyacrylamide gel.
For steady-state kinetic analysis, DNA polymerase activity assays were performed using 2 pmol of duplex DNA incubated with 10 fmol of EhDNApolA and varying dNTP concentrations. The reactions were incubated for 10 minutes at 37°C. Four different DNA duplexes were used to determine the kinetic parameters of each nucleotide opposite to its cognate base. To assure linearity, less than 20% of the substrate was converted to product.
Trophozoites of E. histolytica grown in basal cell culture condition were transferred to glass coverslips. Cells were fixed with 4% paraformaldehyde for 1 hour at 37°C, washed with PBS pH 6.8, permeabilized with 0.5% (v/v) Triton X-100 at 37°C for 60 min, and blocked with 50 mM glycine for 1 h at 37°C and with 1% fetal bovine serum for 15 min. Finally, they were incubated with anti-EhDNApolA antibodies (1 to 75) overnight at 4°C. The cells were washed and conjugated with fluorescein labelled secondary antibodies (Jackson Immuno Research) at 1∶500 dilution. The nucleic acids were stained with DAPI (4′,6′-diamidino-2-phenylindole) washed, and mounted with Vectashield solution (Vector Lab. Burlingame, CA). Light optical sections were obtained through a Nikon inverted microscope attached to a laser confocal scanning system (Leica Microsystems) and analyzed by Confocal Assistant software Image.
A survey of E.histolytica genome with the amino acid sequences of Klenow Fragment and representative family A DNA polymerases revealed that this parasite contains a single open reading frame that codes for a putative family A DNA polymerase. This open reading frame is located at locus EHI_073640 and codes for a protein of 657 amino acids with GenBank accession number XP_653960 and 25% amino acid identity to Klenow fragment. In this work we dubbed this putative polymerase EhDNApolA. The predicted amino acid sequence of EhDNApolA was used as query to search for homologous proteins in the genomes of E. invadens and E. dispar. We found that locus EIN_094210 of E. invadens and locus EDI_083910 of E. dispar also code for putative family A DNA polymerases with 50% and 88% amino acid sequence identity to EhDNApolA respectively. The lack of a conserved 3′-5′ exonuclease active site in the DNA polymerases of the genus Entamoeba indicates that these polymerases are not related to mitochondrial DNA polymerases. A phylogenetic analysis of 37 DNA polymerases (Table S2) positions the DNA polymerases from the genus Entamoeba in a separate branch with respect to other subfamily A DNA polymerases. In this division, family A DNA polymerases are grouped into five separate branches or subfamilies (Figure 1A). The high bootstrap value of each branch validates this division. (Figure 1A). Family A DNA polymerase from Entamoeba have a clear conservation of the catalytic motifs present in the polymerization domain. (Figure 1B). The disappearance of the exonuclease domain is a common feature in some family A DNA polymerases, including DNA polymerase ν, DNA polymerase θ, and several bacterial polymerases. The crystal structure of Klenow fragment bound to duplex DNA in its exonuclease domain was used as template to build a homology model of EhDNApolA [32]. The structural model of EhDNApolA depicts the modular organization present in family A polymerases. In this model, EhDNApolA adopts a structure that resembles a cupped right hand in which the three subdomains of the polymerization domain (fingers, palm, and thumb) form a DNA binding cleft (Figure 1C). Although this structural model depicts high degree of conservation in the polymerization domain, it should not be interpreted as an experimental structure.
In order to test the biochemical properties of EhDNApolA, its open reading frame was cloned into the pCOLD I vector (Takara). Heterologous protein expression was enhanced with the use of the E. coli strain BL21- Rosseta II (Figure 2A, lanes 1 and 2, and data not shown). The recombinant EhDNApolA was soluble (Figure 2A, lane 4) and purified nearly to homogeneity using Ni2+-NTA affinity chromatography as a first chromatographic step (Figure 2A, lane 8). To assure the purity of the recombinant protein and avoid a possible contamination with endogenous DNA polymerases, we performed a second chromatographic step using a phosphocellulose chromatography. After this step, the recombinant protein was more than 95% pure (Figure 2A, lane 9). Our structural model of EhDNApolA was used to design epitopes to raise polyclonal antibodies. The best epitope candidate was a peptide located at the thumb subdomain (residues 286 to 297) of EhDNApolA. The pre-immune serum did not unveil any reactivity against total extracts of E. histolytica (Figure 2B, lane 3) and recombinant expressed polymerase (data not shown). The raised antibodies recognized a single band of 75 kDa in bacterial extracts expressing recombinant EhDNApolA and in total extracts from E. histolytica (Figure 2B, lanes 2 and 4). As observed in Figure 2B, the raised polyclonal antibodies were highly specific for EhDNApolA and did not present any cross reactivity that could compromise the localization of EhDNApolA in vivo.
We tested the ability of EhDNApolA to shift a fixed amount of primer-template (3 nM) by increasing the EhDNApolA concentration from equimolar amounts to 60 fold excess (Figure 3A, lanes 2 to 7). The appearance of a major retarded band that increased in intensity according to the amount of added recombinant protein indicates that EhDNApolA is able to recognize a primer-template substrate. A few minor bands were also detected, however they had low abundance in comparison to the more abundant complex. It is possible that these bands resulted from some alternate binding mode of EhDNApolA to the primer-template, for instance a binding that resembled an editing complex [32], [33].
In order to test if recombinant EhDNApolA displays polymerization activity, we measured its ability to incorporate deoxynucleotides to an annealed primer-template. The presence of elongation products indicates that the recombinant EhDNApolA is a functional DNA polymerase (Figure 3B, lanes 2 and 3). The experimental setup placed the first template thymine at position 36 (Figure 3B). Thus, if dGTP, dCTP and ddATP were added as the only nucleotides in the reaction mixture, it is expected that elongation would stop at position 36. EhDNApolA readily incorporates ddATP and it is halted at position 36 (Figure 3B, lane 2). This is in contrast to Klenow fragment that did not efficiently incorporate ddATP, and replicates beyond the first thymine template (Figure 3B lane 5). Mutagenesis studies demonstrated that residue F762 of Klenow fragment is responsible for ddNTPs selectivity [34]. DNA polymerases with a tyrosine in the corresponding position incorporate ddNTP efficiently because the hydroxyl group of the tyrosine compensates for the missing 3′ OH of the ddNTPs [34]. The corresponding residue of Klenow fragment's F762 in EhDNApolA is a tyrosine (residue Y485). Thus, as it is observed, EhDNApolA efficiently incorporates ddNTPs during primer extension (Figure 3B lane 2). Several bands of lower molecular weight were observed during primer extension reactions. These bands may indicate that, like Klenow Fragment, EhDNApolA is a poorly processive DNA polymerase (Figure 3B, lanes 2–3 and 5–6).
Some family A DNA polymerases, like DNA polymerase γ and DNA polymerase ν are capable of strand displacement. We tested the strand displacement capabilities of EhDNApolA in comparison to other DNA polymerases. The strand displacement activity of EhDNApolA was measured in a primer-template substrate containing a gap of six nucleotides and this activity corresponds to the appearance of primer elongation products longer than 24nt (Figure 3C). φ29 DNA polymerase has strong strand displacement capabilities and is a highly processive polymerase. According to these characteristics, φ29 DNA polymerase is not halted at position 24 (Figure 3C, lane 2). Taq DNA polymerase and T7 DNA polymerase are DNA polymerases with moderate strand displacement, as some polymerase's population are blocked at positions 24 and 23 (Figure 3D, lanes 3 and 4). We found that EhDNApolA was able to perform strand displacement at similar levels that Taq DNA polymerase (Figure 3C, lanes 3 and 5). However, in contrast to Taq and T7 DNA polymerases, EhDNApolA has weak primer-template affinity during strand displacement, as evidenced by the apparition of bands from 25 to 30 nt (Figure 3C lane 5).We tested the ability of the purified EhDNApolA to degrade a labeled primer-template. No detectable 3′-5′ exonuclease activity was observed even after 8 minutes of incubation with EhDNApolA (data not shown). This is in agreement of our in silico prediction which indicates that EhDNApolA does not contain the motifs needed for 3′-5′ exonuclease activity [35].
An important step to measure kinetic parameters is to determine the optimal reaction conditions. Thus, we determined the optimal salt concentration, pH, MgCl2 concentration, and temperature for EhDNApolA activity. EhDNApolA is strongly inhibited by NaCl. The optimal NaCl concentration for EhDNApolA activity is from 0 mM to 50 mM NaCl (Figure S1A, lanes 2 to 5). Increasing the NaCl concentration to 100mM only permitted the incorporation of a single nucleotide (Figure S1A, lane 6). EhDNApolA was not active at 200mM NaCl, a concentration that is similar to physiological concentrations. In this respect, EhDNApolA resembles Klenow fragment which has decreased activity at concentrations higher than 50 mM NaCl [36]. The optimal MgCl2 concentration was 2.5 mM (Figure S1B). This metal concentration was similar to the optimal concentration of Thermus aquaticus and Klenow Fragment DNA polymerases. The optimal pH for polymerization activity is 7.5. EhDNApolA has approximately 80% of activity between pH 7 and 8 (Figure S1C). As expected for an enzyme from a mesophilic organism, the optimal temperature for EhDNApolA activity was 37°C (Figure S1D). Using the optimal buffers, we determined the kinetic parameters for EhDNApolA activity using steady-state kinetics.
The Km of the incoming nucleotide varied from 1.49 to 2.3 µM and the Vmax varied between 2.9 to 3.3 nMol/min (Table S3). The kinetic constants of EhDNApolA were similar to several family A DNA polymerases including the DNA polymerase from Bacillus stereothermophilus, Klenow Fragment and human DNA polymerase ν [20], [37], [38].
Family A DNA polymerases are highly variable in their DNA replication accuracy. Polymerases from bacteriophages, bacteria, and mitochondria are high fidelity polymerases. In contrast, human DNA polymerases θ and ν are low fidelity polymerases. For instance, human DNA polymerase ν misincorporates thymine across from a guanine template with a frequency of 0.45 [20]. To test the fidelity of EhDNApolA, we used a set of primer-templates in which the first template base is different from the following templated base (Figure 4). EhDNApolA selectively incorporated the incoming nucleotide according to the Watson-Crick rules at all four template bases (Figures 4A, 4B, 4C, and 4D). EhDNApol does not extensively misincorporate at canonical templates. This is in contrast to DNA polymerases of subfamilies ν and θ that are low fidelity polymerases. Although an extensive kinetic analysis is needed to quantify the fidelity of EhDNApol, it is evident that EhDNApol follows the Watson-Crick rules during nucleotide incorporation at canonical templates.
DNA lesions can be classified as non-blocking and strong blocking. DNA lesions like 8-oxo guanosine are readily bypassed by the majority of family A DNA polymerases. On the other hand, DNA lesions like thymine glycol, abasic site, and thymine dimers are strong blocks to replication. Seeming exceptions are the cases of DNA polymerase ν that efficiently bypasses 5S-thymine glycol and DNA polymerase θ that bypasses abasic sites [19], [23]. To measure translesion DNA synthesis by EhDNApolA, we tested increasing amounts of the polymerase in a control template thymine and in several DNA lesions. To permit the relative extension comparison, less than 50% of the control thymine template was extended at the lower polymerase concentration. EhDNApolA extended a thymine template to the final 45mer product with an efficiency of 42% at the higher polymerase concentration (Figure 5, lane 4). EhDNApolA efficiently bypasses 8-oxo guanosine, as 26% of the labeled primer was extended to the final 45mer product at the higher polymerase concentration (Figure 5, lanes 5 to 8). EhDNApolA bypasses 5S, 6R thymine glycol with an efficiency of 6%, this efficiency is low in comparison to the thymine template, but is significantly larger than other DNA polymerases, like RB69 that which is completely blocked by this lesion [20], [39]. The stalled 17mer product constitutes 80% of the total labeled DNA in the reaction (Figure 5, lanes 9 to 12). Similar results have been observed for an exonuclease deficient Klenow fragment [20], [40]. EhDNApolA bypasses the 5R, 6S thymine glycol with an efficiency of 4% (Figure 5, lanes 13 to 16). As in the case of the 5S, 6R thymine glycol lesion, this efficiency was low in comparison to the control thymine but is more efficient than DNA polymerase ν [20] or any other family A DNA polymerase characterized to date. The stalled 17mer product represents 33% of the product (Figure 5, lane 16). EhDNApolA is unable to bypass the CPD and the 6-4 photoproduct (Figure 5, lanes 17 to 24). EhDNApolA incorporates only one nucleotide opposite an abasic site (Figure 5, lanes 25 to 28) and some bypass occurs at higher polymerase concentrations (data not shown).
EhDNApolA was able to bypass 8-oxoguanosine and to incorporate across from an abasic site (Figure 5, lanes 6 to 8 and 26 to 28). In order to test the fidelity of lesion bypass, we tested the incorporation of each deoxyribonucleotide across from each lesion. 8-oxoguanosine is a dual code lesion that can template for dCTP and dATP. The syn conformation of 8-oxoguanosine mimics a thymine template that allows dATP incorporation[41]. EhDNApolA incorporated dATP across from 8-oxoguanosine more efficiently than dCTP (Figure 6A, lanes 2 and 5). The rationale for this incorporation resides in the nature of a specific residue at the fingers subdomain. A bulky residue like K635 in T7 DNApol dictates the incorporation of dCTP [41] whereas a glycine residue in B. stearothermophilus DNA polymerase may dictate the incorporation of dATP [38], [42]. EhDNApolA contains a serine in the position corresponding to residue K635 of T7 DNA polymerase, thus the preferential incorporation of dATP is predicted. Family A DNA polymerases preferentially insert dATP across from an abasic site, a phenomena known a as the “A-rule” [43]. EhDNApolA incorporates preferentially dATP (Figure 6B, lane 3) and dGTP (Figure 6B, lane 6) opposite abasic sites. EhDNApolA only incorporates a purine across from the lesion, but it does not extend from the lesion (Figure 6B). This characteristic is conserved with other family A DNA polymerases, like Klenow fragment [44] or DNA polymerase ν [20]. However, DNA polymerase θ is able to bypass abasic sites [23]
In contrast to replicative DNA polymerases, like DNA polymerase RB69, that stall at thymine glycol lesion, EhDNApol is able to bypass this lesion. Family A DNA polymerases, like an exonuclease deficient Klenow fragment bypasses the 5S, 6R thymine glycol lesion, but are halted at the 5R, 6S-thymine glycol lesion [20]. Although EhDNApol readily incorporates across from a 5S, 6R thymine glycol lesion, it is severely hampered during its elongation. The 5R, 6S thymine glycol lesion is also bypassed by EhDNApol, although with different properties than the 5S, 6R thymine glycol lesion. The accumulation of the first incorporated nucleotide occurs less efficiently than in the 5S, 6R lesion. Structural studies suggest that thymine glycol prevents primer extension by obstructing the next 5′ templated base to stack against it [39].
EhDNApolA is able to accurately bypass thymine glycol (Figures 6C and 6D). EhDNApolA inserts dATP opposite 5S, 6R thymine glycol (Figure 6C, lane 3) and 5R, 6S thymine glycol (Figure 6D, lane 3). EhDNApolA did not incorporate any other nucleotide opposite 5S, 6R or 5R, 6S thymine glycol (Figure 6C, lanes 4 to 6 and Figure 6D, lanes 4 to 6). EhDNApolA incorporates dATP at the 5S, 6R thymine glycol lesion and in this context misincorporates dATP opposite template dCTP. A similar phenomenon occurs in human DNA polymerase κ [40] indicating that a subtle DNA distortion originated by the lesion may influence nucleotide incorporation fidelity by these DNA polymerases, as observed by the apparent low fidelity of EhDNApolA. This is in contrast to the high fidelity opposite template dCTP in a canonical template (Figure 4C).
In order to verify that the gene of EhDNApolA is transcribed in vivo, we carried out a RT-PCR using specific oligonucleotides that amplified the two conserved motifs A and C of the EhDNApolA gene. The oligonucleotides were designed to amplify a region of 168 bp corresponding to motif A and a region of 156 bp corresponding to motif C of the EhDNApolA gene (Figure 7A, lanes 3 and 4 respectively). The RT-PCR control product of the actin gene control corresponds to 192 bp (Figure 7A, lane 2) and the no RT reaction showed no appearance of a new band (data not shown). The RT-PCR reaction produced the expected products, thus confirming that the EhDNApolA gene is transcribed under basal conditions in E. histolytica. In order to quantify the abundance of the EhDNApolA transcript, we compared the relative transcript under basal conditions in comparison to actin. The average intensity of the EhDNApolA transcript is approximately 70% of the intensity of the actin transcript (Figure S2). Thus, the EhDNApol A gene is expressed at similar levels than the actin gene in basal cell culture conditions. To determine the localization of EhDNApolA in E. histolytica, we carried out Western blot analyses of fractionated cytoplasm and nuclear extracts using the anti-peptide EhDNApol A antibody and anti-actin and anti-C/EBPβ antibodies as controls. The appearance of a single protein band of 75kDa in the nuclear and cytoplasmic fractions using the anti-peptide EhDNApol A antibody indicates that a population of EhDNApolA is translocated from the cytoplasm into the nucleus (Figure 7B, lanes 1 and 2). The same patter is observed with the anti-actin antibody, as actin is a protein with cytoplasmic and nuclear localization [30]. Because nuclear fractions are often contaminated with cytosolic fractions, we used the identification of C/EBPβ, as a control of the nuclear extract purification protocol. The antibody against this protein identifies a double band of approximately 65 kDa in Western blot assays; however this recognition occurs predominantly in nuclear extracts and not in the cytoplasmic fraction [45]. The data indicates that a population of EhDNApolA is imported from the cytoplasm into the nucleus. Confocal microscopy of E.histolytica trophozoites stained with antibodies against the peptide of EhDNApolA corroborates that EhDNApolA is translocated into the nucleus (Figure 7). DAPI staining indicates the localization of nuclear double-stranded DNA in the parasite (Figure 7D) and immunofluorescence analysis using anti-EhDNApolA antibodies shown a possible nuclear localization (Figure 7E). Merged field indicated that EhDNApol A colocalizes with DAPI staining of the nuclear DNA of E. histolytica (Figure 7 F). An analysis of the EhDNApolA amino acid sequence using the pSORT program (http://psort.ims.u-tokyo.ac.jp/) predicted the presence of several nuclear localization signals. DNA polymerization in E. histolytica is inhibited by aphidicolin, which is an inhibitor of family B DNA polymerases and is weakly inhibited by ddNTPs [46], [47]. As EhDNApolA readily incorporates ddNTPs (Figure 3B) and family A DNA polymerase are not inhibited by aphidicolin, EhDNApolA should not play a preponderant role in DNA replication of E. histolytica's genome.
In this work we report the cloning and biochemical characterization of a family A DNA polymerase present in E. histolytica. Although E. histolytica contains a mitocondrial remnant organelle dubbed mitosome, this organelle does not contain DNA. Furthermore, the genome of E. histolytica does not contains a phage-type RNA polymerase and DNA helicase involved in transcription and replication of mitochondrial DNA [9]. EhDNApolA may have evolved from the ancestral mitochondrial DNA polymerase γ or was acquired by horizontal gene transfer from a bacterial family A DNA polymerase. The fact that EhDNApolA is biochemically related to DNA polymerase ν may be a case of convergent evolution as DNA polymerases of subfamily N are only present in vertebrates [19], [48]. Thymine glycol is a DNA lesion formed by chemical oxidation and ionizing radiation [49]. E. histolytica is subject to reactive oxygen species produced at the colonic tissue and by phagocyte release [4], [50]. In eukaryotic organisms, thymine glycol can be bypassed by DNA polymerases κ and η [40], [51]. However, E. histolytica lacks those DNA polymerases. E. histolytica contains genes for Base Excision Repair including functional 8-oxo guanosine and thymine glycol glycosylases (Garcia et al, manuscript in preparation). Although, the in vivo function of EhDNApolA is unknown, its abilities to bypass thymine glycol and nuclear localization suggest a possible role of this enzyme in translesion DNA synthesis. This role is reminiscent of family A DNA polymerases of Arabidopsis thaliana postulated to be involved in DNA repair at the chloroplast [52] and eukaryotic family A DNA polymerases ν and θ [19], [24].
|
10.1371/journal.pbio.1001432 | Structural Mechanism of ER Retrieval of MHC Class I by Cowpox | One of the hallmarks of viral immune evasion is the capacity to disrupt major histocompatibility complex class I (MHCI) antigen presentation to evade T-cell detection. Cowpox virus encoded protein CPXV203 blocks MHCI surface expression by exploiting the KDEL-receptor recycling pathway, and here we show that CPXV203 directly binds a wide array of fully assembled MHCI proteins, both classical and non-classical. Further, the stability of CPXV203/MHCI complexes is highly pH dependent, with dramatically increased affinities at the lower pH of the Golgi relative to the endoplasmic reticulum (ER). Crystallographic studies reveal that CPXV203 adopts a beta-sandwich fold similar to poxvirus chemokine binding proteins, and binds the same highly conserved MHCI determinants located under the peptide-binding platform that tapasin, CD8, and natural killer (NK)-receptors engage. Mutagenesis of the CPXV203/MHCI interface identified the importance of two CPXV203 His residues that confer low pH stabilization of the complex and are critical to ER retrieval of MHCI. These studies clarify mechanistically how CPXV203 coordinates with other cowpox proteins to thwart antigen presentation.
| Viruses encode a wide array of proteins whose principle function is to disable the surveillance and effector functions of the immune system. A common viral target is the MHC class I antigen processing and presentation pathway, which is a potent mechanism used by the host for the detection and killing of virally infected cells. In this study we have delineated the immune evasion mechanism of the cowpox-encoded CPXV203 protein, which is known to potently block the normal trafficking of MHCI from the endoplasmic reticulum (ER) to the plasma membrane. CPXV203 does this by highjacking an ER-retrieval system that usually serves to capture defective, chaperone complexed MHCI proteins in the Golgi and send them to the ER. We show that CPXV203 adopts a compact beta-sandwich structure and engages evolutionarily conserved MHCI determinants that are located under the peptide-binding platform. The viral protein binds a variety of different MHCI proteins weakly at the pH found in the ER, but the affinity and half-life are significantly augmented at the more acidic conditions found in the Golgi. Together these data suggest that CPXV203 works cooperatively with the endogenous ER retrieval process to promiscuously target fully assembled MHCI, thereby preventing T-cell killing of cowpox infected cells.
| Detection of viral infection by CD8 T cells relies on major histocompatibility complex class I (MHCI) presentation of virally derived peptides at the cell surface. Not surprisingly, a wide variety of viruses have evolved mechanisms to disrupt antigen presentation by targeting the assembly and trafficking pathways used by MHCI proteins [1],[2]. The most common immune evasion mechanism appears to be blockade of cytosol-to-endoplasmic reticulum (ER) peptide transport by the transporter associated with antigen processing (TAP) [3]–[10]. However, other viruses target molecular chaperones to impair the quality of peptide loading without curtailing peptide supply [11],[12]. The quality of peptide loading by MHCI is initially controlled by the peptide loading complex (PLC) made up of TAP, tapasin (Tpn), ERp57, and calreticulin (CRT) [13]. Prior to binding an optimal peptide, the PLC retains in the ER nascent MHCI heavy chains (HCs) assembled with beta-2 microglobulin (β2m). Within the PLC, the MHCI-dedicated chaperone Tpn bridges the HC/β2m complex with TAP. Once a peptide of suitable affinity binds to the HC/β2m complex, the fully assembled MHCI is released from the PLC to transit to the cell surface; and perhaps not surprisingly, there are examples of viral immune evasion proteins that impair peptide loading by blocking PLC assembly [11],[12]. In addition to PLC-imposed quality control, non-PLC-associated CRT uses a KDEL-dependent mechanism to retrieve suboptimally loaded MHCI from the early Golgi to the ER to improve peptide binding [14]. This ER retrieval is dependent upon the C-terminal KDEL sequence of CRT that confers binding to the KDEL receptor (KDELR) in the early Golgi in a pH-dependent manner [15].
Several viral immune evasion proteins appear to directly target MHCI proteins, but only adenovirus (AdV) E3-19K and human cytomegalovirus (HCMV) US2 have been shown to directly bind MHCI luminal domains [16],[17]. E3-19K impairs MHCI egress from the ER by either an ER-retention mechanism dependent on its cytoplasmic tail [18] or its ability to prevent Tpn bridging MHCI to TAP [11], while US2 targets MHCI for ER-associated degradation (ERAD) [19]. E3-19K and US2 both exhibit distinct class Ia allele preferences [20]–[22] that may help these viruses evade natural killer (NK) cell cytotoxicity on the basis of missing self [23]. Alternatively, viruses may encode separate proteins to undermine NK cell surveillance [24]. Interestingly, E3-19K has also been reported to target various MHCI assembly intermediates, and mutagenesis analyses suggest that E3-19K may interact with an MHCI surface similar to that bound by US2 [20],[22]. The only structural study of direct MHCI sabotage revealed that US2 uses an Ig-like fold to bind under the MHCI-binding platform near where the N-terminus of the peptide is anchored [25]. Although US2 was crystallized bound to fully assembled MHCI, cellular studies suggest US2 also targets HC before full assembly with peptide and/or β2m [26]. In any case, the structural basis for how US2, E3-19K, or any other viral immune evasion protein discriminates MHCI alleles and/or assembly intermediates has not been previously reported.
While many viruses exhibit strict host specificity, some orthopoxviruses are able to productively infect a wide variety of mammalian species and encode an array of immunomodulatory genes that target both cell intrinsic and extrinsic antiviral responses [27]. Yet until recently, orthopoxviruses were not known to target antigen presentation. The orthopoxvirus cowpox (CPXV) expresses two unrelated immune evasion proteins, CPXV012 and CPXV203 (UniProt [UNP]: Q8QMP2), which use different mechanisms to block antigen presentation in both human and murine cells [28]–[30]. CPXV012 is a small type II transmembrane protein that blocks peptide transport by TAP [29],[30]. By contrast, CPXV203 is a soluble protein that prevents MHCI proteins from trafficking to the plasma membrane by a mechanism dependent upon its C-terminal KTEL sequence, a motif recognized by the KDELR [28]. To initially probe binding partners, Byun et al. (2007) showed that CPXV203 co-precipitated with MHCI and not TAP. These findings implied that CPXV203 binds MHCI lumenal domains or an associated molecule before and/or after peptide assembly [28]. Furthermore, CPXV203 was found to down regulate MHCI proteins in both murine and human cell lines during normal poxvirus infection [29],[30]. This ability to broadly inhibit MHCI by CPXV203 may help explain productive CPXV zoonotic infection of various mammalian species other than small rodents, the apparent CPXV host reservoir [27]. Indeed, mutant cowpox viruses lacking both CPXV012 and CPXV203 demonstrate attenuated virulence in a cytotoxic T lymphocyte (CTL)-dependent manner [29], in contrast to other viral proteins that target MHCI that do not appear to significantly modulate primary infection in vivo [31],[32].
Here we provide a precise understanding of how CPXV203 binds to a broad array of MHCI complexes that includes both classical and non-classical molecules. Biosensor studies indicate that CPXV203 binds MHCI weakly at the pH found in the ER, but the affinity and half-life are significantly augmented at the more acidic conditions found in the Golgi. Crystallographic analysis reveals that CPXV203 adopts a β-sandwich topology reminiscent of poxvirus chemokine-binding proteins, and this domain engages evolutionarily conserved MHCI determinants available only on fully assembled MHCI. We also undertook mutagenesis analysis that supports the structural model and uncovered the critical functional role played by two CPXV203 His residues in the pH regulation of complex stability. Together these data suggest that CPXV203 works cooperatively with the endogenous KDEL-mediated Golgi retrieval process to promiscuously target fully assembled MHCI, thereby preventing T-cell killing of cowpox infected cells.
To ascertain which MHCI assembly state(s) is targeted by CPXV203, association with the HC of murine H-2Kb (UNP: P01901) was monitored by co-precipitation in wild-type and β2m-deficient cells. CPXV203 only co-precipitated with Kb HC in cells expressing β2m (UNP: Q91XJ8) (Figure 1A), suggesting that heterodimer assembly is required for CPXV203/MHCI association. To further assess whether this association was dependent upon the PLC, CPXV203 was expressed by transduction in cells lacking either TAP or Tpn, which present low levels of fully assembled MHCI. As shown in Figure 1B, CPXV203 dramatically reduced MHCI surface expression in cells lacking TAP or Tpn, whereas the TAP inhibitor CPXV012 did not affect surface expression in these PLC-component deficient cells. We also found that CPXV203 comparably downregulates MHCI expression in cells with and without CRT (Figure 1C), suggesting that CPXV203 expression does not grossly disrupt CRT-associated ER quality control as could potentially occur through KDELR competition. In further support of this conclusion, CPXV203 does not interfere with PLC assembly, as shown by normal TAP/Tpn association and normal steady-state levels of CRT (Figure 1D). Previous studies found comparable peptide loading in cells with and without CPXV203, and failed to identify association of CPXV203 with the PLC [28]. Taken together, these findings provide compelling evidence that CPXV203 regulates the surface expression of fully assembled MHCI after dissociation from the PLC without impairing PLC function.
We next sought to examine whether CPXV203 directly binds to MHCI using soluble recombinant proteins in biophysical assays. We observed that CPXV203 binds Kb with an affinity of KD,Kin = 480 nM at pHER 7.4 using surface-plasmon resonance (SPR) (Figure 2A). The expansion of these studies to additional MHCI molecules revealed that CPXV203 exhibits low affinity interactions (KD,Kin = 82–10,500 nM) with five different murine Ia alleles (Db, Dq, Kb, Kd, Ld) and a primate allele (Ceat-B*12) (Table S1). We also examined a non-classical MHC Ib protein, murine thymic leukemia tumor antigen or TL (T3b), which was engaged by CPXV203 with similar affinity and kinetics at pHER 7.4 as observed for Kb. Unlike classical MHCI proteins that require peptide loading to assemble, TL pairs with β2m and is stable in the absence of ligand binding. Thus, it appears that the requirement for peptide binding to classical MHC Ia proteins for CPXV203 engagement is based on the role peptide loading plays in assembly and stability rather than direct recognition. Promiscuous CPXV203/MHCI association fits well with the previously published data that CPXV203 downregulates murine H-2D and -K alleles, though the affinities were weaker than those previously reported for the viral ER retention protein E3-19K (11–18 nM) [22].
The weaker than expected affinity of CPXV203 for MHCI led us to evaluate a variety of buffer conditions that might more closely reproduce ER/Golgi conditions (divalent cations: Ca2+, Mg2+, Zn2+; ATP; pH 6–8). Of these changes, only low pH augmented CPXV203/MHCI affinity with a decrease to pHGolgi 6.0 increasing the affinity ∼50-fold (KD,Kin = 10 nM, Figure 2B). This striking enhancement occurs through both an increased on-rate (ka) and a decreased off-rate (kd) for all tested murine and primate alleles (Table S1). We confirmed these results using a separate biophysical technique, biolayer interferometry (BLI), where the equilibrium response of CPXV203 binding murine (H-2Dk, -Kb,k, -Ld; TL) and primate MHCI (Mamu-A*01, Patr-B*0802) was monitored as a function of pH (7.6–6.0) (Figure 2C). These binding studies demonstrate that the stability of CPXV203/MHCI complexes is pH regulated to favor association in the Golgi rather than the ER. Importantly, similar observations have been made for the binding of KDEL bearing ligands to the KDELR [15].
To address whether enhanced binding of CPXV203 to MHCI at low pH results from changes in the stoichiometry of the complex, multi-angle light scattering (MALS) experiments were undertaken that demonstrated a 1∶1 stoichiometry that was insensitive to pH manipulation from pH 6.5–8.5 (Figure S1). We also undertook circular dichroism spectra analysis that indicated that the conformation of these proteins (alone or in complex) does not change significantly as a function of pH (unpublished data). These results support the 1∶1 binding model used in our biosensor analysis and suggest that CPXV203/MHCI pH regulation likely involves only small local effects.
We next pursued crystallographic studies of CPXV203 in complex with MHCI to better understand the nature of the interaction. Utilizing the observation that CPXV203 binding affinity increases with decreasing pH, we crystallized SeMet-labeled CPXV203 in complex with Kb loaded with SIINFEKL (OVA257–264) at pH 5.55 and determined the structure at 3.0 Å resolution (Figure 3A; Table S2). Initial molecular replacement phases using MHCI alone were greatly improved through cross-crystal averaging (Figure S2A–S2D) [33], which allowed a preliminary backbone trace of CPXV203 to be built. Subsequently, molecular replacement-single-wavelength anomalous dispersion (MR-SAD) was used to identify eight SeMet sites and introduce anomalous phase information (figure of merit [FOM] 0.604) that improved map quality to the point where the complete CPXV203/MHCI complex could be built and refined. The structure reveals that CPXV203 binds below the MHCI peptide-binding platform, contacting both the HC (α2- and α3-domains) and β2m. Comparison of Kb free and bound by CPXV203 did not indicate any significant changes associated with viral protein engagement. The general footprint located below the α2-1 helix of Kb is supported by serological experiments whereby we determined whether CPXV203 competed with monoclonal antibodies specific for well-characterized epitopes. Direct binding competition was observed for two monoclonal antibodies (MAbs) (AF6-88.5.3 and Y-3) that have been mapped precisely to this region, and no competition was observed for three MAbs mapped outside of the CPXV203 footprint (Figure S2F–S2H). We note that while HCMV US2 also binds MHCI below the peptide-binding platform, the CPXV203 footprint is completely distinct and, strikingly, overlaps with that of Tpn, CD8, and NK cell receptors (NKRs) (Figure 3B).
The structure of CPXV203 does not resemble any structurally characterized viral or host protein known to interact with MHCI. The single domain of CPXV203 (209 aa) is a globular β-sandwich that is stabilized by five disulfide bonds conserved in all T4 poxvirus protein family members (Figures 4A, 4B, S3A). The core β-sandwich consists of two parallel β-sheets (β-sheet I: β1, β5, β6, β10; β-sheet II: β2, β3, β4, β7, β8, β9) made up of anti-parallel strands with one parallel strand interaction (β7/β9) bridging the two segments of β-sheet II (Figure 4B). Three of these disulfide bonds appear to stabilize the h4-loop-h5 arrangement used to engage the MHCI α2-domain. A search for structurally similar proteins indicates that the CPXV203 β-sandwich core resembles the structurally characterized poxvirus chemokine binding proteins (CKBPs), such as the vCCI-like protein encoded by ectromelia virus, EVM001 [34], which exhibits an RMSD of 3.0 Å for 143 aligned residues (Figure S3B; Table S3).
CPXV203 and the poxvirus CKBPs engage their ligands using completely distinct binding surfaces located on opposite faces of the β-sandwich core (Figure 4C). While vCCI-like proteins use β-sheet II to bind chemokines, CPXV203 primarily uses β-sheet I elements. Interestingly, the Ectromelia virus CrmD-CTD (SECRET domain) also appears to use β-sheet I elements to bind chemokines, and it shares with CPXV203 a distinct β7–β9 junction relative to vCCI-like proteins that increases β-sheet I accessibility through the conversion of a flexible loop into a β-sheet II strand (Figure 4D). CPXV203 has further differences with the vCCI core with the replacement of vCCI β13-β14 with two α-helices (h4 and h5), a modification that also exposes CPXV203 β10 to interact with β2m. For CPXV203, these topological changes remove potential steric clashes (Figure S3C), increase solvent-accessibility of the conserved β5–β6 loop (source of nearly all α3 contacts), and create the primary sources for both α2 (h4–h5) and β2m (β8 and β10) contacts (Figure 4B; Table S4). Thus, while CPXV203 is clearly structurally related to poxvirus CKBPs, significant modifications are clearly evident that uniquely allow it to recognize MHCI.
To understand the structural basis of how CPXV203 interacts with such diverse MHCI-family proteins, we analyzed the conservation of MHCI contacts and the similarity of these contacts to those used by other MHCI-binding proteins. CPXV203 promiscuously binds MHCI through a large, somewhat nonpolar interface divided into three distinct contact regions (α2, α3, and β2m domains) (Figures 5A, 5B, S4A, S4B). The arrangement of these contact regions is only available in fully assembled MHCI, and as such CPXV203 binds MHCI in an assembly-dependent manner that is peptide-independent as long as MHCI assembly is also peptide-independent, as is the case for TL. Comparison of the CPXV203/MHCI interface to similar interfaces reveals that the total buried surface area (BSA) is significantly larger than most other complexes, CPXV203 buries >200 Å2 more main-chain (MC) than any similar MHCI-binder, and only CPXV203 divides its interface nearly equally among the platform (α1/α2), β2m, and α3 (Figure 5A; Table S5). CPXV203 recognizes MHCI elements that are extremely well conserved in all murine pMHCI: overall, 86%; CPXV203 contacts, 91%; CPXV203 side chain (SC) contacts, 95%; five invariant SC contacts. Further, CPXV203 recognizes core structural features of the MHCI fold by anchoring each of the three domain interfaces through a buried MC-MC hydrogen bond and two to three MC-SC hydrogen bonds (Figure S4C–S4E; Table S4). Through these contacts, CPXV203 recognizes seven backbone positions conserved by the MHCI fold and coordinates conserved MHCI side chains within the α3 interface (Q226, D227, E229) also required for Tpn and CD8 association [35]–[37]. Further, the presence of CPXV203 His residues opposite negatively charged α3 domain residues (H75-D227, H80-E229) suggests these may be pH-regulated interactions, though only H80-E229 is close enough to form a direct contact (3.5 Å versus 7.3 Å). Finally, we have identified that CPXV203 downregulates the non-classical MHCI molecule H2-M3, while mCD1d escapes CPXV203 retrieval (Figure S4F, S4G). Our structural results support the idea that escape by mCD1d is facilitated in part by a charge reversal at position 229 (mCD1d H233 – CPXV203 H80) and the orientation of mCD1d Q230 away from the interface due to an altered CD-loop conformation (Figure S4H). This structural investigation explains promiscuous MHCI retrieval by CPXV203, as it specifically targets a tri-domain interface of evolutionarily conserved contacts that would only be presented by fully assembled MHCI.
We assessed the functional relevance of specific determinants within the CPXV203/MHCI binding interface by extensive mutagenesis of both CPXV203 and Kb. Mutants were assayed for loss of function by rescue of surface Kb expression or lack of physical association by co-immunoprecipitation (co-IP). Single mutations in either Kb or CPXV203 from all three interaction sites (Figure 5) were tested, but only α3 interface mutations Kb E229Y and CPXV203 F76A significantly rescued Kb surface expression (Figure 6A, 6B). Furthermore, double mutations within the α3 interface (Kb D227K/E229Y, CPXV203 H75A/H80A, Kb Q226A/CPXV203 F76A, Kb E229Y/CPXV203 H75A, Kb E229Y/CPXV203 F76A) or the simultaneous mutation of interfaces α2 and α3 (CPXV203 Y161A, F76A) significantly enhanced Kb rescue, with some mutants displaying complete ablation of CPXV203 function (Figure 6A–6C). Physical association (CPXV203-HA/Kb) was impaired more dramatically than Kb rescue by single α3 interface mutations (Figure 6D), though it should be noted that the HA-tag might impair association. In any case, these experiments clearly demonstrate the functional importance of our structurally defined interface in CPXV203-mediated MHCI association and retrieval.
To extend these findings, biosensor studies were undertaken to probe the contribution of individual interface residues in binding and pH regulation. Equilibrium analysis (BLI, pHER 7.4) of CPXV203 and MHCI mutants further confirmed the three-site binding footprint (Figure 5) and clearly distinguished the CPXV203 binding site from those of E3-19K, US2, and MAb Y-3 (Table S6). Alanine mutation of several CPXV203 residues within the α3-domain interface (including His residues 75 and 80) had a pronounced deficit in binding (similar to co-IP). CPXV203 H75 and H80 were selected for kinetic analysis (SPR) based on their chemical environment (Figure S4E) and previously described sensitivity to alanine mutation (functional and association). Alanine mutation of either His ablated the off-rate (kd) enhancement at low pH, while maintaining a similar on-rate (ka) enhancement (Figure 6E; Table S7). Thus at low pH, α3 interface His residues act to extend the CPXV203/MHCI half-life (6 s–73 s), while a separate interaction appears to regulate the faster association (ka) observed at low pH. Consistent with their importance, the double mutant (H75A, H80A) displayed extremely weak affinity that prohibited accurate kinetic analysis, though equilibrium BLI assays clearly support the greater functional deficit observed for this mutant (Table S6). These investigations indicate that CPXV203 engages MHCI through critical pH-regulated interactions with conserved MHCI α3-domain determinants, while the α2 and β2m domain interfaces may enable CPXV203 to bind fully assembled MHCI with broad specificity.
Viral infection of mammalian hosts can be greatly facilitated by viral proteins that confer the ability to evade CTL detection and clearance. Not surprisingly, viruses have evolved a wide variety of strategies to reduce cell surface presentation of viral peptides on MHCI [1],[2]. The cellular, structural, and biophysical results reported here provide a complete picture of one such strategy, as CPXV203 was shown to directly bind fully assembled MHCI in a manner that is regulated via the normal pH gradient that exists between the ER and Golgi compartments (Figure 7). Though CPXV203 makes contacts to three distinct MHCI domains, pH regulation of the complex half-life is critically dependent on CPXV203 His residues that bind to an α3-domain acidic CD loop important for both Tpn and CD8 association. Thus, CPXV203 exploits a cellular pathway to target MHCI surfaces critical for immunological function in a manner that selects for those MHCI molecules most likely to present viral peptides.
Remarkably, CPX203 is not related to any other MHCI-binding protein, but rather it is most structurally related to poxvirus CKBPs. To our knowledge, CPXV203 is the first member of the large T4 poxvirus protein family [38] to be structurally characterized, suggesting a previously underappreciated link between the poxvirus CKBPs and T4 protein families through similarities in their β-sandwich core. Regardless of the evolutionary history, the adaptation of this protein domain to structurally distinct ligands and unrelated functional outcomes suggests the integral role that CPXV203 plays in antigen presentation disruption may not be its only function.
CPXV203 evolved into a promiscuous MHCI-binding protein by targeting MHCI determinants that are largely conserved by virtue of their roles in the recognition by host factors essential to cellular immunity (Tpn, CD8, NKRs). For instance, many α2-domain contacts (R111, Q115, E128, T134) are conserved through Tpn (α2 128–136) [37],[39] and Ly49 (R111, Q115, D122) [40] interactions, while the β2m contacts primarily involve structurally conserved backbone positions within the LIR-1/MHCI interface [41]. The direct overlap of CD8 and Tpn contact sites (Q226, D227, E229) [35]–[37] in the acidic CD loop of the α3 domain is clearly exploited by CPXV203 for MHCI binding, and this interface is precisely where we have identified two His residues in the viral protein that regulate increased kinetic stability at the lower pH of the Golgi.
Previous investigations of pH-dependent endosomal (PRL/PRLr [42], FcRn/Fc [43]) and ER→Golgi (RAP/LRP) [44] trafficking have repeatedly identified His residues as the pH sensitive component of these regulatory mechanisms. Unlike other amino acids, histidine is well suited to serve this function, as small pH shifts can drastically change the charge and hydrogen-bonding potential of this residue. As such, our investigation of CPXV203/MHCI pH regulation focused on interface histidines, which revealed a significant contribution of CPXV203 H75 and H80 to complex half-life at low pH. We suggest that these titratable His residues endow CPXV203 with the ability to regulate fully assembled MHCI in a manner that is complementary to the regulation of PLC-associated MHCI by CPXV012.
The specific binding of CPXV203 to fully assembled MHCI proteins in a pH-dependent manner clarifies mechanistically how CPXV203 coordinates with CPXV012 to effectively block antigen presentation. Previous characterizations showed the CPXV012 functions in a PLC-dependent fashion to block TAP transport of peptide into the ER [29],[30]. However, some MHCI-binding peptides in the ER are not TAP-dependent and the CPXV012 block of peptide transport is likely not absolute. The MHCI proteins that are able to bind peptide in the presence of CPXV012 are left to CPXV203, since it binds fully assembled MHCI through domain-specific conformational determinants conserved in classical and many non-classical MHCI. Among these interactions, the α3 interface is particularly important based on the presence of CPXV203 His residues that impart pH regulation to the CPXV203/MHCI interaction. This pH dependence suggests that CPXV203/MHCI interacts most avidly in the Golgi and not the ER, thus limiting the pool of MHCI that CPXV203 must retrieve.
Interestingly, CRT has a C-terminal KDEL sequence conferring ER retrieval, and non-PLC-associated CRT has recently been implicated in quality control of MHCI peptide loading [14]. More specifically, CRT was shown to accumulate in the cis-Golgi and return peptide accessible MHCI proteins to the ER. Both CRT and CPXV203 retrieve MHCI proteins but with opposite goals. CRT functions in host quality control by retrieving MHCI with suboptimal peptides, whereas CPXV203 functions in immune evasion by retrieving fully assembled MHCI to block antigen presentation. Thus CPXV012 and -203 act sequentially to efficiently block MHCI expression using PLC-dependent versus PLC-independent mechanisms, respectively. As a possible consequence of efficient MHCI downregulation resulting in NK cell susceptibility, CPXV expresses the soluble class I-like protein OMCP that functions as a competitive antagonist of the NKG2D-activating receptor [24]. Indeed, the combined sabotage of both CTL and NK cell detection of virus-infected cells explains why mutant CPXV lacking CPXV012 and 203 demonstrates attenuated virulence in vivo compared to wild-type virus [29].
MHCI-specific MAbs used in SPR competition assays were obtained from the ATCC (H-2Kb: 25-D1.1.6, B8-24-3, Y-3), purchased from BioLegend (Kb: AF6-88.5.3, Kb/Db: 28-8-6, Kd/Dd: 34-1-2S, hβ2m: 2M2), or provided as a kind gift (Kb: 5F1-2-14) from S. Nathenson (Albert Einstein College of Medicine, New York) and L. Pease (Mayo Clinic, Minnesota). MAbs that were not purchased from BioLegend were purified from ascites on a Bio-Rad Profinia FPLC using Protein A or G. MAbs used in flow-cytometry and IP assays have been described previously. MAb footprints in Figure S2H are based on SPR data from this work and available literature (Text S1).
Peptides were synthesized by Fmoc chemistry and then subjected to reverse-phase HPLC for purification. Peptides were resuspended at >1 mM in ddH2O, DMSO, or 6M GuHCl, as dictated by peptide solubility. Peptides were chosen based on available MHCI crystal structures or personal suggestions by A. Stout (NIH Tetramer Core Facility). See Text S1 for a list of all peptides used in this study.
Murine embryo fibroblast (MEF) B6/WT3 (WT3) and mutant MEFs including TAP1-deficient cells (FT1−), Tapasin-deficient cells (Tpn−/−), calreticulin-deficient cells (CRT−/−), β2m-deficient cells (B6.B2M−) and triple knockout fibroblasts (Kb−/− Db−/− β2m−/−; 3KO) were all derived from C57BL/6 (H-2b) embryos and have been described previously [45]. The CPXV203 and Kb mutants were stably expressed in the indicated cells by retroviral expression vectors pMXsIG [28] and pMIN [45], respectively. Cells transduced by pMIN were selected by neomycin while green fluorescent protein (GFP+) cells from pMXsIG transduced lines were enriched by cell sorting.
For co-IPs (TAP1/Tpn and CPXV203-HA/H-2Kb), cells were lysed in PBS with 1.0% digitonin (Wako) and protease inhibitor cocktails (Roche) for 60 min. Post-nuclear lysates were then incubated with indicated antibodies + protein A-sepharose (Sigma) or anti-HA sepharose (sigma) for HA-tagged CPXV203 for 1 h. After washes, coprecipitated proteins were eluted by boiling in lithium dodecyl sulfate (LDS) sample buffer (Invitrogen). For cross-linking treatment (CPXV203-HA/H-2Kb), cells were incubated with 1–2 mM DSP (Thermo) in PBS for 2 h at 4°C. The cross-linking was terminated with 25 mM Tris-HCl pH 7.4 before the cells were lysed in PBS with 1.0% NP-40. Following immunoprecipitation cross-linked proteins were eluted by boiling in LDS sample butter with 2.5% β-mercaptoethanol. Immunoblot of precipitated proteins was performed following SDS-PAGE separation. Specific proteins were visualized by chemiluminescence using the ECL system (Thermo).
All flow cytometric analyses were performed using a FACS Calibur (Becton Dickinson). Data were analyzed using FlowJo software (Tree Star). Staining was performed as described [46]. PE-conjugated goat anti-mouse IgG (BD Pharmingen) was used to visualize MHCI staining. PE-conjugated anti-mouse CD1d (eBioscience) was used to detect surface CD1d. GFP signal representing CPXV203 transduced cells were collected in the FITC channel.
Mammalian CPXV203ΔKTEL (aa 1–205, etgMVI-LHV), bacterial CPXV203ΔKTEL (aa 1–205, maMVI-LHV), and bacterial MHCI were produced using established methods (Text S1). Kb (aa 0–280, mGPH-PST)/H-2Kd (aa 0–283, mGPH-VSN) constructs were produced in house, while H-2Dd, H-2Dk, H-2Dq, H-2Kk, H-2Ld, Mamu-A*01, Patr-B*0802, H-2Q9-H-2Db, H-2Kb-HLA-A*0201, HLA-A*0201-H-2Kb were produced by the NIH Tetramer Core Facility. H-2Db biotinylated monomer (LCMV Gp33, KAVYNFATC) was purchased from Beckman Coulter. Biotinylated constructs included a C-terminal site-specific biotinylation tag and were biotinylated by following established procedures (Avidity). Chimeric mCD1d-Fc produced in a murine cell line was obtained from R&D Systems. All MHCI include human β2m (hβ2m, aa 0–99, mIQR-RDM, UNP: P61769) unless otherwise noted. Signal-peptide/cloning artifacts are indicated as lower-case aa.
CPXV203 (Brighton Red strain, SeMet-labeled)/MHCI(OVA257–264:H-2Kb:hβ2m) complex was prepared for crystallization by size-exclusion chromatography purification of CPXV203/MHCI in low pH/salt buffer (50 mM NaCl, 30 mM MES pH 5.6, 0.01% Azide). Diffraction-quality crystals of CPXV203/MHCI were grown at 20°C by streak seeding into hanging drops of 0.5 µl complex (7.5 mg/ml) +0.5 µl reservoir solution (10% PEG 6000, 4% glucose, 2% ethylene glycol, 0.1 M tri-K citrate pH 5.55, 0.01% Azide). Crystals were dehydrated, flash frozen in liquid N2, and then used for X-ray data collection at the Advanced Light Source (ALS) beamline 4.2.2 (0.97909 Å wavelength). Crystals belong to space group P1 (a = 88.31 Å, b = 88.25 Å, c = 106.42 Å; α = 76.18°, β = 69.29°, γ = 66.69°) with four CPXV203/MHCI complexes per asymmetric unit (ASU) and a solvent content of 52%. The HKL2000 software package [47] was used to index, integrate, and scale the data, yielding an 87.9% complete dataset at 3.0 Å (R-sym = 16.4%) (Table S2). The structure of H-2Kbm8:mβ2m (2CLZ) [48] was used for molecular replacement (MR) using Phaser within the Phenix suite [49], with four MHCI proteins located within the ASU. Electron density for the unique CPXV203 region was improved by cross-crystal averaging [33] allowing an initial model of CPXV203 to be manually built in Coot [50]. MR-SAD using AutoSol (Phenix suite [49]) was subsequently used to identify eight SeMet sites in the four CPXV203 monomers, enabling the introduction of anomalous phase information (figure of merit [FOM] 0.604) that improved map quality allowing for complete CPXV203/MHCI model building. Phenix Refine [49] was used with global non-crystallographic symmetry (NCS) restraints to refine the CPXV203/MHCI structure to a final Rwork of 22.9% and Rfree 25.3% (see Table S2 for complete crystallographic statistics). The final CPXV203/MHCI model contains four CPXV203/MHCI complexes and 72 water molecules. Each complex contains the nearly complete CPXV203 (mature residues 5–190), Kb (mature residue 1–277), human β2m (0M-99), and OVA257–264. All structure figures were created using PyMOL [51]. See Text S1 for comprehensive crystallization, data collection, structure determination, refinement, structural analysis, and figure preparation details.
SPR experiments were run on a Biacore T100 (GE Healthcare) in either standard HBS-EP+ (pH 7.4) or low pH MBS-EP+ (pH 6.0). Kinetic and equilibrium analyses were performed using Biacore T100 Evaluation software using a 1∶1 Langmuir model. BLI experiments were performed on an Octet RED system (ForteBio) using HBS-EP+/MBS-EP+ supplemented with 0.05% (v/v) TWEEN and 1% BSA. Equilibrium BLI data was analyzed using Octet software (V7.0). All biosensor experiments were run at 25°C and followed proper biosensor experimental technique. Size-exclusion chromatography-multi-angle light scattering (SEC-MALS) experiments were run on a Dawn HELEOS-II 18-angle light scattering detector (Wyatt) and Optilab rEX refractive index monitor (Wyatt) linked to a Waters HPLC system. Dynamic light scattering (DLS) was performed on a DynaPro-801TC. Circular dichroism was measured using a Jasco-810 instrument (Jasco Inc.). Detailed methodologies for biosensor, light scattering, and circular dichroism experiments are available in Text S1.
|
10.1371/journal.ppat.1000380 | Haemonchus contortus Acetylcholine Receptors of the DEG-3 Subfamily and Their Role in Sensitivity to Monepantel | Gastro-intestinal nematodes in ruminants, especially Haemonchus contortus, are a global threat to sheep and cattle farming. The emergence of drug resistance, and even multi-drug resistance to the currently available classes of broad spectrum anthelmintics, further stresses the need for new drugs active against gastro-intestinal nematodes. A novel chemical class of synthetic anthelmintics, the Amino-Acetonitrile Derivatives (AADs), was recently discovered and the drug candidate AAD-1566 (monepantel) was chosen for further development. Studies with Caenorhabditis elegans suggested that the AADs act via nicotinic acetylcholine receptors (nAChR) of the nematode-specific DEG-3 subfamily. Here we identify nAChR genes of the DEG-3 subfamily from H. contortus and investigate their role in AAD sensitivity. Using a novel in vitro selection procedure, mutant H. contortus populations of reduced sensitivity to AAD-1566 were obtained. Sequencing of full-length nAChR coding sequences from AAD-susceptible H. contortus and their AAD-1566-mutant progeny revealed 2 genes to be affected. In the gene monepantel-1 (Hco-mptl-1, formerly named Hc-acr-23H), a panel of mutations was observed exclusively in the AAD-mutant nematodes, including deletions at intron-exon boundaries that result in mis-spliced transcripts and premature stop codons. In the gene Hco-des-2H, the same 135 bp insertion in the 5′ UTR created additional, out of frame start codons in 2 independent H. contortus AAD-mutants. Furthermore, the AAD mutants exhibited altered expression levels of the DEG-3 subfamily nAChR genes Hco-mptl-1, Hco-des-2H and Hco-deg-3H as quantified by real-time PCR. These results indicate that Hco-MPTL-1 and other nAChR subunits of the DEG-3 subfamily constitute a target for AAD action against H. contortus and that loss-of-function mutations in the corresponding genes may reduce the sensitivity to AADs.
| Worldwide, sheep and cattle farming are threatened by anthelmintic-resistant gastro-intestinal nematodes. A novel chemical class of synthetic anthelmintics was recently discovered, the Amino-Acetonitrile Derivatives (AADs), which exhibit excellent efficacy against various species of livestock-pathogenic nematodes and, more importantly, overcome existing resistances to the currently available anthelmintics. Haemonchus contortus, the largest nematode found in the abomasum of sheep and cattle, is a blood-feeding parasite that causes severe anemia that can lead to the sudden death of the infected animal; H. contortus is highly susceptible to AADs. In order to elucidate the mode of action of the AADs, we have developed 2 independent H. contortus mutants with reduced sensitivity to monepantel (AAD-1566). Both mutants were affected in their acetylcholine receptor (ACR) genes of the DEG-3 subfamily. In particular, we discovered a panel of mutations in the gene monepantel-1 (Hco-mptl-1) including deletions leading to mis-splicing, insertions and point mutations leading to premature termination of translation of the protein. These findings support the notion that Hco-MPTL-1 and other nAChR subunits of the DEG-3 subfamily are targets of the AADs. The fact that the DEG-3 subfamily of acetylcholine receptors is nematode-specific may explain the good therapeutic index of AADs in mammals.
| Throughout the world, successful livestock production of ruminants is hampered by gastro-intestinal nematodes. Haemonchus contortus in particular is responsible for substantial losses to the global sheep industry [1]. Haemonchus contortus is a blood-feeding nematode that inhabits the abomasum of sheep, producing in acute infections, severe anemia that can lead to the death of infected animals.
Broad spectrum chemotherapy against gastro-intestinal nematodes is restricted to 3 anthelmintic classes: the benzimidazoles, such as albendazole and oxfendazole, the imidazothiazoles, including levamisole and tetramisole and the macrocyclic lactones (e.g. ivermectin, moxidectin, abamectin and doramectin). The increased usage of anthelmintics has contributed to the spread of resistant nematodes with increasing reports of nematodes insensitive to most if not all of the available classes of anthelmintics [2]–[10]. In some countries in the southern hemisphere, sheep farming is severely endangered by such populations [4], further increasing the need for a new class of anthelmintic [11].
Recently, a new class of compounds, the Amino-Acetonitrile Derivatives (AADs) was discovered [12] with good tolerability in mammals and promising activity against drug-resistant nematodes. The AADs are low molecular mass compounds bearing different aryloxy and aroyl moieties on an amino-acetonitrile core [13]. Further studies [14] have allowed the selection of a drug candidate, AAD-1566 (monepantel). In order to investigate the mode of action of this new class of compounds, AAD-resistant Caenorhabditis elegans mutants were generated by EMS mutagenesis. Classical forward genetics revealed that the majority of recuperated AAD-resistant mutants carried mutations in the gene acr-23, a member of the nematode-specific DEG-3 subfamily of nicotinic acetylcholine receptor (nAChR) alpha subunits [12]. Preliminary data had already indicated an involvement of similar acetylcholine receptors in AAD action against H. contortus [12]. Here we report the identification of the gene monepantel-1 (Hco-mptl-1, formerly named Hc-acr-23H) and other members of the DEG-3 subfamily of ACR genes from H. contortus. A panel of different mutations, mis-splicing in particular, in Hco-mptl-1 transcripts from AAD-resistant worms indicates that Hco-MPTL-1 is a target for monepantel action against H. contortus.
The drug-susceptible H. contortus CRA (Hc-CRA) was received in 1984 from the Veterinary Institute of Onderstepoort, Republic of South Africa and has since been passaged in sheep 75 times. The H. contortus Howick isolate (Hc-Howick) was received from the same institute in 2001. This is a multidrug-resistant isolate that is completely resistant to albendazole, rafoxanide, morantel, ivermectin and trichlorfon [6],[15]. The isolate has been passaged in sheep 9 times since being received. The mutant lines Hc-CRA AADM and Hc-Howick AADM were selected from Hc-CRA and Hc-Howick, respectively, by in vitro exposure to increasing doses of AAD-1566 alternatively with propagation in sheep [12].
Haemonchus contortus isolates were propagated in 3–6 month old sheep (‘Blanc des Alpes’), which had been experimentally infected with the nematode. The sheep were kept in groups of 4 and housed indoors off pasture to prevent natural infection. After 14 days, they were transferred to individual cages. Starting on day 21 after infection, eggs were collected from homogenized feces and filtered several times through a 32 µm sieve. Eggs were further purified by floating on 50% sucrose solution, rinsed with water and counted microscopically.
Sheep studies were performed with approval of a Cantonal animal welfare committee (permit number FR 25A/05). Anthelmintic efficacy tests in sheep were performed according to the guidelines of the World Association for the Advancement of Veterinary Parasitology [16]. Each animal was infected intraruminally on study day −21 with 3000 L3-larvae of H. contortus (cultivated in coprocultures). On study day 0, the sheep were treated with single anthelmintics or combinations thereof as an oral drench at the recommended dose. A sheep was classified as ‘cured’ when no more eggs were counted in the feces and no adults were found in the abomasum at necropsy.
Adult worms were recovered from the abomasum of freshly euthanized sheep, washed in Hank's Buffered Salt Solution (HBSS; Invitrogen) and immediately shock-frozen in liquid nitrogen. While frozen, the worms were crushed with a Kontes pellet pestle (Fisher Scientific). The powder was resuspended in 600 µl of lysis buffer (10 mM Tris pH 7.5, 1 mM EDTA, 100 mM NaCl, 0.5% SDS, 100 µg/ml RNase A) and incubated at 37°C for 1 hour. Pronase (100 µg/ml) was added to the mixture and the tubes were incubated at 37°C until the solution became clear. The samples were extracted with equal volumes of phenol∶chloroform (1∶1) and chloroform. The DNA was ethanol precipitated, washed and resuspended in 50 µl of Tris-Cl (pH 7.5). For RNA extraction, worms were homogenized in TRIzol and processed according to the instructions of the supplier (Invitrogen). To remove DNA contamination, the RNA samples were treated with a TURBO DNA-free kit (Ambion). To generate cDNA, 1 µg of total RNA was reverse transcribed to cDNA using a d(T)30 primer and a Moloney Murine Leukemia Virus Reverse Transcriptase (MMLV RT; SMART cDNA library construction kit from Clontech).
A total of 4 µg of mRNA was isolated from a mixture of male and female Hc-CRA using a Oligotex kit from Qiagen. A cDNA library was constructed with the ZAP-cDNA Cloning kit and Gigapack III Gold packaging kit. The library was screened at high stringency (hybridization at 65°C in 5×SSC, 5× Denhardt's solution, 0.1% SDS, 0.1% sodium pyrophosphate, 100 µg/ml salmon sperm DNA; final wash at 60°C in 0.2×SSC, 0.1% SDS) with a 32P-labeled 456 bp fragment of Hco-mptl-1. This fragment had been amplified from cDNA with the primers Hco-mptl-1_frw3 and Hco-mptl-1_rev1 and cloned into pCR®2.1-TOPO® (Invitrogen). Positive phages were taken through 3 rounds of plaque purification with this probe and the phagemid (pBluescript SK+) was excised using the ExAssist helper phage in the E. coli SOLR strain. Inserts were sequenced in both directions with standard M13 forward and reverse primers and the internal primers Hco-mptl-1_frw4 and Hco-mptl-1_rev3. The sequences were read and assembled using 4Peaks (by A. Griekspoor and T. Groothuis; http://mekentosj.com).
The primers used for PCR-amplification, real-time PCR or for cDNA first strand synthesis of H. contortus nAChR genes are summarized in Table S1. For nested PCR on cDNA with spliced leader (SL) primers, the primary products were diluted 50-fold and 2 µl were used for the second PCR with nested primers. The annealing temperature was fixed at 55°C for cDNA and 58°C for genomic DNA template. PCR products were gel purified using the NucleoSpin® ExtactII kit (Macherey-Nagel) and cloned into either pGEM-T easy (Promega) or pCR®2.1-TOPO® (Invitrogen). Plasmid DNA was purified using the QIAprep Spin Miniprep Kit (Qiagen) and sequenced using the standard primers M13 forward and reverse and, if necessary, an additional internal primer to cover long products. For rapid amplification of cDNA ends by PCR (RACE-PCR), an internal reverse primer (Table S1) was combined with splice leader sequence (1 or 2) to obtain the 5′ UTR, or an internal forward primer combined with a poly-dT primer for the 3′ UTR of the transcript.
For real-time PCR, 1 µg of total RNA from adult H. contortus was used to synthesize first-strand cDNA by random priming using Superscript II reverse transcriptase (Invitrogen) in a final volume of 20 µl following the manufacturer's instructions. Reverse-transcribed material corresponding to 40 ng RNA was amplified in 25 µl MESA GREEN qPCR MasterMix Plus for SYBR Assay (Eurogentec) by using the ABI SDS7000 Sequence Detection System under the following conditions: 1 cycle of 95°C for 15 minutes followed by 40 cycles of 95°C for 15 seconds and 60°C for 1 minute. The primer pairs used for the amplification are listed in Table S1 and target the following genes: β-tubulin, Hco-mptl-1, Hco-des-2H and Hco-deg-3H. Three independent total RNA extractions were performed and each was tested in duplicate. Relative expression values were calculated according to Livak and Schmittgen [17]; a 136 bp region within the phosphoglucose isomerase gene was used for normalization, a 122 bp region within the β-tubulin gene was used as a (presumably) non-affected control, and no reverse transcriptase and no template reactions as negative controls. The specificity and identity of individual amplicons were verified by melt curve analysis and visualized on a 2% agarose gel.
In order to study the mode of action of the AADs, we used 2 mutant isolates, Hc-CRA AADM and Hc-Howick AADM selected from parent Hc-CRA and Hc-Howick isolates, respectively. Both mutant isolates showed reduced sensitivity to AAD-1566 (monepantel) in vitro [12]. To test whether the observed loss of susceptibility to AAD-1566 in vitro was relevant for the situation in vivo, Hc-CRA, Hc-Howick and their AADM derivatives were challenged in vivo with single compounds or combinations thereof; AAD-1566 and the commercial compounds were applied at their recommended doses to sheep. Sheep were infected intraruminally with Hc-CRA AADM. Following treatment with AAD-1566 at the proposed minimum dose rate of 2.5 mg/kg body weight [18] eggs were found in the feces and adults seen at necropsy (Table 1). Likewise, nematode eggs and adults were also found in sheep infected with Hc-Howick AADM larvae when treated either with AAD-1566 or albendazole or a combination of AAD-1566 and ivermectin (Table 1). The offspring from the Hc-Howick AADM isolate that survived the AAD-1566 and ivermectin treatment were cultured and challenged with albendazole and levamisole over the following generations (data not shown). Finally, Hc-Howick AADM was able to survive a full simultaneous in vivo treatment with albendazole, levamisole, ivermectin and AAD-1566, administered at their recommended doses (Table 1). Thus the reduction of sensitivity to AAD-1566 induced in vitro was also relevant in vivo for the mutant lines. The AAD-mutant H. contortus apparently did not show any alterations in motility, infectivity to sheep (determined by the numbers of adult H. contortus recovered at necropsy) or egg production, and did not exhibit any phenotype with respect to the ultrastructure (by electron microscopy) of the cuticle, head or tail.
The putative target of the AADs in C. elegans, ACR-23, is a member of the nematode-specific DEG-3 family of nAChR alpha subunits. A tblastn search [19] with DEG-3 members against the (incomplete) H. contortus genome database (http://www.sanger.ac.uk/Projects/H_contortus) returned strong hits from different contigs, coding for a total of 6 different DEG-3 subfamily nAChR subunit homologues. However, the lack of overlap between the different contigs precluded the assembly of full length coding sequences. The predicted H. contortus proteins were named Hco-MPTL-1 (accession number: contig_0024907; contig_0033952; contig_0079482; haem-240m02.q1k; contig_0053297; contig_069357), Hco-DES-2H (contig_0064641), Hco-DEG-3H (contig_0075200; contig_0075201), Hco-ACR-24H (contig_0003482; contig_0064300), Hco-ACR-5H (contig_0106281; contig_0023143) and Hco-ACR-17H (contig_0101516; contig_0101514). For Hco-MPTL-1, Hco-DES-2H and Hco-DEG-3H, full-length coding sequences were obtained by cDNA library screening or RACE-PCR, respectively (see below). Figure 1 shows the position of the H. contortus sequences in a phylogenetic tree of the DEG-3 subfamily nAChR from C. elegans, C. briggsae and Brugia malayi. Note that an incomplete sequence of Hco-MPTL-1 was previously named Hc-ACR-23H [12].
To obtain the full length coding sequence of the Hco-mptl-1 gene, a lambda phage cDNA library from mRNA of adult H. contortus was constructed and screened at high stringency with a radioactive probe from a partial Hco-mptl-1 sequence. After 3 rounds of selection, a clone with the full-length coding sequence, Hco-mptl-1, was isolated and sequenced. The Hco-mptl-1 mRNA is composed of at least 17 exons and 16 introns (1992 bp) with a short 5′ UTR and 3′ UTR (21 bases and 42 bases, respectively). The transcript is trans-spliced as the splice leader 1 (SL1) is present at its 5′ end. Interestingly, a start codon (AUG) is present in exon 1 but is followed after 8 amino acids by a stop codon in frame (UGA). This is a feature found in many other organisms [20]–[22] and it is assumed to play a role in the regulation of translation efficiency. In most cases, upstream AUGs decrease mRNA translation efficiency and have a strong, negative regulatory effect [23]. The longest open reading frame (ORF) in the Hco-mptl-1 gene is obtained when the translation is initiated at the second AUG codon in exon 3 and extends over 1695 bases. Overlapping long range PCR was performed in order to estimate the total size of Hco-mptl-1. The gene was found to be approximately 18.5 kb long with a large intron (about 7 kb) between exons 3 and 4 (see below). The predicted Hco-MPTL-1 protein consists of 564 amino acids and possesses motifs typical for Cys-loop ligand-gated ion channels, including an N-terminal signal peptide of 18 amino acids [24], 4 transmembrane domains and the Cys-loop (2 cysteines separated by 13 amino acids). Loops A to F, which are involved in ligand binding [25] are also present in the protein (Figure S1). In loop C, there are 2 adjacent cysteines, defining Hco-MPTL-1 as a nAChR alpha subunit.
As determined by PCR with gene-specific primers on genomic DNA, Hco-mptl-1 (Hco-mptl-1_frw6 and Hco-mptl-1_rev6), Hco-des-2H (Hco-des2_frw8 and Hco-des2_rev8) and Hco-deg-3H (Hco-deg3_frw1 and Hco-deg3_rev1) are present in the Hc-CRA and Hc-Howick parental isolates (Figure 2). Of the 3 products obtained for the Hco-mptl-1 gene, the smallest one (1478 bp) corresponded to Hco-mptl-1. The same primers were used for reverse transcriptase PCR on total RNA, showing that all 3 genes were expressed and spliced in L3-larvae as well as in adult nematodes (Figure 2).
The predicted Hco-MPTL-1 protein shares 48.5% identity and 66.8% similarity with C. elegans ACR-23 and 60.2% identity and 70.7% similarity with C. elegans ACR-20. The novel H. contortus nAChR was originally named Hc-ACR-23H based on a partial sequence that was most closely related to C. elegans ACR-23 [12]. In the light of the full-length sequence, this nomenclature seems to have been premature since the Haemonchus nAChR turned out to be more closely related to C. elegans ACR-20 (Figure 1). In the absence of a complete record of ACR paralogues from H. contortus, and in analogy to levamisole-insensitive (lev-) mutants in C. elegans [26], we propose to name the gene monepantel-1 (Hco-mptl-1) due to its apparent involvement in monepantel sensitivity.
In order to compare the Hco-mptl-1 sequences from the AAD-susceptible isolates and their AAD-mutant progeny, primers were designed at each extremity of the ORF (Hco-mptl-1_5′_frw3 and Hco-mptl-1_3′end_rev1) and the full length Hco-mptl-1 coding sequences amplified from cDNA from adults. A product of about 1800 bp was obtained for all isolates apart from the Hc-CRA AADM, which produced a shorter product of 1650 bp (Figure 3B). Sequencing clones of the latter revealed that they lacked either exon 4 or exon 15 (Figure 4, Hco-MPTL-1-m2 and m3). This was confirmed with primers flanking either exon 4 (Hco-mptl-1_5′_frw2 and Hco-mptl-1_rev8; Figure 3C) or exon 15 (Hco-mptl-1_frw6 and Hco-mptl-1_rev6; Figure 3D). PCR with a SL1 forward primer and a reverse primer in the Hco-mptl-1 coding sequence (Hco-mptl-1_rev1, product of about 1200 bp; Figure 3A) also produced shorter products (1000 bp and 850 bp; Figure 3A) from Hc-CRA AADM. The 850 bp product turned out to lack both exon 2 and exon 3 while the 1 kb product lacked exon 4 (Figure 4, Hco-MPTL-1-m1 and m2). The 1200 bp product was cloned from Hc-CRA AADM but contained only silent mutations compared to Hc-CRA. Loss of exon 4 caused a frame-shift leading to a premature stop of translation and a predicted Hco-MPTL-1 protein truncated at amino acid 19 (Figure 4). Loss of exon 15 also led to a premature stop codon that truncated the Hco-MPTL-1 protein at amino acid 448 (Figure 4). The mutation Hco-MPTL-1-m1 (loss of exon 2 and 3) did not cause a frame-shift but the loss of the signal peptide and the first 39 amino acids of the extracellular loop.
To understand the molecular basis of exon loss in the Hc-CRA AADM isolate, PCR primers Hco-mptl-1_frw8 and Hco-mptl-1_rev6 (Table S1) were designed to flank the mis-spliced exon 15. PCR was performed using genomic DNA as a template. Sequencing of cloned PCR products revealed a 10 bp deletion upstream of exon 15 in the Hc-CRA AADM mutant that encompasses the predicted splice acceptor site (UUUCAG; Figure 5). Presumably, the splicing machinery is not able to identify the end of intron 14 and uses the next splice acceptor site (intron 15). This would explain why exon 15 is skipped (Figure 4, Hco-MPTL-1-m3). Joining of exon 14 to exon 16 causes a frame-shift leading to a premature stop codon. With primers flanking exon 4 (Hco-mptl-1_frw10/gDNA and Hco-mptl-1_rev8; Table S1), a 323 bp deletion was detected consisting of the end of intron 3 (206 bp) and most of exon 4 (117 bp). Again, loss of the predicted splice acceptor site at the end of intron 3 may explain the observed loss of exon 4 (Figure 4, Hco-MPTL-1-m2), since the splicing machinery will use the next available splice acceptor site (intron 4), joining exon 3 and exon 5. The resulting frame-shift causes a premature stop at codon 19 (TGA), terminating translation after the signal peptide (Figure 4, Hco-MPTL-1-m2).
No obvious mutations such as mis-spliced exons were detected in the Hc-Howick AADM isolates. When sequencing the Hco-mptl-1 coding regions (SL1 and Hco-mptl-1_rev6) from both susceptible and AAD-1566-mutant Howick isolates, a transversion from G277 to T in exon 6 of the Hco-mptl-1 gene was observed that led to a premature stop codon (E93*; Figure 6). Direct sequencing of RT-PCR products (using Hco-mptl-1_frw4 and Hco-mptl-1_rev1 primers) revealed that about 80% of the Hc-Howick AADM cDNAs, as estimated from the electropherogram [27], carried a T at position 277 (Figure 6A). The point mutation underlying E93* creates a restriction site for the endonuclease BfrI (recognition site: CTTAAG) that lent itself for RFLP analysis. Only the PCR product amplified from cDNA of Hc-Howick AADM was digested by BfrI (Figure 6B). As expected from the sequencing, a small proportion (about 20%) of the product was not cut, indicating that not all of the Hco-mptl-1 genes from Hc-Howick AADM population carried the G277T mutation. When this BfrI-unrestricted product from Hc-Howick AADM was excised from an agarose gel, cloned and sequenced, a further polymorphism was detected that led to skipping of exon 8 (Figure 4, Hco-MPTL-1-m6). As this exon is very short (22 bases), it was impossible to discriminate between mutant and parental wild type PCR products (Figure 3). Loss of exon 8 causes a frame-shift leading to a premature stop codon and a predicted Hco-MPTL-1 protein truncated at amino acid 166 (Figure 4). A minority of the Hco-mptl-1 PCR products obtained from Hc-Howick AADM did not contain any major mutations. These sequences could come from AAD-susceptible individuals within the H. contortus Howick AADM populations or from AAD-mutant individuals that carry other, yet to be identified mutations.
As the DEG-3 subfamily gene Hco-des-2H has also been implicated in AAD action in H. contortus [12], we cloned and sequenced the full-length Hco-des-2H coding sequence from H. contortus cDNA by RACE-PCR. Using primers NheI_des2_frw1 and XhoI_des2_rev1 (Table S1), 2 products were obtained from the four H. contortus isolates. Cloning and sequencing revealed the smaller transcript to lack 168 bases coding for part of the internal loop between TM3 and TM4, possibly indicating alternative splicing of the Hco-des-2H gene. The predicted protein (full version) consists of 534 amino acids and shows 69% identity and 80% similarity with C. elegans DES-2. Hco-DES-2H possesses motifs typical for Cys-loop ligand-gated ion channels (4 transmembrane domains, a Cys-loop and loops A to F) and the 2 adjacent cysteines in the C-loop, defining Hco-DES-2H as a nAChR alpha subunit (Figure S2).
When comparing Hco-des-2H coding sequences (Table 2) obtained from Hc-CRA and Hc-CRA-AADM, respectively Hc-Howick and Hc-Howick-AADM, no mutation was found to correlate perfectly with AAD-susceptibility. Nevertheless, using the SL1 primer and 2 internal reverse primers (Hco-AcRa_rev3 and Hco-AcRa_rev2) in a nested PCR experiment, an insertion of 135 bp was detected in the 5′ UTR of the Hco-des-2H gene from the Hc-CRA AADM and Hc-Howick AADM isolates, creating 2 additional start codons. Both start codons are followed by an early stop codon in frame.
In the C. elegans genome, DES-2 and DEG-3 are encoded on the same operon and both subunits are co-expressed to form a functional channel [28],[29]. Performing RACE-PCR on H. contortus (adults) cDNA we identified Hco-deg-3H encoding a protein of 569 amino acids that shows 68.4% identity and 78% similarity to C. elegans DEG-3. Again, Hco-DEG-3H carried all the hallmarks of nAChR alpha subunits (Figure S3). No mutations were detected for Hco-deg-3H in the AAD-mutant H. contortus isolates compared to the parental isolates. The Hco-deg-3H mRNA carries a spliced leader type 2 (SL2) sequence at its 5′ end. To test whether Hco-des-2H and Hco-deg-3H are also on an operon in H. contortus, a long range PCR was performed using a forward primer designed at the end of Hco-des-2H (Hco-des2_frw11) and a reverse primer at the beginning of Hco-deg-3H (Hco-deg3_2r). A band of approximately 6 kb was obtained for the 4 isolates confirming that Hco-des-2H and Hco-deg-3H are encoded on a single operon. However, the distance between the 2 genes is 10 times larger in H. contortus than in C. elegans.
The steady-state mRNA levels of the DEG-3 subfamily acetylcholine receptor genes Hco-mptl-1, Hco-des-2H and Hco-deg-3H were quantified by real-time PCR (Figure 7). For the Hc-CRA AADM isolate, a small, statistically not significant (p>0.05) decrease in the mRNA level was observed for Hco-mptl-1 (−21%) and Hco-des-2H (−16%). In contrast, the relative mRNA level of the Hco-deg-3H gene was higher (69%; p<0.01) in this mutant. For Hc-Howick AADM, a significant (p<0.01) down-regulation of the 3 measured DEG-3 subfamily members was observed: −70% for Hco-mptl-1, −77% for Hco-des-2H and −92% for Hco-deg-3H. The relative expression level of the β-tubulin gene was measured in both mutant isolates as a (presumably) non-affected control. No statistically significant changes were observed.
A new chemical class of synthetic anthelmintics, the AADs, was recently discovered [12]. The AADs exhibit excellent efficacy against various species of livestock-pathogenic nematodes and more importantly, can control nematodes resistant to the currently available anthelmintics [30],[31]. To get insights into the mode of action of the new AADs, a classical ‘forward genetic’ screen for AAD-resistant C. elegans mutants was performed previously [12]. As a result, AADs were proposed to act through the nAChR ACR-23, a member of the nematode-specific DEG-3 subfamily [32]. By screening the currently available (but incomplete) H. contortus genome sequence for DEG-3 nAChR homologues, it was found that this subfamily is conserved between C. elegans and H. contortus. Six paralogous proteins out of 8 in C. elegans or C. briggsae were identified (Figure 1), in contrast to only 2 in the genome of B. malayi [33]. The AADs possess a unique mode of action: the nAChR subunits involved in AAD action are different from those targeted by imidazothiazoles [34],[35] and there is no cross-resistance between the 2 chemical classes [12].
Two independent AAD-mutant H. contortus lines were used to screen for mutations in ACR genes of the DEG-3 subfamily. Two genes were found to be affected: The H. contortus des-2 homologue Hco-des-2H, where all AAD-mutant H. contortus carried an insertion in the 5′ UTR introducing 2 additional, out-of-frame start codons, and the gene monepantel-1 (Hco-mptl-1), for which a panel of different mutations were detected in AAD-mutant (AADM) H. contortus. Apart from 1 nonsense mutation discovered in Hc-Howick AADM nematodes (Hco-MPTL-1-m5; Figure 4), the detected mutations all involved mis-splicing resulting in loss of exon(s) from the mRNA as indicated by shortened reverse transcriptase-PCR products (Figure 3). This unusual mechanism has not been described before in H. contortus. In the genetic screen performed on AAD-resistant C. elegans [12], 2 mutants bearing a G-to-A transition of the conserved G nucleotide in the 3′ splice acceptor sites of either the second or third introns were described; these mutations are predicted to cause an increase in the size of the mRNA due to the lack of splicing of the affected intron. In another study [36], a single base pair change in the first intron of the lev-8 subunit gene was identified in a partially levamisole-resistant C. elegans mutant. This mutation leads to alternative splicing and introduction of a premature stop codon. In the case of mutations Hco-MPTL-1-m2 (loss of exon 4), Hco-MPTL-1-m3 (loss of exon 15) or Hco-MPTL-1-m6 (loss of exon 8), exon skipping creates a frame-shift that leads to a premature stop codon (Figure 4). These mutations, including the Hco-MPTL-1-m5 (stop codon) are predicted to result in a truncated, non-functional Hco-MPTL-1 protein and/or, if the mutant mRNA is recognized by the nonsense-mediated mRNA decay (NMD) machinery [37], degradation of the mRNA (some known genes of the NMD machinery in C. elegans have orthologues in the H. contortus genome; Rufener and Mäser, unpublished). Measuring the expression levels of the 3 DEG-3 subfamily genes Hco-mptl-1, Hco-des-2H and Hco-deg-3H in adult H. contortus, we found statistically significant differences in the steady state level of mRNA in AAD mutant worms. In the Hc-CRA AADM isolate, a significant increase of the Hco-deg-3H transcript was observed. A possible explanation may be that this compensates for the loss of the Hco-MPTL-1 subunit since no more full-length Hco-mptl-1 transcript was detectable in Hc-CRA AADM. In the case of Hc-Howick AADM, all 3 nAchR genes were down-regulated compared to Hc-Howick. Although we cannot give a result-based explanation, we interpret the finding that the expression of DEG-3 subfamily nAChR genes is affected in H. contortus as further evidence for the involvement of these genes in AAD susceptibility.
The mutations Hco-MPTL-1-m1 (loss of exon 2 and 3) and Hco-MPTL-1-m4 (partial loss of exon 4 and 15) did not cause a frame-shift, but the loss of the signal peptide and the first 39 amino acids of the extracellular loop for the first mutation, and a truncated protein for the second mutation. Interestingly, 1 of the previously identified AAD-resistant C. elegans mutants also carried a mutation in the signal peptide of the Cel-ACR-23 protein [12]. Receptors are assembled in the endoplasmic reticulum (ER) [38] and interference with the signal peptide could result in mis-localization of the protein or in improper interactions with ER-resident, ACR-specific chaperones [25], [39]–[41]. Furthermore, it is known that the expression, assembly and transport to the surface of ACR subunits is subject to stringent quality controls that guarantee the functional competence of the final product [42]–[44]. Truncated nAChR proteins are likely to be targeted to the lysozyme and degraded.
In summary, we have detected a large number of different mutations affecting the Hco-mptl-1 gene and transcript, respectively, from AAD mutant H. contortus (Table 2). For the benzimidazoles, a variety of different mutations in the target protein ß-tubulin are associated with drug resistance, 3 so far from H. contortus [15],[45],[46] and many more from phytopathogenic fungi [47]. These are point mutations, that are thought to interfere with benzimidazole binding while preserving microtubular function. The mutations have less drastic effects on the predicted protein than those described here for Hco-mptl-1 of H. contortus. At present, we do not know whether Hco-mptl-1 is an essential gene in H. contortus, but our findings indicate that it may not be. There were no mutations in common between H. contortus CRA-AADM and Howick-AADM, indicating that the genetic screen was not saturated. However, for Hco-des-2H, an insertion of 135 bp creating 2 additional start codons was present in the 5′ UTR from both AADM isolates. While Hco-des-2H mRNA levels were significantly lower in Hc-Howick AADM (compared to Hc-Howick), no effect was observed on Hco-des-2H expression in Hc-CRA AADM. It is interesting to note that in C. elegans, mutant worms lacking a functional DES-2 did not exhibit any AAD resistance [12]. The in vitro protocol used to breed AAD-mutant H. contortus is very focused using a large number of individuals and a border line subcurative exposure concentrations over extended time period. This protocol is different from the situation in the field, e.g. after multiple generations exposed to subcurative treatment in sheep, we have so far not been able to obtain AAD-resistant H. contortus (Pradervand and Kaminsky, unpublished data).
In conclusion, several independent mutations were found in the Hco-mptl-1 gene from H. contortus mutants with reduced sensitivity for monepantel, implicating Hco-MPTL-1 as a likely target for AAD action against H. contortus. However, this hypothesis is difficult to test since H. contortus is not readily amenable to genetic manipulation [48]. The AADs are very well tolerated by sheep or cattle [14]. The absence of DEG-3 subfamily acetylcholine receptors in mammals might explain the selective toxicity of AADs to nematodes.
|
10.1371/journal.ppat.1003341 | Differences in Gastric Carcinoma Microenvironment Stratify According to EBV Infection Intensity: Implications for Possible Immune Adjuvant Therapy | Epstein-Barr virus (EBV) is associated with roughly 10% of gastric carcinomas worldwide (EBVaGC). Although previous investigations provide a strong link between EBV and gastric carcinomas, these studies were performed using selected EBV gene probes. Using a cohort of gastric carcinoma RNA-seq data sets from The Cancer Genome Atlas (TCGA), we performed a quantitative and global assessment of EBV gene expression in gastric carcinomas and assessed EBV associated cellular pathway alterations. EBV transcripts were detected in 17% of samples but these samples varied significantly in EBV coverage depth. In four samples with the highest EBV coverage (hiEBVaGC – high EBV associated gastric carcinoma), transcripts from the BamHI A region comprised the majority of EBV reads. Expression of LMP2, and to a lesser extent, LMP1 were also observed as was evidence of abortive lytic replication. Analysis of cellular gene expression indicated significant immune cell infiltration and a predominant IFNG response in samples expressing high levels of EBV transcripts relative to samples expressing low or no EBV transcripts. Despite the apparent immune cell infiltration, high levels of the cytotoxic T-cell (CTL) and natural killer (NK) cell inhibitor, IDO1, was observed in the hiEBVaGCs samples suggesting an active tolerance inducing pathway in this subgroup. These results were confirmed in a separate cohort of 21 Vietnamese gastric carcinoma samples using qRT-PCR and on tissue samples using in situ hybridization and immunohistochemistry. Lastly, a panel of tumor suppressors and candidate oncogenes were expressed at lower levels in hiEBVaGC versus EBV-low and EBV-negative gastric cancers suggesting the direct regulation of tumor pathways by EBV.
| Epstein-Barr virus (EBV) is detected in roughly 10% of gastric carcinoma (GC) cases worldwide. Despite a strong link between EBV and gastric carcinoma, the contribution of EBV to the tumor environment in EBV associated gastric carcinoma is unclear. We performed a global assessment of EBV and host cell gene expression in gastric carcinoma tumors from 71 patients to link EBV genes (and expression intensities) to cell and microenvironmental changes. In addition to the finding that EBV is associated with down-regulated tumor regulatory genes, this study revealed that samples with high levels of EBV gene expression (hiEBVaGCs) displayed elevated immune cell infiltration with high interferon-gamma (IFNG) expression compared to samples with low or no EBV gene expression. Despite this evidence of increased immune posturing, hiEBVaGC samples also showed elevated expression of the potent immune cell inhibitor, IDO1. This finding may partly explain the persistence of these virus associated tumors in the face of local immune cell concentration. Importantly, the small molecule IDO inhibitor, 1MT (1-methyl Tryptophan), has been shown to reverse the tolerance inducing effects of IDO1 in other tumors. We propose that stratification of gastric carcinomas into EBV-negative, EBV-low and EBV-high may provide indicator value for the use of IDO1 inhibitors as adjuvant therapies against hiEBVaGCs.
| Epstein-Barr virus (EBV) is a herpes virus that infects most humans by adulthood. EBV is associated with several human malignancies, including malignancies of epithelial origin. The first report showing EBV's association with lymphoepithelioma-like carcinomas of the stomach was in 1990 by Burke and colleagues using polymerase chain reaction (PCR) [1]. Since that time, several studies have investigated the association between EBV and gastric carcinomas using a variety of methods (PCR, Southern blotting, and in situ hybridization (ISH)). In 1992, Shibata and Weiss reported EBV infection in 16% of gastric adenocarcinomas using PCR primers to the EBNA 1 gene and by ISH using probes against the EBV encoded small RNAs, EBERs [2]. Another report from Japan detected EBV in 6.9% of gastric carcinoma cases using EBER ISH [3]. Attributed to regional/country differences, the highest incidence of EBV-associated gastric carcinoma (EBVaGC) (16%) has been reported from the United States [2] while the lowest incidence (1.3%) is from Papua New Guinea [4]. Despite these landmark studies showing the association between gastric carcinomas and EBV, the mechanisms of EBV pathogenesis in gastric carcinoma are unclear.
Previous studies have shown the sensitivity of high throughout sequencing for detecting infectious agents [5], [6], [7] and for the new discovery of exogenous agents associating with human cancer [6], [8]. Merkel cell virus has been linked to Merkel carcinoma [8] and Fusobacterium has recently been associated with colorectal carcinoma [6]. In line with other reported methods for investigating pathogen associations in human cancers, we have previously developed a computational pipeline for the identification of exogenous sequences in RNA-seq data called PARSES [9]. Using PARSES, we examined two B-cell lines, Akata and JY, which are commonly used as model systems for EBV studies. Analysis of these cell lines revealed the presence of EBV in both cell lines as expected, but it also revealed the presence of the murine leukemia virus, MuLV in the JY but not Akata cell lines [7].
We have improved PARSES to include the utilization of parallel computing either on a local cluster or large-scale clusters, and we have included features that allow the user to simultaneously analyze the human cellular genes in addition to pathogen discovery (recently coined as ‘dual RNA-seq’ by Westermann and colleagues [10]). Here we utilized this pipeline, RNA CoMPASS (RNA comprehensive multi-processor analysis system for sequencing, Xu et al., unpublished), for the detection of viral pathogens in clinical tumor samples by analyzing a cohort of gastric carcinomas generated by the Cancer Genome Atlas initiative (SRA035410). EBV was detected in 12 out of 71 gastric carcinoma samples and the depth of coverage was sufficient to assess overall transcriptome structure in four cases. To our knowledge, this is the first study to globally assess both the EBV and host transcriptomes in gastric carcinomas using RNA-seq (although a recent paper has shed light onto this EBV specific host cell changes using a real time RT-PCR approach [11]). Our analysis led to insights into viral-host interactions and mechanisms through which EBV alters its local tumor environment. Further, this analysis revealed significant differences in the degree of host responses depending on the level of EBV gene expression. This raises the idea that the magnitude may be a more important clinical indicator than the simple detection of EBV in the selection of therapeutic regimens and the prediction of therapeutic responses in gastric carcinomas.
RNA-seq data from The Cancer Genome Atlas (TCGA) gastric adenocarcinoma cohort (SRA035410) was first analyzed using RNA CoMPASS (Figure S1 and Xu et al., unpublished) to assess the virome for each of the 71 data sets. This initial screening was performed using a single lane of sequencing data from each patient. Most samples contained relatively low numbers of reads matching non-human viral sources (e.g. enterobacteria phage T4T) that possibly represent environmental contamination (Figure 1A–B). Of the known human viruses detected, one sample (BR-4298, Figure 1A) contained 6 reads attributed to Hepatitis C virus. Further inspection of these reads showed high homology to the human immunoglobulin light chain variable region (Table S1). These reads likely represent human sequences rather than reads derived from Hepatitis C virus. Twelve samples showed evidence of human cytomegalovirus (HCMV) with read numbers ranging from 5 to 132. Individual BLASTing of selected HCMV reads showed high homology to HCMV genomes but not to human sequences indicating that these are bona fide HCMV derived reads. The relatively low numbers of HCMV reads in these samples (relative to the numbers of EBV in some samples, see below) suggests that these reads are derived from a low number of HCMV infected cells or that the virus is not expressing substantial numbers of polyadenylated RNAs in these tumor samples.
EBV was detected in 12 out of the 71 (17%) gastric carcinoma cases with varying levels of reads. To further analyze the EBV-associated gastric carcinoma (EBVaGC) samples, the two lanes of sequence per sample were combined to attain greater sequencing depth. These sequence files were aligned against a modified EBV B95-8 genome that contains Raji genome sequences inserted into a deleted region of the B95-8 genome (Genbank accession number AJ507799) plus the hg19 assembly of the human genome. Alignments were carried out using Novoalign V2.07.18 [-o SAM, paired-end, default options]. Based on the assembly to the human genome, sample quality and throughput was found to be consistent across all samples with the numbers of human mapped reads ranging from 128 to 159 million. Eight of the 12 EBV positive samples were found to have less than 200 reads per sample (inset of Figure 1C), three were found to contain more than 30,000 reads and one sample was found to contain 1,194 reads (Figure 1C). We tentatively considered the 8 cases with less than 200 reads to represent nominal infections similar to that observed with CMV (above). The 4 samples with higher read numbers, BR-4253, BR-4271, BR-4376, and BR-4298, were taken for more in depth transcriptome analysis.
Notably, while three of the four EBV positive samples with high numbers of EBV reads were classified as the more common Type I strain of EBV, one of these samples, BR-4253, was classified as the type II strain (Figure 1A). Since the strain defining regions of EBV, EBNA2 and EBNA3A/3B/3C [12] are largely not expressed in EBVaGC, we were concerned that the reads from sample BR-4253 could be misclassified as type II. We analyzed a few of the reads defined by MEGAN as type II from sample BR-4253 using manual BLAST and the majority of reads aligned to both type I (B95-8/Raji) and type II strains (AG876 (Genbank accession number DQ279927)) with some of these showing better homology to AG876 (data not shown). Despite this, the small number of reads derived from the EBNA2 and EBNA3A/3B/3C loci were more homologous to the type I than the type II strain. Therefore, this sample was likely misclassified as the type II strain because of greater similarity to the AG876 genome at highly expressed regions outside of the EBNA2 and EBNA3A/3B/3C loci.
EBV transcript quantification and genome coverage information was generated for samples, BR-4253, BR-4271, BR-4376, and BR-4298 using the transcriptome analysis software, SAMMate (note that the sequencing libraries were generated from polyA selected RNA which precludes the sequencing of EBER genes) [13]. Genome coverage information was first visualized by displaying the number of reads across each genomic position in the Circos plot shown in Figure 2A. Because of disparate coverage intensities, the Circos graph in Figure 2A is plotted in log scale to allow simultaneous visualization of the less abundantly expressed regions of the genome (expandable non-log and log plots are provided in Figures S3 and S4). Notably, coverage across the BamHI A region was high relative to other parts of the genome with greater than 96% of total reads corresponding to the BamHI A region in each case (Figure 2B).
Evidence for transcription of the essential episomal replication factor, EBNA1 is observed in samples BR-4253, BR-4271 and BR-4376 (Figure 2A (upper left region of the figure) and Figure 2B). No EBNA1 reads were detected in BR-4298 most likely owing to the significantly lower read numbers in this sample (Figures 2A–B).
Evidence for transcription of the immediate early genes, BZLF1 and BRLF1, is similarly seen in BR-4253, BR-4271 and BR-4376 but not in BR-4298 (again, possibly due to the low overall read numbers). Despite the detection of BZLF1 and BRLF1 reads in samples BR-4253, BR-4271 and BR-4376, there is a remarkable absence of reads for most other downstream lytic genes in these samples. In Figure 3, we plotted the ratio of lytic gene transcripts (sans lytic genes in the BamHI A region) relative to the level of BZLF1 RNAs in BR-4253, BR-4271 and BR-4376 and compared this to the corresponding relative levels of these gene transcripts in reactivating Akata cells [14]. This comparison indicates that while the BZLF1 and BRLF1 immediate early genes are expressed in these tumors, there is a clear lack of lytic cycle progression; reflecting abortive lytic replication in this in vivo setting.
Consistent with previous reports of LMP2 expression in gastric carcinomas [15], [16], we similarly see evidence of LMP2 transcription in samples BR-4253, BR-4271 and BR-4376 (Figures 2A–B and Figure 4A). LMP1 has been previously reported to be expressed at low levels or to be not expressed in gastric carcinomas [17], [18], [19]. We similarly find low albeit detectable levels of LMP1 in BR-4253, BR-4271 and BR-4376 (Figures 2B and 4A). Strikingly, however, sample BR-4253 has a very high number of reads corresponding to the early BNLF2A/B locus, which overlaps the LMP1 3′ untranslated region (Figures 2B, 3 and 4A). No BNLF2A/B reads are detected in BR-4271, BR-4376, and BR-4298 (Figure 2B) suggesting that this is unique to BR-4253. The high expression level of the early BNLF2A/B genes in BR-4253 is surprising because it occurs in the absence of most other early genes. This suggests the possibility that BNLF2A/B is expressed in this patient through an alternative mechanism possibly mediated through a viral genetic alteration.
The most actively polyA transcribed region of the EBV genome, the BamHI A region (Figures 2A–B), shows excellent coverage across most of the RPMS1/A73 exons with apparent additional coverage observed for the regions spanning the leftward transcribed genes, BALF5, BALF3, and BALF4 (Figure 4B). Coverage across these leftward genes is unexpected because they are thought to be lytic genes and not expressed during latency. We therefore performed directional sequencing of a naturally occurring EBV positive gastric carcinoma cell line, SNU-719, to allow us to determine the orientation of transcripts across this region. EBV read coverage for SNU-719 was remarkably similar to that observed for the tumor specimens (Figure 4B). Outside of a small blip of leftward transcription noted near the RPMS1 exon 1b, there is little leftward transcription across this region. This indicates that the transcription observed across this region in the tumor specimens are likely rightward oriented and to a large extent related to RPMS1 and/or A73 but not BALF5, BALF3, BALF4, BILF1, LF1, or LF2.
Also notable in Figure 4B is rightward coverage across the introns between exons 4 and 5 and exons 6 and 7 of the RPMS1 gene (boxed regions in SNU-719 tracks). This coverage likely does not represent intron fragments generated after transcript splicing because this coverage is observed in sequencing libraries generated from polyA selected RNA (upper SNU-719 tracks). In contrast, there is no coverage of the first 4 RPMS1 introns on the polyA track whereas there is substantial coverage across these regions when ribo-depleted RNA was used for sequencing (Figure 4B). Therefore, the rightward coverage between exons 4 and 5 and between exons 6 and 7 likely represent bona fide previously unannotated rightward exons/transcripts. The read coverage between exons 6 and 7 may arise from mature RPMS1 isoforms that retain this intron (forming a unique RPMS1 isoform). The coverage between exons 4 and 5 starts near the middle of this intron suggesting that this is a site of transcription initiation or a that it is a splice acceptor site. Since splice mapping (see below) did not identify candidate splicing events near the beginning of this intron coverage, it is possible that this coverage arises from transcription initiation from an unknown upstream promoter.
As mentioned above, more than 96% of all EBV reads align to the BamHI A region. Further, RPMS1 exon coverage ranks within the top seven percent of expressed cellular genes in samples BR-4253, BR-4271, and BR-4376 with expression that is more than five times the median cellular gene expression level (Figure 5). We conclude that not only is expression of this region high relative to other EBV encoded genes, but the expression is also high relative to cellular genes. In contrast, it is notable that the LF3 gene which is within the BamHI A locus and which has been found to be expressed at very high levels in other systems [20], shows no evidence of expression in these in vivo gastric carcinoma tumor datasets.
To assess splicing events in this region, alignments were performed using the junction mapper, TopHat [21]. Consideration of the most abundant splice junction reads indicates the predominance of sequential splicing from exons 1-2-3-4-5-6-7 (Figure 6). Nevertheless, there is significant evidence of intra-exonal splicing at exons 3 (3a to 3b), 5 (5a to 5b), and 7 (7a to 7b) (Figure 6). Although splicing from exons 1 to 2 is the most predominant 5′ region splicing order, there is also good evidence of alternative splicing to exon 1a (i.e. splicing of exon 1 to exon 1a to exon 2) (Figure 6). In samples BR-4253 (Figure 6) and SNU-719 (data not shown), we also noted evidence of splicing initiating from the middle of the newly identified coverage in the intron between exons 4 to 5 to the start of exon 5. This indicates additional complexity in this new region whereby some of these transcripts splice to exon 5 while some read through to exon 5.
EBV likely contributes to gastric carcinoma through the subversion of at least some of the oncogenic pathways required for the development of gastric carcinoma. However, the way that EBV subverts these pathways is likely distinct from the mechanism of pathway disruption in the absence of EBV (e.g. through genetic alterations). Since cellular gene expression is typically responsive to altered signaling mechanisms, differences in gene expression profiles can be used to not only classify cell populations but also infer upstream signaling events within certain cell populations.
To investigate influences of EBV dependent alterations in tumor signaling pathways, we analyzed global cellular gene expression in all 12 EBV positive specimens plus an additional 20 randomly selected EBV negative samples. EBV gene expression data was not included in this analysis to ensure that clustering occurred based only on differences in cellular gene expression (i.e. that it occurred independently of biases incurred by the presence of EBV gene expression signatures). Strikingly, when the set of samples were analyzed using hierarchical clustering, the four gastric carcinoma samples with higher numbers of EBV reads (BR-4253, BR-4271, BR-4376, and BR-4298) formed its own well-separated group (Figure 7A). One of the EBV negative samples, BR-4294, clustered independently of the others and subsequent analysis revealed that this sample was likely an outlier (Figure S5). Nevertheless, this sample was retained in the subsequent differential expression analysis as a conservative measure.
Human transcript counts from the EBVaGCs with high EBV read levels were compared to the EBVaGCs with low EBV read numbers and with the EBVnGCs. Using this approach, 490 genes were found to have statistically significant differential expression in the “high” EBVaGC (hiEBVaGC) samples relative to both EBVnGC and “low” EBVaGCs (loEBVaGC) samples (Figure 7A–B and Table S2). These genes separated into five distinct clusters with clusters 1, 3, and 5 showing genes that were predominately expressed at higher levels in hiEBVaGCs and clusters 2 and 4 containing genes that were predominantly expressed at lower levels in hiEBVaGCs (Figure 7A). We also performed an additional clustering analysis using only the EBV genes across the 12 EBVaGC. This analysis revealed that the 4 hiEBVaGC samples cluster distinctly from the other EBVaGC samples (Figure S6). This apparently distinct gene expression pattern observed in the 4 hiEBVaGC samples raises the possibility that these samples represent infection of a unique cell type relative to the other samples (possibly tumor cells versus stroma or B-cells).
Ingenuity Pathway Analysis software (IPA: Ingenuity Systems) was used to assist the analysis of pathways and known molecular functions associated with differentially expressed genes. Twenty four percent (116) of the 490 genes with statistically significant differential expression were found to be immunologically related genes (Figure 8A). The vast majority of these genes were expressed at higher levels in hiEBVaGCs with IDO1 and IFNG ranking among the top (38-fold and 16-fold, high v. negative). The differentiation and other cell surface marker profiles are consistent with the presence of cytotoxic T-cells (CTLs) and/or natural killer (NK) cells in hiEBVaGC. Further, CTLs and NK cells are key producers of granzymes and perforin, which are found to be elevated in the hiEBVaGC (Figure 8A).
The interferon gamma (IFNG) pathway was analyzed further using IPA to determine the extent of IFNG pathway involvement in hiEBVaGC. We observed marked involvement of the IFNG pathway with 156 of the 490 differentially expressed genes associated with the IFNG pathway, the majority of which were elevated (Figure 8B).
The analysis of IDO1 levels for each of the 32 gastric carcinomas showed that the samples with the highest number of EBV reads had the highest levels of IDO1 expression (Figure 9A). To further explore the link between EBV and IDO1, we analyzed a separate cohort of Vietnamese gastric carcinoma samples by real time RT-PCR. RPMS1 was detected in two of these samples (CZRDPREA and WZQ1TALM) (Figure 9B) and these samples ranked among the highest for expression of IDO1 (27 and 17 fold relative to the average of the 5 normal adjacent tissue samples). Further, in these samples, normal adjacent tissue showed lower RPMS1 expression and lower IDO1 expression compared to their tumor counterparts. Notably, one of the EBV negative samples, W31AB410, showed the highest level of IDO1 (43 fold). Nevertheless, this sample was notable in that like the two EBV positive samples, the pathology report for this sample similarly noted high levels of immune cell infiltration which may result from the presence of another infectious agent.
In Situ Hybridization for EBER was performed on a gastric carcinoma tissue array (ST2091; US Biomax) in order to assess the presence of EBV. In the strongly EBV positive cases, EBV was detected in the epithelial cells (e.g. F8 in Figure 9C). A high level of immune cell infiltration is observed in EBV positive (e.g. F8, Figure 9C) but not the tumor grade matched EBV negative sample, A15 (Figure 9C) with a high proportion of the immune cells in F8 showing intense IDO1 staining.
Analysis of the 178 down regulated genes showed that 19 tumor suppressor genes and 13 candidate oncogenes were found to be expressed at lower levels in hiEBVaGC (Table S3). Furthermore, we observed several inhibitors of the hedgehog and Wnt pathways to be expressed at lower levels in hiEBVaGCs suggesting additional components to the complex interactions involved in EBVaGC pathogenesis.
Consistent with the Shibata and Weiss study for the incidence of EBVaGC in the United States using ISH against EBERs [2], we detected EBV in 12 of the 71 (17%) gastric carcinoma samples from The Cancer Genome Atlas (TCGA) cohort using RNA CoMPASS. The detection of EBV using EBER ISH is widely used and the similar detection levels between the Shibata and Weiss study [2] and our work suggest that both methods are accurate for determining the presence of EBV in biological specimens. Importantly, however, the use of RNA-seq data allowed us to also infer the magnitude of local environmental signaling influences for different levels of EBV infection/viral gene expression. While the four samples with higher levels of EBV transcripts formed a clearly distinct cellular gene expression cluster, the eight samples with low numbers of EBV reads clustered in a mixed fashion among the EBV negative specimens. We propose that these two classes of EBV infection should be considered functionally distinct with possible implications in therapeutic intervention decisions and/or therapeutic response predictions.
RNA CoMPASS has the potential to simultaneously allow for the investigation of all pathogens present in tumor samples. In addition to EBV, we detected low levels of enterobacteria phage T4T, HCMV, Hepatitis C virus, and Helicobacter pylori (data not shown). The detection of enterobacteria phage T4T and Hepatitis C virus transcripts should be met with caution due to the likely possibility of environmental contamination and misclassification of these reads, respectively. While the HCMV reads likely represent true HCMV infection of cells within the tumor sample, the low read levels suggest either low numbers of HCMV infected cells or limited expression of polyadenylated viral RNAs. Finally, we detected H. pylori in three of the gastric carcinoma samples but the number of reads was very low in each case. Since bacterial RNAs are typically not polyadenylated or have limited numbers of polyadenylated RNAs [22], [23], [24], [25], this low detection level probably results from the sequencing libraries being prepared from polyA selected RNA rather than an absence of H. pylori in these samples.
Of the 12 EBVaGCs, there was sufficient EBV read coverage in four of the samples to carry out more detailed transcriptomic analysis. LMP2, EBNA1, and LMP1 expression was detected in three of the EBVaGCs and these results are generally consistent with the findings of other groups [15], [16], [17], [18]. The magnitude of expression from the BamHI A region relative to the transcription levels of other EBV genes is striking, however. This result is consistent with a previous report using a naturally infected EBV positive gastric carcinoma cell line [26]. Nevertheless, our analysis makes this observation in the natural in vivo setting of the tumor, and the use of RNA-seq facilitated the evaluation of transcript structures and the magnitudes of BamHI A region gene expression relative to other viral and cellular genes.
Although others have been unable to detect protein from naturally expressed BamHI A rightward transcripts [27], [28], the high expression level of these transcripts in hiEBVaGC samples suggests a functional role in gastric adenocarcinomas; possibly as long non-coding RNAs (lncRNA). These rightward BamHI A transcripts also encode as many as 44 intronic microRNAs (miRNAs) [29], [30]. The function of the BART miRNAs in the EBV life cycle and in EBV associated malignancies is currently unclear but a recent study by Raab-Traub's group provided evidence that the BART miRNAs contribute to the tumor phenotype in EBVaGC [31]. In Raab-Traub's study, several lines of evidence supported this contention. First, very little EBV latent protein expression was detected and inhibition of the small amount of LMP1 expressed did not affect the cell's phenotype. Second, they observed that the majority of the significant cellular gene expression changes following infection of AGS (a gastric carcinoma cell line) cells with EBV were down regulated, many of which were significantly enriched in both experimentally and bioinformatically predicted BART miRNA targets [31], [32]. Based on this evidence and the fact that the BamHI A rightward transcripts are expressed at high levels in gastric carcinomas, it seems likely that the BART miRNAs play an important role in modulating the cellular phenotype in this tumor type. Nevertheless, many lncRNAs are involved in repressive complexes raising the possibility that the high levels of spliced rightward BamHI A transcripts that we detect in vivo may function as lncRNAs which similarly contribute to repression of cellular gene expression in hiEBVaGCs. Our strand specific RNA-seq analysis of SNU-719 cells further support our contention of high level expression of the rightward RPMS1 and A73 related transcripts in gastric carcinomas. This analysis also demonstrated the presence of additional rightward exons/genes within this region that may similarly play a role in lncRNA mediated regulation of viral and/or cellular signaling.
Although EBV primarily exhibits latent gene expression patterns in EBV associated tumors, recent studies using EBV associated lymphoma models suggest that a small portion of tumor cells express lytic transcripts that promote tumor growth [33], [34], [35], [36]. The Kenney lab showed that B cells harboring an EBV BZLF1 knock out mutant grew slower than wild type infected cells in a SCID mouse xenograft model [33]. In a separate study, they showed that a mutant EBV over expressing BZLF1 induces lymphomas with abortive lytic EBV infection in a humanized mouse model [36]. By assessing global EBV gene expression, we provide evidence for an abortive lytic phase in vivo; in the context of the natural setting of a human tumor. This supports the lymphoma animal studies from the Kenney group and raises the possibility that an abortive lytic phase may also play a role in EBV associated epithelial tumors.
One EBVaGC sample (BR-4253) was found to express high levels of BNLF2A/B. In the absence of significant expression of other lytic genes, the detection of BNLF2A/B expression in this sample was unexpected. One of the simplest models to explain this observation is a possible viral genetic alteration that juxtaposes this gene with an active viral promoter; in a manner reminiscent of the previously identified hetDNA (BZLF1 gene recombined to an active latency promoter) [37], [38], [39]. Alternatively, this could result from a rare viral integration event positioning the BNLF2A/B gene downstream from an active cellular promoter. Just as advantageous genetic alterations evolve in the cellular genome during cancer progression, a genetic event that resulted in the activation of BNLF2A/B may be an example of an advantageous viral genetic alteration that was selected during tumor evolution. BNLF2A was shown previously to function as an immune evasion protein through HLA class I down regulation (via blocking of TAP activity) [40]. This anti-immune function may have been selected for during tumor evolution and may support viral/tumor survival in this patient.
Cellular RNA expression profiling provided strong evidence for immune cell infiltration in hiEBVaGCs. This can be seen in tissue sections from EBV positive specimens (e.g. see Figure 9C) and is further supported by the pathology reports from the two EBV positive gastric carcinoma samples from the Vietnamese cohort which indicate high levels of immune cells (Table S4). This observation is in line with previous studies using standard hematoxylin and eosin staining of tumor sections [3], [41] where lymphocyte infiltration was found to be predominately CD8+ T cells [42], [43]. Notably, however, despite this apparent robust immune response in hiEBVaGC, EBV and the infected tumor cells are able to persist in these patients. This suggests that these tumors may have compensatory immune evasion strategies that allow virus/tumor survival in this setting [44]. First, the limited expression of viral protein coding genes in EBVaGC likely contributes to the avoidance of viral antigen targeting [45]. Second, although the EBV encoded protein, EBNA1 is required for viral episomal replication/maintenance and therefore must be expressed in proliferating cells, it encodes a glycine-alanine repeat domain that blocks its proteasomal processing for CTL presentation [46], [47]. Third, here we found that expression of the interferon-gamma (IFNG) inducible CTL and NK inhibitor, indoleamine 2,3-dioxygenase (IDO1) is high in hiEBVaGC. IDO1 is a rate-limiting enzyme involved in the catabolism of tryptophan (Trp) [48]. CTLs and NK cells are uniquely sensitive to Trp depletion leading to the induction of stress responses and the inhibition of proliferation and activation [49], [50]. IDO1 functions to cause local tryptophan depletion under physiological and pathogenic immune tolerance settings such as during placentation and cancer [51], [52] where it is considered to be critical for establishing local immune tolerance. Among other candidate effectors, increased IFNG has been shown to induce IDO1 expression [53], [54]. Therefore, despite the apparent increase in CTL and NK cells in hiEBVaGCs, the activated IFNG signaling may counteract this response through IDO1 mediated Trp depletion (Figure 10); allowing tumor survival.
The findings of high IDO1 levels in several cancers and studies showing that IDO1 is critical for tumor survival has led to intense interest in the potential of anti-IDO1 based immunomodulatory therapeutics [55], [56], [57], [58]. IDO1 inhibitors, such as the small molecule inhibitor, 1MT, have shown anti-tumor potential in combination with conventional chemotherapeutic drugs [57], [58]. This raises the important possibility that the therapeutic response for at least the subset of hiEBVaGCs may similarly be enhanced by the addition of IDO1 targeting therapeutics.
In our study, 156 of the genes found to be differentially expressed in EBVaGCs are linked to the IFNG pathway. The EBV encoded small RNAs, EBERs, have been shown to induce the expression of IFNG [59], and they likely play a significant role in the active IFNG response observed here. Despite this, the extensive level of secondary structure guiding the processing of the BamHI A rightward introns during the miRNA processing steps may similarly contribute to the IFNG response in EBVaGCs observed here (Figure 10).
EBVaGCs exhibit extensive nonrandom DNA methylation at the promoter regions of various cancer-related genes [60], [61] and has been classified as having the CpG island methylator phenotype (CIMP) [62]. Several studies have investigated possible mechanisms of promoter hypermethylation of host genes in association with EBV infection. LMP1 mediated activation of DNA methyltransferase 1 (DNMT1), through either activation of c-Jun NH2-terminal kinase (JNK)-activator protein-1 (AP-1) signaling [63] or through RB-E2F pathway activation [64], have been proposed as mechanisms in some systems. However, EBVaGCs do not typically express significant levels of LMP1. A study by Hino et al. demonstrated DNMT1 activation via LMP2A [65] raising the possibility that a LMP2A/DNMT1 mechanism could be involved. Nevertheless, a study by Chong et al. showed that DNMT expression was suppressed in EBVaGC and that the methylation of specific genes occurs through a mechanism independent of DNMT1 activation [66]. Based on this observation and on our findings of relatively low levels of LMP1 and LMP2A expression in EBVaGCs, we propose that methylation/imprinting may be downstream of more direct EBV inhibitory mechanisms. The robust expression levels of the BamHI A transcripts in EBVaGCs put them high on the radar as candidates for this type of regulation, possibly through lncRNA mediated chromatin imprinting based mechanisms.
Multiple tumor suppressors were expressed at lower levels in EBVaGCs including five (TFF2, RBP4, HOXA9, LRRN1, and RAP1GAP) that are known to be hypermethylated in cancers [67], [68], [69], [70], [71]. Another gene expressed at lower levels in EBVaGCs was HNF4A, a cell-specific transcription factor known to regulate a large number of genes in liver, intestine, pancreas, and stomach [72]. Decreased expression of HNF4A has been shown in renal cell carcinoma [73] and has recently been shown to regulate key genes involved in cellular proliferation [72]. A recent study by Lucas and colleagues suggest that HNF4A acts as a tumor suppressor [72].
In addition to tumor suppressors, we also observed several candidate oncogenes to be expressed at lower levels in EBVaGCs including 4 (CDH17, CDX1, ETV4, and PPP1R1B) known to be over expressed in gastric carcinoma and gastrointestinal cancers [74], [75], [76], [77]. Although EBV clearly contributes to cancers, its oncogenic properties are a byproduct of its life cycle rather than an evolved tumor promoting function. In line with this concept, the lower levels of these oncogenes in EBVaGCs may be a byproduct of EBV's life cycle. Conversely, it is possible that the non-EBV mediated gastric carcinoma oncogenic pathway occurs through the up-regulation of these genes whereas the EBV assisted oncogenic path does not. Regardless of which of these principles may explain this observation, the lower levels of oncogenes in EBVaGC may partly explain the more favorable prognosis that is often observed in EBVaGC. Similarly, the lower levels of USP2, a negative regulator of p53, may help explain the normal to elevated levels of p53 found in EBVaGC [78], [79] and possibly the better responses to chemotherapeutics.
An increase in sonic hedgehog (SHH) expression and its activation in gastric carcinoma, especially H. pylori associated gastric carcinomas has been well established [80]. In our study, several inhibitors of both the SHH and Wnt pathways were found to be lower in hiEBVaGC including HHIP (SHH) and SHISA3, NKD2 and LRP4 (Wnt). The decrease in SSH inhibitor, HHIP, [81] suggests that Hedgehog activity may be higher in hiEBVaGC. Down regulation of HHIP in pancreatic cancer has been shown to be mediated through epigenetic CpG hypermethylation within the promoter region [82]. This raises the possibility of a specific methylation process by EBV, since we observe a significantly lower level of HHIP reads in the hiEBVaGC compared to loEBVaGC and EBVnGC. Hypermethylation of the promoter region of NKD2 has been established in malignant astrocytic gliomas [83], and a CpG island within the SHISA3 and LRP4 promoter regions have been identified [84]. This suggests that epigenetic silencing of these Wnt pathway inhibitors may also occur through an EBV mediated mechanism.
All human specimens were de-identified prior to acquisition. Total RNA from 21 Vietnamese gastric carcinoma samples and 5 normal adjacent samples were obtained from Biospecimen Repository at Bioserve (Beltsville, MD). Demographic and clinical data is available in table S4. RNA-seq data from 71 gastric carcinoma samples generated through the National Institutes of Health, The Cancer Genome Atlas (TCGA) project were obtained from the NCBI Sequence Read Archive (SRA035410, now available through the Cancer Genomics Hub managed by the University of California, Santa Cruz (UCSC)). Demographic and clinical data for each sample is available through the TCGA data portal (http://cancergenome.nih.gov/). Briefly, samples were obtained from non-Hispanic White Russians with no previous treatment. The mean age was 69 years with a range of 43 to 90 years. Total RNA was isolated from each sample using mirVana RNA kit according to TCGA. High quality RNA was polyA selected and sequenced using an Illumina Genome Analyzer II machine running paired end 51 base sequencing reactions with two lanes of sequence per sample.
SNU-719 gastric adenocarcinoma cells were obtained from the Korean Cell Line Bank. They were grown in RPMI 1640 (Thermo Scientific; Waltham, MA) plus 10% fetal bovine serum (Invitrogen-Gibco; Grand Island, NY) with 1% penicillin-streptomycin (Invitrogen-Gibco; Grand Island, NY). Cells were grown at 37°C in a humidified, 5% CO2 incubator.
Total RNA was extracted from SNU-719 cells using the miRNeasy Mini Kit (Qiagen, Hilden, Germany) according to manufacturer's instructions. Two separate cDNA libraries were prepared from polyA selected and from Ribo-Zero selected RNAs using the Illumina Truseq Stranded Total RNA Sample Prep Kit (RS-122-2101). 101-base paired-end sequencing was performed using an Illumina HiSeq 2000 instrument. The SNU-719 RNA-seq data used in this publication have been deposited in NCBI's Gene Expression Omnibus [85] and are accessible through GEO Series accession number GSE45453 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE45453).
RNA CoMPASS (Xu et al., unpublished) a graphical user interface (GUI) based parallel computation pipeline, RNA comprehensive multi-processor analysis system for sequencing (RNA CoMPASS) for the analysis of both exogenous and human sequences from RNA-seq data (Figure S1). Briefly, for the analysis of both exogenous and human sequences, raw sequence data is first processed through an in house de-duplication algorithm. Following de-duplication, reads are aligned to a reference genome containing human (hg19; UCSC) and abundant sequences (which include sequence adapters, mitochondrial, ribosomal, enterobacteria phage phiX174, poly-A, and poly-C sequences). Novoalign V2.07.18 (www.novocraft.com) [-o SAM, default options] is used to map reads to the reference genome and to eliminate low-quality reads (QC<20). TopHat V1.4.0 [default options] [21] is used to identify and isolate all sequences that map to human splice junctions. The results from these programs are compiled and separated into mapped reads (used for human transcriptome analysis) and unmapped reads (used for exogenous sequence analysis). Human mapped reads are analyzed using SAMMate [13] to quantify gene expression and to generate genome coverage information. Unmapped reads are subjected to consecutive BLAST V2.2.24 searches against the Human RefSeq RNA database (an additional “pre-clearing” step) and then to the NCBI NT database to identify reads corresponding to known exogenous organisms [86]. Results from the NT BLAST searches are filtered to eliminate matches with an E-value of less than 10e−6. The results are fed into MEGAN 4 [87] for convenient visualization and taxonomic classification of BLAST search results.
RNA CoMPASS is designed to take advantage of parallel processing at several key steps to speed processing times. In our case, we used a four node, 12 core, Intel Xeon Mac Pro (64GB of memory per node) cluster.
Samples containing evidence of EBV were identified using RNA CoMPASS. Since each sample contained sequence data from two runs, data from both runs were combined in order to generate a greater sequencing depth for transcript quantification. In addition, 20 EBV negative samples were randomly chosen for analysis. Samples were aligned to a reference genome containing human (hg19) and a modified EBV B95-8 genome that contains Raji genome sequences inserted into a deleted region of the B95-8 genome (Genbank accession number AJ507799) using Novoalign V2.07.18 (www.novocraft.com) [-o SAM, default options]. Transcript data from Novoalign was analyzed using SAMMate for transcript quantification of human and EBV genes and to generate coverage (wiggle) files for visualization of read distributions. Splice junction data was generated using the junction aligner, TopHat V1.4.0 [default options]. Coverage data was visualized using the Integrative Genomics Viewer (IGV) [88].
Total RNA was reverse-transcribed using the SuperScript III First-Strand Synthesis System for RT-PCR (Invitrogen, Carlsbad, CA). Random hexamers were used along with 250 ng RNA in a 20 µl reaction volume according to manufacturer's instructions. For the incubation steps (25°C for 10 min followed by 50°C for 50 min) a Mastercycler ep (Eppendorf, Hamburg, Germany) was used. For real-time PCR, 1 µl of the resulting cDNA was used in a 10 µl reaction mixture that included 5 µl of 10× SsoFast EvaGreen supermix (Bio-Rad, Hercules, CA), 1 µl of 10 µM forward and reverse primer mix (Integrated DNA Technologies, Coralville, IA), and 3 µl of PCR grade water. The IDO1 primers amplified a 112 base pair product. Forward primer 5′-CAAATCCACGATCATGTGAACC-3′ and reverse primer 5′-AGAACCCTTCATACACCAGAC-3′ were used previously by Prachason et al [89]. The RPMS1 primers amplified a 181 base pair product consisting of exon 6 and exon 7. Forward primer 5′-CCAGGTCAAAGACGTTGGAG-3′ and reverse primer 5′-CACCACGGTGCAGCCTAC-3′ were used. The GAPDH primers amplified a 297 base pair product. Forward primer 5′- CAATGACCCCTTCATTGACC-3′ and reverse primer 5′- GACAAGCTTCCCGTTCTCAG-3′ were used. Each sample was performed in triplicates. No-template controls and no-reverse transcription controls were also included in each PCR run. Thermal cycling was performed on a CFX96 Real Time System (Bio-Rad, Hercules, CA) and data analysis was performed using the CFX Manager 3.0 software. Cycling conditions included an initial incubation at 95°C for 30 seconds followed by 40 cycles consisting of 95°C for 5 seconds, and 60°C for 5 seconds. Melting curve analysis was performed at the end of every qRT-PCR run.
Chromogenic In Situ Hybridization (CISH) was performed by the Tulane Molecular Pathology Lab using the HistoSonda EBER probe kit (American Master Tech, Lodi, CA) according to manufacture's instructions. The tissue array was deparaffinized and rehydrated in a graded solution of Xylene and alcohol. Tissue array was deproteinized using Proteinase K and incubated with Digoxigenin EBER probe. Tissue array was subsequently washed with water and PBS. The tissue array was incubated with Anti-digoxin and anti-mouse horse radish peroxidase to form a duplex with the Digoxigenin EBER probe. For colorimetric staining, slides were then incubated in 3,3′-Diaminobenzidine (DAB; Vector Laboratories), washed with dH2O, counterstained with hematoxylin, and rinsed with PBS (pH 7.4). Slides were dehydrated in a graded solution of alcohol and Xylene and sealed with Permount Mounting Medium (Sigma). To visualize the tissue array, slides were scanned into ScanScope CS2 (Aperio, Vista, CA) and images were acquired with ImageScope (Aperio).
Formalin-fixed, paraffin-embedded (FFPE) gastric tumor tissue array (ST2901) was purchased from U.S. BioMax (Rockville, MD). Demographic and clinical data can be found on the U.S. BioMax website (http://www.biomax.us/tissue-arrays/Stomach/ST2091). The tissue array was deparaffinized, and rehydrated in a graded solution of Sub-X clearing medium (Leica Biosystems, Buffalo Grove, IL). Antigen retrieval was performed with Tris-EDTA Buffer, consisting of 10 mM Tris Base, 1 mM EDTA Solution, and 0.5% Tween 20 (pH 9.0), for 30 minutes. The tissue array was then quenched with 3% H2O2 (Sigma), rinsed with TNT washing buffer made of 0.1 M Tris-HCl, 0.15 M NaCl, and 0.5% Tween-20 (pH 7.5), blocked with blocking reagent purchased from Perkin Elmer (Waltham, MA) and stained with goat-anti-human IDO (Abcam, Cambridge, MA) overnight at 4°C. Tumor sections were subsequently washed in TNT, incubated with donkey-anti-goat HRP conjugated secondary antibody (Santa Cruz, Dallas, TX) for 1 hour at room temperature, and washed with TNT. For colorimetric staining, slides were then incubated in 3,3′-Diaminobenzidine (DAB; Vector Laboratories), washed with dH2O, counterstained with hematoxylin, and rinsed with PBS (pH 7.4). Slides were dehydrated in a graded solution of Sub-X clearing medium and sealed with Permount Mounting Medium (Sigma). To visualize the tissue array, slides were scanned into ScanScope CS2 (Aperio, Vista, CA) and images were acquired with ImageScope (Aperio).
Transcript counts were imported into the R software environment and analyzed using the edgeR package [90]. Genes with low transcript counts (less than 1 CPM (count per million)) in the majority of samples were filtered. The Manhattan (L-1) distance matrix for the samples was computed using the remaining transcript counts, and this was taken as input for hierarchical clustering using the Ward algorithm. The well separated cluster of four EBV positive samples were found to be those with the highest numbers of EBV reads and were classified as “high EBV”. The remaining samples were then classified as “EBV-negative” or “low EBV”. The glmFit function was then used to fit the mean log(CPM) for each group and likelihood ratio tests were used to identify those genes that were differentially expressed in any of the three possible comparisons, with adjusted p<0.05 following the Benjamini-Hochberg correction for multiple testing. The fitted log(CPM) values for the subset of genes that were differentially expressed in the high EBV samples relative to both the low EBV and EBV-negative samples were then clustered using the Euclidean distance and complete linkage algorithm to detect groups of co-expressed genes. The EBV transcript counts from all positive samples were imported in MeV [91] for hierarchical clustering using the Manhattan distance matrix and average linkage clustering algorithm.
|
10.1371/journal.pgen.1003433 | Coordinated Cell Type–Specific Epigenetic Remodeling in Prefrontal Cortex Begins before Birth and Continues into Early Adulthood | Development of prefrontal and other higher-order association cortices is associated with widespread changes in the cortical transcriptome, particularly during the transitions from prenatal to postnatal development, and from early infancy to later stages of childhood and early adulthood. However, the timing and longitudinal trajectories of neuronal gene expression programs during these periods remain unclear in part because of confounding effects of concomitantly occurring shifts in neuron-to-glia ratios. Here, we used cell type–specific chromatin sorting techniques for genome-wide profiling of a histone mark associated with transcriptional regulation—H3 with trimethylated lysine 4 (H3K4me3)—in neuronal chromatin from 31 subjects from the late gestational period to 80 years of age. H3K4me3 landscapes of prefrontal neurons were developmentally regulated at 1,157 loci, including 768 loci that were proximal to transcription start sites. Multiple algorithms consistently revealed that the overwhelming majority and perhaps all of developmentally regulated H3K4me3 peaks were on a unidirectional trajectory defined by either rapid gain or loss of histone methylation during the late prenatal period and the first year after birth, followed by similar changes but with progressively slower kinetics during early and later childhood and only minimal changes later in life. Developmentally downregulated H3K4me3 peaks in prefrontal neurons were enriched for Paired box (Pax) and multiple Signal Transducer and Activator of Transcription (STAT) motifs, which are known to promote glial differentiation. In contrast, H3K4me3 peaks subject to a progressive increase in maturing prefrontal neurons were enriched for activating protein-1 (AP-1) recognition elements that are commonly associated with activity-dependent regulation of neuronal gene expression. We uncovered a developmental program governing the remodeling of neuronal histone methylation landscapes in the prefrontal cortex from the late prenatal period to early adolescence, which is linked to cis-regulatory sequences around transcription start sites.
| Prolonged maturation of the human cerebral cortex, which extends into the third decade of life, is critical for proper development of executive functions such as higher-order problem-solving and complex cognition. Little is known about changes of post-mitotic neurons during this prolonged maturation period, including changes in epigenetic regulation, and more broadly, in genome organization and function. Such knowledge is critical for a deeper understanding of human development, cognitive abilities, and psychiatric diseases. Here, we identify 1,157 genomic loci in neuronal cells from the prefrontal cortex that show developmental changes in a chromatin mark, histone H3 trimethylated at lysine 4 (H3K4me3), which has been associated with regulation of gene expression. Interestingly, the overwhelming majority of these developmentally regulated H3K4me3 peaks were defined by rapid gain or loss of histone methylation during the late prenatal period and the first year after birth, followed by slower changes during early and later childhood and minimal changes thereafter. The genomic sequences showing these dynamic changes in H3K4me3 were enriched with distinct transcription factor motifs. Our findings suggest that there is highly regulated, pre-programmed remodeling of neuronal histone methylation landscapes in the human brain that begins before birth and continues into adolescence.
| Prolonged maturation of the human cerebral cortex, which extends into the third decade of life, is critical for proper development of executive functions, such as higher order problem solving and complex cognition [1], [2]. Little is known about changes within the nuclei of post-mitotic neurons during this prolonged maturation period, including possible changes in epigenetic regulation of DNA and histone proteins. This lack of knowledge is remarkable given that neurogenesis and subsequent permanent exit from the cell cycle by newly generated cortical neurons takes place within the first 3–4 months of prenatal development. It is likely that dynamic changes in epigenetic regulation of neuronal gene expression extend far beyond the first trimester of pregnancy. For example, various periods of growth and differentiation in early and late childhood have been found to be followed by highly dynamic waves of pruning and remodeling of synapses and neuronal connectivity in the prefrontal cortex (PFC) and related areas [3].
Human postmortem brain studies have revealed robust transcriptional changes during the transition from the prenatal period to childhood and into adulthood, with broad implications for inhibitory and excitatory neurotransmission, myelination, metabolism, and various other cellular functions [4], [5]. Thus, it is reasonable to hypothesize that the developmental milestones of cortical neurons are linked to ‘pre-programmed’ changes in neuronal gene expression. These mechanisms are likely to be associated with the process of neural differentiation because genes controlling the process of cell division show a general decline in expression during prenatal development and infancy, while genes associated with synaptic functions and neurotransmission show an increased expression during the same period [5]. Likewise, adolescence and early adulthood (15–25 years) are accompanied by a transient increase in transcripts for energy metabolism as well as protein and lipid synthesis, in conjunction with a further decline in genes implicated in neurodevelopment and plasticity [6]. Each of these developmentally defined transcriptional changes simultaneously involves hundreds to thousands of genes distributed throughout the entire genome, raising the question of whether there is a coordinated unfolding of neuronal gene expression as the PFC matures.
Quantifying developmentally regulated changes in neuronal transcriptomes and epigenomes, however, is confounded by the rapidly occurring and dramatic shifts in cell composition during maturation of the human cerebral cortex. For example, there is a post-natal increase of several fold in the number of astrocytes generated by cell division from local precursors [7]. Although the prenatal cortical plate is overwhelmingly comprised of post-mitotic neurons, neuron-to-glia ratios in the mature primate cerebral cortex are in the range of 0.6 (human) to 1.7 (macaque) [8]. Developmental changes in the cortical transcriptome may largely reflect this underlying change in cell composition and, when explored in tissue homogenate, could mask cell type-specific regulation. To circumvent the potential confound of shifting cell compositions, we have developed the technique of sorting and separating neuronal and non-neuronal nuclei of human post-mortem PFC specimens for subsequent preparation of mono-nucleosomes (nucleosomes are the fundamental units of chromatin, composed of a histone core—the H2A/H2B/H3/H4 octamer—and 146 bp DNA wrapped around it) for genome-wide histone methylation mapping [9], [10]. For this work, we focused on histone H3 trimethylated at lysine 4 (H3K4me3), a histone modification sharply regulated around transcription start sites (TSS) and other regulatory sequences [11], [12]. The H3K4me3 mark broadly correlates with RNA polymerase II occupancy at sites of active gene expression [13], and is largely non-overlapping with promoter-associated DNA cytosine methylation and other repressive marks [14], which provides an additional component of transcriptional regulation [15], [16].
In an earlier study, we presented high-resolution maps of H3K4me3 in neuronal and non-neuronal nuclei collected from the PFC of 11 individuals ranging from 0.5 to 69 years [9]. We identified thousands of genes that showed highly enriched H3K4me3 levels in neuronal but not non-neuronal PFC chromatin. We also identified hundreds of genes (many function in developmental processes) with decreased H3K4me3 during the first year after birth. Here, building upon our earlier results, we present H3K4me3 maps of 36 human PFC specimens collected from the late prenatal period to 81 years of age, with an emphasis on comparing the H3K4me3 landscape in prenatal PFC with that of older ages. We present evidence that 1157 loci within the neuronal epigenome of the PFC are subject to ongoing remodeling of the local H3K4me3 landscape. This remodeling is defined by a unidirectional trajectory of progressive gain or loss of H3K4me3, with steeper changes in the late prenatal period until the first year after birth, and slower changes thereafter extending into the adolescence and adult period. The genes near these age-dependent H3K4me3 peaks showed similar temporal expression patterns, i.e., genes near regions with increasing H3K4me3 also show increasing expression as a function of age and vice versa. Furthermore, in neuronal chromatin, genes near developmentally up-regulated H3K4me3 peaks were enriched in various functional categories related to mature neurons. In contrast, non-neuronal chromatin was defined by a progressive H3K4me3 increase at transcription start sites (TSS) associated with myelination and other glial-related functions. Our results draw a connection between cis-regulators of chromatin structure and function as well as the molecular mechanisms governing the maturation of the human prefrontal cortex, starting with prenatal development and continuing into adolescence and beyond.
We analyzed 36 datasets of H3K4me3 chromatin immunoprecipitation followed by sequencing (ChIP-seq) of sorted nuclei from the human prefrontal cortex. Thirty-one of these datasets were for neuronal NeuN+ nuclei and as control, the other five datasets were for non-neuronal (NeuN–) nuclei. The neuronal (NeuN+) samples covered an age range from 34 gestational weeks (gw) to 81 years (yr), including three prenatal samples (C1–C3) and three infant samples (<1 yr old; C4–C6). The NeuN– samples spanned an age range of 40 gw to 69 yr, including one prenatal sample and one infant sample. Among these datasets, 16 neuronal samples (including the 3 prenatal) and two NeuN– samples were newly generated and sequenced for this study, and the remaining 18 datasets were taken from our previous publications [9], [10]. We also used one input dataset based on the NeuN+ cells of C28 [10] that had been previously generated, but for which no antibody was added though processed in the same way as the above ChIP-seq samples (Table 1).
Samples were sequenced using Illumina GAII and a total of 390 million (M) 36-bpreads were obtained, where 318 M reads mapped to unique locations in the human genome (Table 1). The data are available at http://zlab.umassmed.edu/zlab/publications/ShulhaPLOSGen2013.html. On average, 50% uniquely mapped reads in ChIP-seq samples were in promoters (within 2 kb of the TSS) of RefSeq genes (http://www.ncbi.nlm.nih.gov/RefSeq/; 35,519 transcripts in total), and 4% reads of the input sample were in promoters, consistent with the prior knowledge that H3K4me3 is strongly enriched around TSS [17]. We used the MACS algorithm [18] to detect the genomic regions significantly enriched in H3K4me3 (called peaks) in each ChIP-seq sample compared with the input sample. We then constructed a pool of 47,281 neuronal peaks detected in at least one neuronal sample, and another pool of 42,693 peaks detected in at least one NeuN– sample. We performed further analysis on these two pools of peaks. For each peak, we counted the total number of reads in each sample, normalized by the total number of reads in that sample that mapped to within 2 kb of all RefSeq TSSs.
To investigate promoter H3K4me3 occupancy profiles in neurons, we constructed a vector for each neuronal sample with 21,084 elements (the total number of unique RefSeq TSSs excluding chromosome Y), each element being the normalized number of reads in a RefSeq promoter. Then we computed the Pearson correlation coefficient for all pairs of samples and presented the results as a heatmap (Figure 1A). All 31 samples are highly similar, with correlations above 0.9. The most distinct samples are the three prenatal samples C1–C3, followed by the three infants C4–C6, as shown by the darker shades of the first six columns in Figure 1A. Thus, from the prenatal period to less than 1 year of age to older childhood and adulthood, there is an age-dependent progression of the genome-wide H3K4me3 profile in annotated promoter regions. These findings are consistent with observations we previously reported [9], [10] using much smaller cohorts and a narrower age range.
To determine how the overall population of H3K4me3 peaks (regardless of their association with any annotated TSS) differ across the neuronal samples, we performed principal components analysis on all H3K4me3 peaks as previously described [10]. Principal components analysis is a mathematical method that reduces the dimensionality of data, in our case, a 31 by 31 matrix for which each element is the number of H3K4me3 peaks found in one sample contrasted with another sample. The analysis identifies directions, called principal components, which maximize the variation in the data. Samples are then plotted along the principal components to reveal clusters. Figure 1B shows the 31 neuronal samples plotted against the first two principal components, which in combination account for 95% of the variation in our data. Strikingly, the first principal component separates the prenatal samples (in green) away from all other samples, and the second principal component separates the prenatal samples and the infant samples (in blue) from the remaining samples (in red). The third and fourth principal components do not further separate the samples older than 1 yr into subgroups (data not shown). In order to investigate whether the clear separation is due to the larger number of samples in the >1 yr group, we also performed PCA with 12 samples, 3 samples in each age range: gestational, 0–1 years, 3–14 years, and 15–25 years. The results looked very similar, namely the first two principal components could clearly distinguish the gestational and infant groups from the remaining samples (Figure S1). Thus, neurons in the prefrontal cortex undergo substantial changes in their H3K4me3 landscapes during the transition from the last 2 months of gestation to postnatal life, to a somewhat lesser degree during the first year after birth, and comparatively minor changes for the remainder of the lifespan. With the exception of the three prenatal samples who were all females, all other age groups comprised males and females. However, neither sample-to-sample Pearson correlation analysis of RefSeq promoters nor PCA analysis (Figure 1A–1C) showed any evidence for a role of gender in the age-dependent remodeling of neuronal and non-neuronal H3K4me3 landscapes in the PFC. Notably, a recent study in a cohort of 269 postmortem specimens collected across the lifespan reported developmentally regulated changes in gene expression and a consistent molecular architecture of the PFC across the human race [5]. Our histone methylation studies support this conclusion because each of our age groups of prenatal, infant, older children and adults included subjects from at least two races (information on race was available for 14 brains: 8 Caucasian and 6 African American).
To identify H3K4me3 peaks subject to age-dependent regulation, we systematically compared the three prenatal samples with the 25 samples older than 1 year of age and found 742 (415) peaks with at least two-fold higher or lower levels in the prenatal samples (p-value<0.05; see Methods for details). We called these ‘down’ and ‘up’ peaks, reflecting the change of H3K4me3 upon aging (Table S1). We computed the average H3K4me3 levels for the up and down peaks in each sample. We observed a progressive increase (for the up peaks; Figure 2A) and decrease (for the down peaks; Figure 2B) as a function of age that continued at least for the first 10–20 years of life. Note that each of the three infant samples (<1 yr; blue dots in Figure 2A, 2B) was positioned in between the prenatal samples and the older child/adult samples although these three samples were not used to define the up and down peaks. Figure S2 uses boxplots to illustrate the distributions of H3K4me3 levels for the peaks within the ‘up’ and ‘down’ groups, respectively. It is clear that the variation among genes within each group is smaller than the difference between the two groups. From these results, we conclude that the PFC neuronal population undergoes a steady and continuous developmental remodeling of H3K4me3 peaks, starting during prenatal life and extending into the first childhood years.
These up and down peaks were within 2 kb of the TSSs of 247 and 508 RefSeq genes respectively, and using the DAVID tool, we asked whether these two groups of genes were enriched in any gene ontology (GO) categories [19]. The 508 genes near the down peaks were enriched in 117 GO processes with a false discovery rate (FDR) of less than 5% (Table S2). Some of the most significant GO categories found in the down peaks included ‘anatomical structure’, ‘organ’, ‘systems’, ‘nervous system’ development (FDR ranges from 1.1e-9 to 1.6e-6), and many other processes of critical importance of early embryonic development (Figure 3A, Table S2). A representative example is the transcription factor SOX11 (SRY-related HMG-box) which among other functions, regulates neuronal differentiation during early development [20] (Figure 3B). The 247 genes near the up-peaks were enriched in 28 GO categories with FDR<5% (Table S2), and the most significant categories included neuron projection (FDR = 2.2e-5), synapse (FDR = 6.1e-4), and axon (FDR = 2.5e-3, Figure 3A). Thus, H3K4me3 levels near neuronal genes related to the function of mature neurons, including synaptic transmission and connectivity are up-regulated during the transition from pre- to post-natal life. A representative example is synaptopodin (SYNPO), encoding an actin-associated protein enriched in dendritic spines and postsynaptic densities [21], [22] (Figure 3C). The genes near the down peaks and genes near the up peaks were similarly enriched in 11 GO categories, mostly involving cell-cell communication and signal transduction function (colored blue in Table S2); however, distinct sets of genes led to the enrichment of each shared category.
Some of these genes contain multiple promoters that overlapped with multiple age-dependent H3K4me3 peaks. The first group contained eleven genes whose alternative promoters overlapped H3K4me3 peaks with opposing developmental changes (one up peak and one down peak). Some of these genes are essential for normal brain development, including SHANK2, encoding a synaptic scaffolding protein that when mutated is responsible for mono- or oligogenic causes of autism and other neurodevelopmental disease [23], [24]; LFC/ARHGEF2, encoding a Rho-specific guanine nucleotide exchange factor important for neurogenesis and dendrite spine morphology [25], [26]; and PLEKHG5, a Pleckstrin homology domain-containing gene that is responsible for an autosomal recessive form of motor neuron disease [27] (Table S3).There were also five genes each with two promoters subject to a developmental increase in H3K4me3, including the calcium sensor CABP1,which regulates voltage-gated Ca2+ channels and synaptic short term plasticity [28], and SEPT9, a member of the cytoskeleton-related septin family, which is responsible for hereditary neuropathic syndromes such as amyotrophic neuralgia [29] (Table S3).Furthermore, the third set of nine genes harbored multiple promoters that were subject to a coordinated developmental downregulation of H3K4me3, including the trans-membrane protein NOTCH3, which is essential for preventing premature death of young differentiating neurons [30] and responsible for some forms of vascular disease in the mature brain [31], the Wingless-Int (WNT) protein, WNT7B, which shows distinct regional expression patterns during human brain development, including progenitor cells and sub-layers of the developing cortical plate [32], and DPYSL3, which encodes for a dihydropyrimidinase-like protein important for neuronal differentiation and regulated by NMDA glutamate receptors [33] (Table S3).
Earlier studies reported that the process of differentiation from pluripotent stem cells was associated with dynamic changes in the shape of the H3K4me3 profile, due to spreading into neighboring nucleosomes in differentiated tissue, thereby resulting in broader peaks [34], [35]. To find out whether maturation of PFC is associated with similar changes in neuronal H3K4me3 landscapes, we determined for each subject numbers and proportion of peaks across six different length categories (from 500 bp to 5 kb). However, no age-dependent trends emerged, because peaks <1 kb in length comprised the large majority of the total pool of peaks, while longer peaks (>4 kb) contributed little (3% or less) to each sample (Table S4). This result, however, is unsurprising, because the marker used in our study for nuclei sorting (NeuN) specifically labels postmitotic neurons and excludes stem cells. Furthermore, in good agreement with the above studies [34], [35] reporting that genes specifically expressed in differentiated tissues show wider histone methylation peaks, the broadest peaks (>4 kb) in neurons in the present study also showed a highly significant enrichment of Gene Ontology categories defining neurons. These categories include nervous system development, neurogenesis, neuron differentiation, axonogenesis, neuron projection, synapse, and postsynaptic density among others (Table S5).
The up and down H3K4me3 peaks described above show roughly monotonic changes of H3K4me4 levels as a function of age (largely driven by rapidly changing H3K4me3 levels during the transitions from the prenatal period to infancy and from infancy to early childhood). To test whether our dataset harbors genomic loci with a different, or non-monotonic age profile, we performed k-means clustering, an unsupervised learning technique to separate all H3K4me3 peaks into five clusters (k = 5) so that peaks within each cluster have similar age profiles. In Figure S3, peaks were shown to be increased in Cluster 1 and decreased in Clusters 2 and 3 with age (with Cluster 2 showing a greater decrease than Cluster 3, but both clusters are defined by the largest shifts occurring within the first few years of life). Clusters 4 and 5 showed age-invariant H3K4me3 profiles. We also performed clustering using larger k values but the results were qualitatively the same, with additional clusters showing similar patterns. Therefore, consistent with the hypothesis-driven method described above for identifying the up and down peaks, unbiased k-means clustering again resulted in the two main patterns of age-dependent H3K4me3 changes, with some loci going monotonically up and other loci going monotonically down upon aging. Thus we concluded that the overwhelming majority of the developmentally regulated H3K4me3 changes in neuronal chromatin of the prefrontal cortex are unidirectional and monotonic, with the changes during the successive transitions from the prenatal period to early infancy and from infancy to later childhood ages being much more pronounced than any shifts that may occur later in life.
We also performed GO enrichment analysis on the genes near Cluster 1 and 2 peaks, and the results are shown in Table S6. Because the peaks in Clusters 1 and 2 match the up and down peaks defined in the previous section, the results of GO enrichment are highly consistent. Specifically, the genes near Cluster 1 peaks are enriched in developmental processes, with ‘anatomical structure’, ‘system’, ‘organ’ development, ‘neurogenesis’ (and many other GO categories in Table S2) again among the most significantly enriched, and the genes near Cluster 2 peaks are enriched for many neuron-related categories, including ‘axon’, ‘neuron projection’ and others.
H3K4me3 level is a good indicator of gene expression, and we wanted to further test whether the genes near the age-dependent H3K4me3 peaks show similar age-dependent expression patterns. Colantuoni et al. performed microarray experiments to assay the genome-wide transcription levels in the prefrontal cortex of 269 subjects spanning the majority of the human lifespan, including 38 prenatal samples (14–20 gws) and 18 infants (<1 yr) [5]. Using their data, we plotted, in Figure 4, the average expression levels of the 202 and 419 genes that are near our up and down H3K4me3 peaks which also assayed by Colantuoni et al. as a function of age. The expression signal as defined by Colantuoni et al. is the log2 density ratio of a particular sample over the reference sample produced by pooling all test samples [5]. It is clear that the genes that are near the up peaks are more highly expressed in older samples (Figure 4A) and the genes that are near the down peaks are expressed at lower levels in older samples (Figure 4B).
Similarly, we plotted the average expression patterns for the genes near each of the H3K4me3 ChIP-seq peaks in the five clusters determined by k-means clustering (Figure S4). Indeed, the average expression profile for genes in each cluster follows the same trend as the average H3K4me3 profile: genes near Cluster 1 peaks increased expression upon aging, genes near Cluster 2 and 3 peaks decreased expression upon aging with Cluster 2 showing a greater decrease than Cluster 3, and genes near Cluster 4 and 5 peaks show invariant expression across the age span.
Note that the prenatal gene expression data in Colantuoni et al. were for the 14–20 gw period [5], earlier than the stage of our H3K4me3 data (34–40 gw). They reported four groups of genes that showed significant changes in expression across age: genes that increased expression in both prenatal and infant stages (up-up genes), genes that decreased expression in both stages (down-down genes), genes that increased expression during the prenatal stage and then decreased expression during the infant stage (up-down genes), and genes that decreased expression during the prenatal stage and then increased expression during the infant stage (down-up genes). We plotted the H3K4me3 profiles for these four groups of genes (Figure S5). Indeed, both the up-up and down-up genes show a monotonic increasing H3K4me3 pattern (the left two panels of Figure S5), and both the down-down and up-down genes show a monotonic decreasing H3K4me3 pattern (the right two panels of Figure S5) However, the changes in these patterns are not as pronounced as our up and down H3K4me3 peaks (Figure 2 and Figure S3). Detailed examination of the expression patterns in Figure 2 of Colantuoni et al. indicates that the reversal of the expression patterns for the down-up and up-down genes occurs at around 19 gw. Because our H3K4me3 data are for later time points (34–40 gw), our two H3K4me3 patterns are consistent with the four expression patterns by Colantuoni et al.
In the human cerebral cortex, the genome-wide distribution of H3K4me3 is largely anti-correlated with methyl-cytosine densities [14] and many genes in the genome show a robust DNA methylation increase in CpG dense sequences, including those residing in proximal promoters [36], [37]. Using the datasets from [37], prenatal to postnatal changes in methyl-CpG densities in cortical tissue homogenates were available for 394 (744) CpGs that were in the promoters of genes subject to developmental H3K4me3 up- (down-) regulation in our study on PFC neurons (Table S7 and Table S8). Much higher percentages of the CpGs (59.9% and 60.1% respectively) were in the promoters of the genes near the up or down H3K4me3 peaks than the remaining 26,446 CpGs (42.0%) which show significantly different (FDR<0.05) methylation levels between prenatal samples and samples older than 1 yr. Moreover, 15.7% of CpGs associated with an increase in H3K4me3 during development showed a significant decrease in methyl-CpG levels (FDR<0.05), while only 6.2% genes associated with declining H3K4me3 levels also showed a significant decrease in methyl-CpG levels (Table S9). Thus, there is a significant anti-correlation between the age-dependent change of H3K4me3 levels and the age-dependent change in DNA methylation levels (p-value = 7.0e-8; hypergeometric test). These results are robust regardless of the cutoff for calling a DNA methylation as significantly different between prenatal samples and samples older than 1 yr (the lower half of Table S9 shows the results for the cutoff of FDR<1e-10).Examples of genes with significant and opposing changes in H3K4me3 and DNA methylation include ARC, NR4A1 and other transcription factors with critical roles in synaptic plasticity, learning and memory [38], [39], ANK3, a psychiatric susceptibility genes encoding a synaptic scaffolding proteins [40], and ligand-gated ion channels including the GABAA receptor subunit GABRD, which has been linked to mood and seizure disorders [41], [42].
We also tested whether the sequence motifs of any transcription factors (TF) are enriched in the up or down peaks using the Clover algorithm [43] with all TF binding motifs in the TRANSFAC database [44]. We used two types of background sequences, one was generated by shuffling the sequences of the up or down peaks while preserving dinucleotide frequencies, and the other was the H3K4me3 peaks that did not vary their intensities across age (peaks in Clusters 4 and 5). Table S10 lists the motifs that are significantly enriched in the up or down peaks according to both types of backgrounds (FDR<0.05). Among them, the motif ofAP-1 is significantly enriched among the up peaks. AP-1 (heterodimer of c-Jun and c-Fos) is a classical early response transcription factor and master regulator of the axonal response in neurons. It also functions as a negative regulator of myelination in Schwann cells (SCs) and is strongly reactivated in SCs upon axonal injury [45]. The Pax motif is highly enriched in the down peaks. Among the Pax family of transcription factors, Pax2 and Pax8 specify GABAergic cell fate [46], Pax6 is the master regulator of the visual system [47], and Pax8 is important for hindbrain neurons [48]. The motif of Rp58 is also enriched in the down peaks, and Rp58 is recently shown to be essential for the patterning of the cerebellum and for the development glutamatergic and GABAergic neurons [49]. We also noticed significant enrichment for Signal Transducer and Activator of Transcription (STAT) motifs in the down H3K4me3 peaks from PFC neurons, which agrees well with the finding that STAT-dependent signaling pathways primarily promote astrocytic and other non-neuronal differentiation in the developing cortical plate [50]–[52].
To further explore which, if any, of the above mentioned developmental H3K4me3 changes are specific to PFC neurons, we explored H3K4me3 landscapes in non-neuronal (NeuN–) nuclei obtained from one prenatal, one infant and three adult specimens (Table 1). Similar to the findings in the 31 neuronal samples described in Figure 1A, computation of Pearson correlation coefficient for all five pairs of NeuN– samples revealed higher correlations between samples of similar age (Figure 1C). The smaller number of NeuN– samples prevented us from using the same approach for identifying up and down H3K4me3 peaks as for NeuN+ samples, i.e., directly comparing prenatal samples with samples older than 1 yr, due to the lack of statistical power. Instead, we used the k-means clustering algorithm to identify 2224 TSS-proximal peaks (within 2 Kb of a TSS) subject to decline upon aging, and 785 TSS-proximal peaks that increased upon aging. The genes whose TSSs are proximal to the NeuN– H3K4me3-down peaks partially overlapped with the genes that are proximal to the NeuN+ H3K4me3-down peaks (1889;241;392 for the NeuN– unique, shared, and NeuN+ unique genes; chi-square test p-value<2.2e-16). Yet, these two sets of H3K4me3-down genes fall into similar GO categories (compare Tables S6 and S11). We determined the GO categories that each set of genes was enriched in and observed a strong correlation between the enrichment scores of the two sets of GO categories (Pearson correlation coefficient R = 0.70; p-value<1e-6). A GO category was included in the calculation if the FDR was less than 0.85 for either gene set and the enrichment score was defined as the –log10(FDR). Both gene sets were highly enriched in five GO categories related to development (multicellular organismal development, systems development, developmental process, anatomical structure development, and nervous system development). These results indicate that similar functional pathways are pruned epigenetically between neuronal and non-neuronal cells even though the genes that belong to these pathways differ between the two cell types.
Similar analysis for H3K4me3-up genes revealed a different picture. The genes whose TSSs are proximal to the NeuN– H3K4me3-up peaks overlapped partially with the genes that are proximal to the NeuN+ H3K4me3-up peaks (675;83;615 for the NeuN– unique, shared, and NeuN+ unique genes; chi-square test p-value<2.2e-16). However, these two sets of genes were enriched in mostly non-overlapping GO categories (compare Tables S6 and S11). The Pearson correlation coefficient between the enrichment scores of the two sets of GO categories was –0.11 (p-value = 0.24). The most enriched GO categories for NeuN– H3K4me3-up genes included compact myelin (FDR = 0.0029) and myelin sheath (FDR = 9.0e-5), consistent with one important function of glial cells, which make up the vast majority of NeuN– cells, and these GO categories were not enriched in any of the other three groups of genes (NeuN+ H3K4me3-up, NeuN+ H3K4me3-down, or NeuN– H3K4me3-down). On the other hand, the NeuN+ H3K4me3-up genes were more enriched in axon, neuron projection, and signal transduction (FDR = 0.0013–0.0052) than the NeuN– H3K4me3-up genes. We conclude that while both neuronal and nonneuronal cells undergo a major remarking of TSS-associated histone methylation during PFC development and maturation, different areas of the genome are affected, depending on cell type.
The present study provides detailed analyses into the developmental regulation of a transcriptional mark, H3K4me3, in neuronal and non-neuronal chromatin during the extended course of PFC maturation. There were 1157 loci, including 768 TSS-proximal and many other regulatory sequences that showed evidence for the developmental remodeling of H3K4me3. Strikingly, these peaks showed similar kinetics as defined by an unidirectional course with the largest decline or increase occurring within the first 1–2 years of postnatal life, followed by a gradual slowing of age-related changes that apparently continue at least until early adolescence or even beyond. We show that these developmentally regulated H3K4me3 peaks at transcription start sites are associated with age-related changes in the expression of the corresponding RNA. We further showed that a subset of regulatory motifs, including Pax and AP-1 transcription factor recognition sites, are overrepresented among the developmental regulated peaks showing a decrease (Pax, Stat), or increase (AP-1) respectively, during the course of PFC maturation. Finally, the developmental remodeling of H3K4me3 landscapes in PFC is cell type specific. Collectively, our results draw a connection between cis-regulators of chromatin structure and function and the molecular mechanisms governing maturation of the human prefrontal cortex.
The findings presented here, when taken in conjunction with studies exploring developmental changes in DNA methylation [36], [37] and gene expression across the lifespan of the human PFC [5], paint a picture in which immature PFC, during the weeks and months preceding and following birth, undergoes a larger scale reprogramming of transcriptomes and promoter-associated epigenetic regulators, including promoter-bound DNA and histone methylation. This reprogramming involves hundreds of TSS that define cellular functions that are either characteristics of differentiated neurons and glia (e.g. synaptic transmission, myelination) or functions related to earlier stages of development (e.g. neurogenesis, nervous system development) that decline as the PFC matures. By charting a developmental map for the neuronal and non-neuronal constituents of the PFC separately, the present study and our earlier studies [9], [10] largely extends the previous work on tissue homogenate that allows for limited data interpretation due to age-related shifts in neuron-to-glia ratios and other confounding factors. The present study further emphasizes the prenatal stage and shows that developmental remodeling of TSS-associated histone methylation in PFC neurons rapidly changes in prenatal and infant stages but continues at a slower pace deep into the second decade of life.
Our study faces several important limitations. Given that no technique with single cell resolution exists to map histone methylation levels at specific loci, our assays by design only inform about H3K4me3 profiles averaged across millions of cell type-specific nuclei that are required for the ChIP-seq assays. Thus, the cell population-based developmental kinetics of H3K4me3 with the simple and unidirectional exponential curves, as presented here, leave open whether individual PFC neurons would show a more dynamic interplay between H3K4 methylation and demethylation. Furthermore, our prenatal specimens were limited to the mid- and late stages of the third trimester, and it remains possible that brains of an earlier prenatal age could show a more complex or multi-layered regulation of the H3K4me3 marks, compared to what is reported here. For example, there are four groups of genes reported by Colantuoni et al., who compared RNA expression at earlier stages of gestation (14–20th week of pregnancy) with those of infant and older brains [5]. Nonetheless, these four groups of genes could be recognized by their age-dependent H3K4me3 profiles in our study (Figure S5), reaffirming the view that these gene expression networks are co-regulated both on the level of the transcriptome and the epigenome.
The present study suggests that the developmental remodeling of TSS-associated histone methylation in PFC neurons in the weeks and months before and after birth continues at a slower pace deep into the second decade of life. Presently, nothing is known about the molecular ‘clocks’ or ‘pacemakers’ that orchestrate such steadfast remodeling of TSS-associated H3K4me3 during the first 10–20 years of life and, to the best of our knowledge, these phenomena await further investigation in laboratory animals. Indeed, a recent H3K4me3 ChIP-seq study in whole tissue of four macaque PFC specimens found evidence for histone methylation changes during the course of maturation and aging [53]. Insights into these mechanisms bear great promise for a better understanding of normal development and the pathophysiology of schizophrenia and autism and other neurodevelopmental disorders associated with DNA and histone methylation changes in the PFC [10], [54]–[58]. To this end, it is interesting that AP-1 transcription factor motifs are enriched in H3K4me3 peaks that are upregulated during the course of PF maturation. Of note, antipsychotic drug treatment administered over the course of 2–3 weeks resulted in lasting increases of AP-1 protein and transcript in rat rostromedial cortex (considered the functional homologue to primate PFC) [59]. Because AP-1 in PFC and other brain regions is highly regulated by neuronal activity [60], up-regulation of AP-1 expression and AP-1 mediated transcriptional activity, either during normal development or in the context of psychopharmacological treatments, could serve as one of the molecular drivers for the regulation of H3K4 trimethylation at AP-1 bound promoters and other active TSS in the mature PFC.
All postmortem tissue work was done in compliance with the Institutional Review Board regulations of the University of Massachusetts Medical School and the Mount Sinai School of Medicine. Freshly frozen (never fixed) tissues from the rostral prefrontal cortex of subjects ranging in age from the 34th week of gestation to 81 years, was provided by four independent brain banks (Table 1). Tissue aliquots (200–500 mg/subject) were extracted in hypotonic lysis buffer, purified by ultracentrifugation and resuspended in 1x PBS, immunotagged with anti-neuronal nucleus (anti-NeuN, Millipore) antibody and sorted into NeuN+ and NeuN– fractions using a FACSVantage SE flow cytometer, as described [61], [62].
Mononucleosomal preparations from at least 1×106 sorted nuclei were prepared for subsequent chromatin immunoprecipitation with anti-H3K4me3 antibody (Upstate/Millipore), and ChIP-seq libraries prepared from the immunoprecipitated DNA by blunt-ending, A-tailing and ligation to adaptors (Genomic Adaptor Oligo Mix, Illumina) and PCR amplification and sequencing on an Illumina Genome Analyzer II platform, as described [9], [63].
All of our sequencing libraries contained single-end 36-bp reads and we mapped them to the human genome with Bowtie (version 0.11.3) [64]. We allowed up to one mismatch and mapped all sequences to the gender appropriate genome HG19. Reads that mapped to multiple locations were discarded. Unique mappers constitute 66–90% of all reads. Detailed statistics is presented in Table 1. As previously reported, H3K4me3 levels at promoters did not show correlations with postmortem interval and tissue pH [9], [65]. Critical ChIP-seq parameters, including the proportion of uniquely mappable sequence tags, and the percentage of uniquely mappable sequences at gene promoters were very similar between samples from the four brain banks, without significant differences (Table 1) (%uniquely mappable, % uniquely mappable at promoters: Harvard Brain Tissue Resource Center (HBTRC): 82.3 ± 4.9, 45.3 ± 19.3; Maryland Psychiatric Research Center (MPRC) 84.4 ± 2.9; 51.9 ± 11.4; University of California at Irvine/Davis (UCI/UCD) 78.7 ± 8.7; 60.3 ± 8.7; University of Maryland Brain and Tissue Bank for Developmental Disorders (UM-BTB),80.7 ± 4.5; 58.2 ± 10.5).
To construct the heatmap in Figure 1, we calculated Pearson correlation coefficients for each pair of samples. We took the genomic coordinates of all TSSs (except chrY) from RefSeq and expanded them in both directions by 2 kb. If some regions overlapped with each other, we merged them together. For each of these 21,084 non-overlapping regions, we computed the total number of mapped H3K4me3 ChIP-seq reads and normalized by size of the region. The resulting read densities were used to compute Pearson correlation coefficients.
The MACS software (version 1.3.5) [18] was used to identify statistically enriched H3K4me3 regions (called H3K4me3 peaks or peaks in short). We contrasted each sample against the input sample using bw = 230; tsize = 36 and default values for the remaining parameters in MACS.
Principal Component Analysis was performed on a matrix that contains peaks unique to each sample. Each sample was compared against every other sample using the MACS software with parameters mentioned above. The H3K4me3 peaks thus obtained were further filtered using the following criteria: (1) MACS p-value must be less than 1e-20, (2) read density ratio between the two samples must be greater than 4, and (3) normalized read density must be greater than 0.005.
To search for age-dependent H3K4me3 peaks, all peaks from NeuN+ samples were combined and overlapping peaks merged, resulting in 47,281 peaks. The 742 down peaks (≥1 k bp) were defined as: (1) the average read density in prenatal samples must be greater than or equal to 0.01, (2) the ratio of average read density of prenatal samples and the 25 samples older than 1 year must be greater than or equal to 2, and (3) the t-test p-value for comparing the three prenatal samples with the 25 samples older than 1 year must be less than or equal to 0.05. The reciprocal criteria were used for defining the 415 up peaks.
We performed k-means clustering on the age profiles of all H3K4me3 peaks. We limited our calculations to regions that were ≥1 k bp and had an average H3K4me3 read density ≥0.01 in prenatal samples or in the 25 samples older than 1 year. The small number of peaks in chrY was excluded from this analysis. This resulted in 14,708 regions that were further normalized by the strongest signal (across all samples) for each region. We then used the “k-means” procedure from the R software with 5 clusters. We averaged the H3K4me3 levels across the regions in each cluster and plotted the resulting average H3K4me3 profile for each cluster in Figure S3. We used the same approach to perform k-means clustering for NeuN– samples.
We used the DAVID web-server for detecting enriched Gene Ontology categories. For each set of H3K4me3 peaks, genes whose TSSs were within 2 k bp of an H3K4me3 peak are included in the analysis. We used false discovery rate (FDR) to quantify statistical significance, because it accounts for multiple testing correction.
To compare our H3K4me3 data with DNA methylation [37] we downloaded the data from the author's website (http://braincloud.jhmi.edu/downloads.htm).The dataset contains methylation level for CGs in a set of gene promoters. CGs in every gene that matched our proximal H3K4me3 peaks were analyzed for age-dependent changes (Tables S7 and S8). We performed a t-test for every gene, comparing all prenatal samples with all samples older than 1 year, and computed false discovery rate (FDR) after multiple testing correction. Cases with significant DNA methylation changes (using two cutoffs, FDR<0.05 or FDR<1e-10) were used to establish relationship with H3K4me3 (Table S9).
|
10.1371/journal.ppat.1002103 | The Binding of Triclosan to SmeT, the Repressor of the Multidrug Efflux Pump SmeDEF, Induces Antibiotic Resistance in Stenotrophomonas maltophilia | The wide utilization of biocides poses a concern on the impact of these compounds on natural bacterial populations. Furthermore, it has been demonstrated that biocides can select, at least in laboratory experiments, antibiotic resistant bacteria. This situation has raised concerns, not just on scientists and clinicians, but also on regulatory agencies, which are demanding studies on the impact that the utilization of biocides may have on the development on resistance and consequently on the treatment of infectious diseases and on human health. In the present article, we explored the possibility that the widely used biocide triclosan might induce antibiotic resistance using as a model the opportunistic pathogen Stenotrophomonas maltophilia. Biochemical, functional and structural studies were performed, focusing on SmeDEF, the most relevant antibiotic- and triclosan-removing multidrug efflux pump of S. maltophilia. Expression of smeDEF is regulated by the repressor SmeT. Triclosan released SmeT from its operator and induces the expression of smeDEF, thus reducing the susceptibility of S. maltophilia to antibiotics in the presence of the biocide. The structure of SmeT bound to triclosan is described. Two molecules of triclosan were found to bind to one subunit of the SmeT homodimer. The binding of the biocide stabilizes the N terminal domain of both subunits in a conformation unable to bind DNA. To our knowledge this is the first crystal structure obtained for a transcriptional regulator bound to triclosan. This work provides the molecular basis for understanding the mechanisms allowing the induction of phenotypic resistance to antibiotics by triclosan.
| The wide utilization of biocides for different purposes, including toothpastes, soaps, house-hold compounds surfaces' disinfectants and even their use as additives of different materials (from textiles to concrete used in germ-free buildings) to avoid their colonization by microorganisms, poses a concern on the impact of these compounds on natural bacterial populations. Furthermore, it has been demonstrated that such biocides can select, at least in laboratory experiments, bacteria resistant to antibiotics. This situation has raised concerns on the impact that the utilization of biocides may have on the development on resistance and consequently on the treatment of infectious diseases. In the present article we study whether biocides can induce phenotypic resistance to antibiotics, a process that would be barely detectable unless purposely searched out. In the article, we present functional, biochemical and structural data showing that the widely used biocide triclosan induces antibiotic resistance, mediated by the binding of the biocide to SmeT, the transcriptional regulator of the expression of the Stenotrophomonas maltophilia multidrug efflux pump SmeDEF, which can extrude an ample range of antibiotics. Our study provides an unambiguous link between the presence of this biocide and the increased efflux of antibiotics by the opportunistic pathogen S. maltophilia.
| The widespread use of biocides in toothpastes, soaps, household cleaning agents, surface disinfectants and as additives in different materials (from textiles to the concrete used in germ-free buildings) etc., all with the aim of preventing microbial colonization [1]–[5], could have an undesired impact on natural bacterial populations [1], [6]–[8]. Biocides have been associated with the in vitro selection of bacterial mutants showing reduced susceptibility to antibiotics (cross-resistance) without the need for any antibiotic-selective pressure [9]–[12], although whether this occurs in the wild is less clear. Triclosan is one of the most widely used biocides [13]. Using different models it has been shown that resistance to triclosan can be conferred by the expression of multidrug (MDR) efflux pumps capable of expelling antibiotics [9], [11], [14], [15]. Mutants overexpressing MDR efflux pumps are easily obtained under antibiotic selective pressure [16]–[18]. It has also been shown that triclosan can select for mutants that constitutively overproduce such pumps and which are thus less susceptible to antibiotics [9], [11], [14], [15]. The constitutive overexpression of MDR efflux pumps is very often due to mutations in the local transcriptional regulators that control pumps expression or, in a few cases, to mutations in their operator DNA sequences [19]–[21].
The expression of chromosomally-encoded MDR efflux pumps is tightly controlled by specific transcriptional regulators (usually repressors). Under normal growing conditions in the laboratory, MDR efflux pumps are expressed at a very low level (if they are expressed at all) [9], [11], [14], [15]. However, their expression can be activated by the binding of effectors to their repressors and the consequent inhibition of the binding of such repressors to their operators [22]–[27]. Although most work on bacterial efflux pumps has focused on their impact on antibiotic resistance, antibiotics are not always the natural inducers of their expression [20]. In fact, in spite of the broad range of substrates that efflux pumps can expel, only a narrow group of ligands can act as effectors capable of triggering the transcription of the operons encoding these pumps.
The present work explores whether the biocide triclosan can activate the expression of MDR efflux pumps. Previous work has shown that triclosan selects mutants that overproduce the Stenotrophomonas maltophilia MDR efflux pump SmeDEF [11]. This efflux pump belongs to the resistance-nodulation-cell division family and is a tripartite efflux pump formed by an inner membrane protein, which is the transporter itself (SmeE), an outer membrane protein (SmeF) and a membrane fusion protein (SmeD). S. maltophilia is often isolated from the rhizosphere and from water sources [28], [29]. Besides this environmental origin, this bacterial species is an opportunistic pathogen, which presents low susceptibility to several antibiotics [30], [31], and is involved in different types of infections with a considerable mortality rate [32]. Infections by S. maltophilia include bacteremia [33], endocarditis [34], infections in patients with cancer [35] and respiratory tract infections, including those suffered by cystic fibrosis patients [36]–[38] among others. The genome of this bacterial pathogen harbors a large number of antibiotic resistance determinants [39], [40], including antibiotic inactivating enzymes [41]–[43], a qnr determinant [44]–[46] and different MDR efflux pumps, like SmeABC, SmeDEF, SmeJKL and SmeYZ, being SmeDEF the most important MDR efflux pump known to confer antibiotic resistance in S. maltophilia [47]–[52]. The expression of SmeDEF is regulated by SmeT, a transcriptional repressor encoded upstream of smeDEF in the complementary DNA strand [53], [54]. SmeT belongs to the TetR family of transcriptional repressors. The members of this family show a characteristic helix-turn-helix DNA-binding motif at their N-terminal end and a C-terminal region involved in both dimerization and effector binding [55]. The structural analysis of SmeT has revealed this repressor to have close similarities to TetR, QacR and TtgR [53] and to a lesser extent with CprB [56], EthR [57], CmeR [58], AcrR [59], ActR [60], IcaR [61], members of the TetR family of repressors. However, unlike them, SmeT has extensions at its termini that might modulate its interaction with DNA as well as the nature and size of the effector-binding pocket (when empty, this pocket is the smallest of all those of the TetR family members). SmeT binds to a 28 bp-long pseudopalindromic region in the promoter regions of smeDEF and smeT with a Km (app.), calculated from the data presented in [53] in the range of 1 µM, which is similar to that found for the TetR regulator TtgR [62]. The binding of SmeT to its operator region simultaneously represses smeDEF and smeT transcription by the steric interference of RNA polymerase binding to DNA [54]. Constitutive overexpression of smeDEF occurs in mutants selected by triclosan or antibiotics, and these show changes in SmeT that preclude the binding of the repressor to its operator [48], [51]. The possibility of the binding of effectors to SmeT inducing smeDEF expression has been suggested [53], but never demonstrated.
SmeDEF has a wide range of substrates that includes antibiotics, solvents, biocides and dyes [47], [49]. However, no information is available on the inducers of this efflux pump. To ascertain whether the biocide triclosan, which is a substrate of SmeDEF, might also activate its expression, a number of functional and structural analyses were performed. The data collected support the idea that triclosan can induce the expression of smeDEF and consequently reduce the susceptibility of S. maltophilia to antibiotics. This induction is due to the binding of triclosan to the pump repressor SmeT, which impedes its binding to its operator region, and triggers the expression of the most important MDR system in S. maltophilia, SmeDEF. The X-ray crystal structure of the SmeT-triclosan complex indicates that the biocide stabilizes the protein structure in a conformation unable to bind DNA. To our knowledge, this crystal structure is the first structural evidence of the ability of triclosan to act as an effector via its binding to a transcriptional regulator. Given that this regulator (SmeT) mediates the susceptibility of S. maltophilia to antibiotics by repressing smeDEF expression, the present results provide information that aids our understanding of the molecular basis of biocide-induced antibiotic resistance.
To measure the binding kinetics of triclosan to SmeT, we determined the changes of fluorescence of SmeT in the presence of triclosan. This method has been previously used for analyzing the interactions of triclosan with the enoyl-acyl carrier protein reductase [63] and is a good alternative to isothermal titration calorimetry for molecules with low solubility in water as triclosan. The triclosan addition to SmeT samples resulted in a concentration-dependent quenching of the intrinsic protein fluorescence. The fluorescence variations, relative to the untreated samples, were best fitted by a single hyperbola, assuming an stoichiometry of two molecules of triclosan per SmeT dimer and yielding an apparent Kd value of 0.63±0.15 µM (Figure 1). Values for the Kd in the low micro molar range have been described for the binding of drugs to different transcriptional regulators of the TetR family [24], [64]. These results confirmed that in the solution state triclosan interacts with SmeT at concentrations similar to those described for known effectors of other members of the TetR family of transcriptional regulators.
It has been shown that the binding of other members of the TetR family to their cognate DNA operators is modulated in response to effectors such as antibiotics, detergents or plant exudates [55]. To determine whether triclosan is able to induce a conformational change in SmeT and thus modify its DNA binding properties, EMSA was performed with SmeT and a 30-bp DNA fragment containing its operator either in the presence or absence of the biocide. In vivo, SmeT is usually bound to its operator thus repressing transcription of smeDEF. However, the entrance of an effector into its binding pocket might release the effector-SmeT complex from the DNA. To mimic this situation, triclosan was added to preformed SmeT-DNA complexes. As shown in Figure 2, the addition of triclosan to the DNA-SmeT complex resulted in the loss of the retarded band, indicating the separation of the components. The addition to the preformed complex of ciprofloxacin, which is a substrate of SmeDEF [49], did not release SmeT from its operator (not shown). These results suggest that the structural changes suffered by SmeT upon triclosan binding render it unable to bind to its cognate operator DNA.
Since the addition of triclosan precludes the in vitro binding of SmeT to its cognate operator, smeDEF expression ought to be induced by this biocide. To ascertain whether our in vitro data match the physiological in vivo response, the levels of the mRNA from the smeD gene in the presence and absence of the biocide were measured by real time RT-PCR (Figure 3). The expression of smeD was also measured in S. maltophilia D457R. This strain is a natural mutant, derived from the wild-type D457, which has been selected in the presence of antibiotics [50]. The multidrug resistant strain D457R harbors an inactive allele of SmeT as the consequence of a Leu166Gln change [51], [54]. Because of this, the strain D457R constitutively expresses high levels of smeDEF [49] and it is thus a good control for measuring the level of expression of smeDEF under non-repressing conditions. As shown in Figure 3, triclosan increased 8.7-fold the expression levels of smeD compared to the levels observed for cells growing without the biocide. In comparison, S. maltophilia D457R, in which smeDEF transcription is fully de-repressed, showed a 13.7-fold increase for smeD mRNA. These results indicate that triclosan de-represses the transcription of smeDEF in agreement with the data obtained with the EMSA assays described above.
The genome of S. maltophilia harbors genes encoding several putative MDR efflux pumps [40]. Among them, SmeABC, SmeJKL and SmeYZ are known to be involved in antibiotic resistance in S. maltophilia [40], [65], [66]. To determine whether the effect of triclosan was specific for smeDEF or whether other MDR pumps are induced by the biocide, the expression of smeC, smeJ and smeY was examined. As shown in Figure 3, none of these MDR pumps were induced by triclosan, indicating that the effect of this biocide is specific for smeDEF.
Since triclosan induced the expression of smeDEF, it was predicted that the susceptibility of S. maltophilia to antibiotics would be lower in the presence of the biocide. To test this, Etest assays were performed with the biocide and ciprofloxacin. Ciprofloxacin was chosen because the constant over-production of SmeDEF in the strain S. maltophilia D457R, which harbors a defective SmeT repressor, results in an 8-fold increase in the MIC value for this quinolone [48], and because ciprofloxacin does not induce smeDEF expression (AH, unpublished results).
For these assays, a square of dried Whatman paper previously soaked with triclosan was placed just below the point of the Etest strip corresponding to the minimal inhibitory concentration (MIC) of ciprofloxacin, and the MICs in the presence or in the absence of the biocide were determined. A 2.5-fold increase in ciprofloxacin MIC was observed in the presence of triclosan (from 0.75 µg/ml to 2 µg/ml), indicating that the biocide induced resistance to antibiotics, although the increase in MIC was lower than that observed in the S. maltophilia D457R mutant, which constitutively expresses smeDEF at high level [48]. To further confirm that triclosan transiently reduces the susceptibility of S. maltophilia to quinolones, growth curves were plotted for S. maltophilia cultures with or without triclosan in the presence or absence of sub-MIC concentrations of these antibiotics. As shown in Figure 4, at the tested concentrations the presence of the biocide alone slightly slowed the growth of S. maltophilia. However, when bacterial growth was inhibited by adding the antibiotics, the presence of the biocide favored bacterial growth, antagonizing the inhibitory effect of the quinolones. This indicates that the biocide exerts a dual effect on bacterial viability and thus on the susceptibility of S. maltophilia to antibiotics (Figure 4). Triclosan inhibits bacterial growth but simultaneously induces the expression of drug-detoxification elements. This mixed effect might be the cause of the moderate increase in MIC values observed in the presence of triclosan, in spite of highly increased smeD expression.
To gain more insight into the structural basis of the induced expression of SmeDEF by triclosan, SmeT was co-crystallized with this biocide and the structure solved by X-ray diffraction. The structure of the SmeT-triclosan complex involves one homodimer in the asymmetric unit, as seen for other TetR family members [55]. Each SmeT polypeptide chain is composed of 9 α-helices (α1–α9) divided into two structurally distinct domains (Figure 5). The smaller N-terminal domain, composed of the first 3 helices (α1: residues 14–27, α2: residues 24–41, α3: residues 45–49) and the beginning of the fourth, mediates DNA binding through the N-terminal helices α2 and α3, which are almost perpendicular to each other and constitute the DNA-binding helix-turn-helix motif. The larger C-terminal domain, which is mainly involved in ligand binding and dimerization, is composed of helices 4 to 9 (α4: residues 54–76, α5: residues 85–102, α6: residues 104–114, α7: residues 127–149, α8: residues 159–179 and α9: residues 186–201). The dimerization surface, which is mostly formed by α8 and α9 helices, has a hydrophobic character despite the highly negatively charged solvent-exposed surface of the C-terminal domain. This interface involves 48 residues of chain A and 46 residues of chain B, an area of 1718 Å2 and 1796 Å2 respectively. The hydrophobic interactions are complemented by a network of at least eight hydrogen bonds and two salt bridges between Arg134 (chain A) and Glu180 (chain B) and between Arg164 and Asp189 of both monomers. The monomers are almost identical, with a root mean square deviation (rmsd) of 0.708 Å for all atoms (0.680 Å for Cα atoms). The biggest difference between monomers is found at the N terminal domain, with a rmsd value of 0.712 Å compared to 0.611 Å for the C-terminus.
The SmeT-triclosan structure revealed several structural differences with respect to the apo SmeT structure [53] that shed light into the role of the biocide at the binding pocket and its subsequent stabilization of the protein folding specially at disordered regions that could not be modeled in the apo SmeT structure (Figure 6a). Residues 10–12, 21–34 and 45–55 that appeared disordered in apo SmeT become well ordered upon triclosan binding. Therefore, a new interface area between the loop connecting helices α6 and α7 (residues Ser116 – Arg123) and the top of helix α1 (residues His25 – Gly28) and the beginning of the loop connecting α1 and α2 (residues Val29 – Thr33) could be seen as a result of this ordering (Figure 6b). In this area, poorly defined in the SmeT structure [53], hydrogen-bonds formed between Ala120 and Asn121 of one monomer and His25 and Glu26 of the other, respectively, help to stabilize this region.
A closer inspection of the electron density maps in the C-terminal domain showed a clear density in one of the subunits (A) (Figure 7a). This density was readily interpretable and allows the unambiguous placement of two molecules of the biocide (in agreement with the stoichiometry determined by fluorescence measurements) and the subsequent refinement. This is the first structural evidence showing a transcriptional regulator in complex with triclosan. However, no equivalent dense area was seen in subunit B. The volume of the triangle-shaped ligand binding cavity is 1030 Å3, which represents a notable increase with respect to the volume calculated for the apo SmeT structure (630 Å3), but similar to the calculated values for other TetR proteins.
One of the triclosan (TCL1) molecules binds to the bottom of the ligand binding site, in close contact with helices α6, α7 and α8. The phenolic hydroxyl group forms a strong hydrogen bond with the Nδ atom of His167 (placed at 2.2 Å), anchoring this molecule almost parallel to helix α8. In the apo SmeT structure this residue is placed at an alternative conformation into the empty binding site, with the Cδ and Nε atoms pointing to the interior of the cavity. The position and orientation of the triclosan molecule is further stabilized through its 2,4-dichlorophenoxy ring via two edge-to-face aromatic ring interactions, on one side with Phe133 (3.9 Å) and on the other with the 2,4-dichlorophenoxy ring of the second triclosan molecule (TCL2) (4.1 Å). In this scenario the chlorine atoms engage stabilizing interactions with hydrophobic residues at the binding site. Thus, the 5-Cl atom makes contact with the side chains of Val170 (3.5 Å), Leu166 (4.1 Å), Met140 (3.3 Å) and Met93 (3.5 Å). The 2-Cl atom faces Gly132 (3.2 Å), the Cε atom of His67 (3.3 Å), and the phenol ring of TCL2 (4.4 Å). The 4-Cl atom is surrounded by the side chains of Met110 (3.4 Å), Leu114 (3.6 Å) and Met113 (3.6 Å). This atom, together with the phenolic ring of TCL2, displaces these last two residues to expand the active site cavity (see above) and maybe to accommodate a second triclosan molecule (Figure 7b). The hydroxyl group of the phenolic ring of TCL2 forms an H-bond with the Nδ atom of His67 (2.3 Å). The 2,4-dichlorophenoxy ring of TCL2 stacks against the phenol ring of Phe70 (3.4 Å). This residue was modeled in double conformation in the apo SmeT structure but only the so-called open conformation is seen in the structure of the SmeT-triclosan complex (Figure 7c). The density for this residue in subunit B is weaker than in subunit A and a single conformation of the side chain was modeled with an occupancy of 50%. None of the triclosan molecules interacts with residues from subunit B.
In recent years, the possibility that widely-used biocides might co-select for antibiotic resistance has been suggested to pose a potential risk to the successful treatment of infectious diseases [1], [6]–[8]. Although presently there is no clear evidence of the selection of antibiotic-resistant mutants by biocides in the wild, risk-assessment studies are required since in vitro experiments have shown that exposure of bacterial populations to certain biocides, such as triclosan, indeed leads to selection for mutants with reduced susceptibility to antibiotics [11], [14], [67], [68]. On most occasions on which the molecular basis of this resistance has been explored, it has been acquired as a consequence of the stable de-repression of MDR efflux pumps [9], [11], [14], [15].
Since the expression of chromosomally-encoded MDR efflux pumps is usually strongly down-regulated by local repressors [19], and their expression can be triggered by specific effectors [27], the present work explores the possibility that triclosan, a known substrate of the S. maltophilia MDR efflux pump SmeDEF [11], might induce the expression of this antibiotic resistance determinant and thus render an inducible phenotype of antibiotic resistance that would be barely detectable unless searched for. Triclosan-induced phenotypic antibiotic resistance [69] is a possibility since previous work by our group has shown that the exposure of S. maltophilia to triclosan selects for antibiotic resistant mutants that overexpress SmeDEF [11]. The present results show that triclosan binds the repressor of smeDEF transcription SmeT, and that this binding induces a conformational change in SmeT, which is released from its DNA operator. Although additional crystal structure of SmeT-DNA is needed to further confirm this mechanisms. These in vitro data correlate with the in vivo induction of the expression of smeDEF in the presence of triclosan, indicating that the latter is a good inducer of the expression of the MDR efflux pump smeDEF. It might be predicted that this increased expression of smeDEF would lead to a reduction in the MIC similar to that observed in the strain S. maltophilia D457R, which harbors a defective allele of SmeT that cannot repress smeDEF expression. However, the increase in triclosan-induced resistance, although consistent, is modest (2.5-fold) in comparison to that observed for the S. maltophilia D457R mutant that overexpresses smeDEF (8-fold). Although the induction of smeDEF expression achieved by triclosan (8.7-fold) is slightly lower than that observed in the S. maltophilia D457R multidrug resistant mutant compared to its isogenic wild-type strain S. maltophilia D457 (13.7-fold), it is unlikely that these small differences can account for the differences in MICs observed when the expression of smeDEF is induced by triclosan or when it is constitutively de-repressed in the S. maltophilia D457R mutant. The small increase in the observed triclosan-induced resistance might be the consequence of the methods used in the assay. However, several compounds have a concentration-dependent effect on their target organism (beneficial at low concentrations and harmful at high concentrations), a behavior known as hormesis [70]. In the case of triclosan, the biocide induces expression of resistance determinants but simultaneously inhibits bacterial growth (Figure 8). In this case, the dual role described in hormesis occurs within too narrow a window of triclosan concentrations for both effects to be distinguished. This allows a working model to be proposed in which triclosan induces the transient over-expression of smeDEF, but the phenotypic consequences of this overexpression with respect to antibiotic resistance are counteracted by the inhibitory activity of the biocide itself (Figure 8). The observed increase in antibiotic resistance is therefore lower than might be predicted.
The resolution by X-ray diffraction of the structure of the SmeT-triclosan complex crystal allows more insight to be gained into the structure-function relationship of the SmeT transcriptional repressor. This paper is the first to describe the structure of triclosan bound to a transcriptional regulator and of SmeT complexed with a ligand. The SmeT-triclosan complex and the apo SmeT (PDB code: 2W53) structures are almost identical, with an rmsd value of 0.345 Å. However, in the present SmeT-triclosan complex structure nearly all the residues could be modeled while most of those in the N-terminal domain of subunit B were disordered in the apo SmeT structure, even though they were crystallized under the same conditions and show a similar crystal packing lattice (Figure 5). This observation supports the hypothesis that the N-terminal domain of SmeT is intrinsically disordered, being stabilized upon ligand or DNA binding. It is in the interface between the N-terminal domains of both subunits where this effect is most remarkable. The stabilization of this domain due to the binding of the effector reveals a new interface area between the loop connecting helices α6 and α7 and the top of helix α1 (Figure 6b). An analogous contribution to the dimer interface through the DNA binding domain has been described in the TtgR complex [23].
Due to the stabilization of the SmeT structure, the distance between the α-3 recognition helices of both subunits could be measured in the SmeT-triclosan complex, with 44.6 Å between the C-alpha atoms of Tyr49 (the N terminal of subunit B in the apo SmeT structure had to be modeled in order to estimate this length). This value is 10 Å longer than the distance between the B-form DNA major grooves, where these helices should bind (Figure 9). This conformation prevents the repressor from binding to DNA, as observed for the QacR dimer [25]. Further, the conformational change associated with the binding of triclosan causes the release of SmeT from its DNA operator, as seen in the EMSA assay.
The present work shows that two molecules of triclosan enter the hydrophobic binding pocket of SmeT. The binding of two molecules to one subunit of a dimer has been reported for other member of the TetR family, e.g. the binding of two molecules of phloretin to TtgR [23] or the binding of two different drugs to QacR [64]. One molecule of triclosan binds at the bottom of the binding pocket, parallel to helix α6, in a geometry similar to the phloretin molecule when binding to the high affinity site of TtgR. The second molecule, binds close to the dimer interface, and also interacts with this helix through the phenolic ring (Figure 6b). The first molecule of triclosan stacks against the phenol ring of Phe70, a residue identified as the gatekeeper of the ligand binding pocket. Supporting this hypothesis, the so-called open conformation is the only one seen in the structure of the SmeT-triclosan complex. In contrast, Phe70 has a double conformation in the apo SmeT structure.
The ligand binding site of SmeT is mainly formed by hydrophobic residues - there are just three polar residues: His67, His167 and Ser96. This lack of charged residues may be related to the polyspecific substrate recognition of SmeT. Additionally the binding site can increase its volume almost 2-fold by displacing the side chains of residues with double conformations in the apo SmeT structure namely Phe70, Met113 and Leu114. The different side chains that can be exposed, or covered, may also facilitate interactions with a wide range of compounds. Remarkably, some residues displaced by ligand binding (Met110 and Met113) belong to helix α6, which interacts through hydrogen bonds with residues from helix α1. These interactions are described as decisive in QacR and TetR deactivation due to ligand binding [19]. Through these interaction networks between helices, the QacR and TetR changes the relative orientation of the N-terminal domains to generate a ligand bound structure unable to bind DNA. The present data suggest that SmeT may modify its shape after triclosan binding, provoking the release of DNA by a similar mechanism (Figure 9).
The structure of the complex SmeT-triclosan, together with the experimental data supporting its role in smeDEF pump derepression, is the first structural evidence that triclosan can act as an effector of an MDR system regulator, and provides an unambiguous link between the presence of this biocide and the increased efflux of antibiotics by the opportunistic pathogen S. maltophilia.
SmeT protein was expressed and purified as previously described [53]. Briefly, the smeT gene was cloned into the pTYB1 vector (IMPACT-CN system, New England Biolabs) and the protein expressed in the E. coli strain ER2566 (IMPACT-CN system, NEB). After induction with 0.5 mM IPTG, the culture was grown overnight at 15°C. Cells were disrupted by sonication and centrifuged. The supernatant with the SmeT-Intein-CBD fusion protein was loaded into a Poly-Prep chromatography column (Bio-Rad) containing chitin beads (NEB). After overnight incubation in cleavage buffer (20 mM Tris-HCl, 0.5 M NaCl, 100 mM dithiothreitol (DTT), 1 mM EDTA, pH 8.0) at 4°C, SmeT was eluted and the remaining DTT present in the sample removed by dialysis against Tris-buffered saline. Finally, the sample was loaded into a Sephacryl S100 gel filtration column (GE Healthcare) in a buffer containing 300 mM NaCl and 20 mM Tris pH 8.0.
Prior to concentration, the sample was incubated with triclosan (5-chloro-2-(2,4-dichlorophenoxy) phenol), previously dissolved in DMSO, for 4 h at 4°C in a 1∶10 molar ratio. The SmeT-triclosan complex was concentrated to 5 mg/ml and crystallized using the sitting-drop vapor diffusion technique in a solution containing 175 mM Li2SO4, 100 mM Tris pH 8.5 and 27% v/v PEG MME 2K. Crystals grew to approximately 500×200×20 µm within 3 days at 22°C. These crystals were transferred into a cryoprotectant solution (175 mM Li2SO4, 100 mM Tris pH 8.5 and 30% v/v PEG MME 2K) and flash-cooled at 100°K.
Crystallographic data were collected at the Beamline ID23-2 of the European Synchrotron Radiation Facility (ESRF) and processed using Mosflm [71] and SCALA software [72]. The crystal parameters were equivalent to those of the apo SmeT structure [53]. The crystals belonged to the space group P21 and had unit cell dimensions of a = 56.5, b = 59.5, c = 84.8 Å, and β = 102.7°. The Mathews coefficient for a dimer in the asymmetric unit was 2.82 Å3/Da, which corresponds to a solvent content of 56.4%. Details of data collection, crystal parameters and data-processing statistics are given in Table 1.
Difference Fourier techniques were used to solve the structure since unit cell dimensions are almost identical to those of the SmeT structure (PDB code 2w53 [53]). Molecular Replacement using MOLREP [73] has confirmed that the crystal packing of the SmeT-TCL complex is identical to that of the SmeT protein. The translation-libration-screw (TLS) groups (3 per chain) were defined using the TLSMD server [74]. Iterative cycles of manual building and TLS-restrained refinement cycles were performed using COOT [75] and REFMAC software [76] to final Rfree and Rcryst values of 25 and 20% respectively (Table 1). The final model comprised residues 6 to 218 of chain A, 10 to 218 of chain B, 225 water molecules, 2 sulphate anions and 2 triclosan molecules. The overall electron density map was of high quality but of poor definition for residues 30–31 and 120–121 in chain B, perhaps due to the high flexibility of these regions. Analysis of the geometry for the final model was performed using Molprobity software [77]; 98.3% of the residues fell within the favored regions of the Ramachandran plot and none within the disallowed regions. The anisotropy of the atomic displacement parameters was analyzed using the PARVATI server [78]; the mean anisotropy for the dimer was 0.46±0.19 (0.43±0.19 for chain A and 0.48±0.19 for chain B). Protein surfaces were analyzed using the PISA server [79]. Cavity volumes were determined using Casp software [80]. Figures were produced using PyMOL software [81].
The dissociation constant for SmeT was obtained from the change in the intrinsic fluorescence of the protein upon triclosan binding. Mixtures consisting of 0.5 µM SmeT with various concentrations (7–1000 nM) of triclosan were prepared in 50 mM NaCl, 20 mM Tris pH 8.0 and 5% (v/v) ethanol. Fluorescence emission spectra were recorded at room temperature in a QuantaMaster QM-2000-7 model spectrofluorometer (Photon Technology International) using a 1-cm-pathlength quartz cell (Hellma). The excitation monochromator was set at 280 nm and emission was recorded between 310 and 360 nm. Three independent titration curves were carried out. The data were analyzed using a non-linear least-squares fit assuming an stoichiometry of two molecules of triclosan per SmeT dimer.
The γ-[32P] dATP labeling of the 30-bp oligonucleotide 5′ GTTTACAAACAAACAAGCATGTATGTATAT 3′ that contained the operator site, and the SmeT purification, were performed as described in earlier work [53]. EMSA assays with or without triclosan were performed by incubating the 5′ end-labeled 30 bp double stranded DNA (2 nM, 10000 cpm) with 0.2 µM SmeT for 20 min at room temperature. The binding buffer used was 10 mM Tris-HCl, 50 mM KCl, 10 mM MgCl2, 1 mM EDTA, pH 7.2, 50 µg/ml bovine serum albumin, 1 mM dithiothreitol, 5% (v/v) glycerol, and 100 µg/ml poly(dI-dC) as nonspecific competitor DNA. Increasing concentrations (0.1 mM and 0.2 mM) of the biocide were then added and incubated at room temperature for 15 min more. Retarded complexes were separated on a 6% non-denaturing polyacrylamide gel (40∶1 acrylamide∶bisacrylamide). The electrophoretic conditions were: 100 V for 90 min at 4°C. The buffer used was TE (89 mM Tris-borate, 2 mM EDTA pH 8.0). Gels were dried before autoradiography.
15 µl of overnight cultures of strains S. maltophilia D457 or S. maltophilia D457R were used to inoculate flasks containing 15 ml LB broth with or without sub-inhibitory concentrations of the biocide (15 µg/ml). When the OD600≈1.0, cells were spun down at 6000× g for 10 min at 4°C and immediately frozen on dry ice and stored at −80°C. Total RNA was extracted from the pellets using the RNeasy Mini Kit (QIAGEN) according to the manufacturer's instructions. TURBO DNA-free (Ambion) was used to eliminate any remaining DNA. RNA integrity was verified on a 1% agarose gel and the absence of DNA confirmed by real time PCR using gyrA(+): 5′CCAGGGTAACTTCGGTTCGA3′ and gyrA(−): 5′GCCTCGGTGTATCGCATTG3′ primers. cDNA was obtained from 1 µg RNA using the High Capacity cDNA Reverse Transcription Kit (AB Applied Biosystems). Real time RT-PCR was performed as described elsewhere [82]. Briefly, a first denaturation step was allowed at 95°C for 10 min followed by 40 cycles (95°C for 15 s, 60°C for 1 min) for amplification and quantification. Primers were designed using Primer Express 3.0 software (AB Applied Biosystems) with the default settings. RT-smeC(+): 5′TCACTGGATGCCTCGAAGATT3′ and RT-smeC(−): 5′CAGGGCATCGGCCACTT3′ amplify a 93 bp fragment of the smeC gene. RT-smeD(+): 5′CGGTCAGCATCCTGATGGA3′ and RT-smeD(−): 5′TCAACGCTGACTTCGGAGAACT3′ amplify a 76 bp fragment of the smeD gene. RT-smeJ(+): 5′TCGAACGCGCCTGAGTATC3′ and RT-smeJ(−): 5′CGCTTTCGTACTGTGCCACTT3′ amplify a 96bp fragment of the smeJ gene. RT-smeY(+): 5′AGCTGCTGTTCTCCGGTATCA3′ and RT-smeY(−): 5′CACCAGGATGCGCAGGAT3′ amplify a 65 bp fragment of the smeY gene. RT-gyrA(+) and RT-gyrA(−) were used to amplify a 60 bp fragment of the housekeeping gene gyrA [83]. Differences in the relative amounts of the mRNA for the smeC, smeD, smeJ and smeY genes were determined using the 2−ΔΔCt method [84]. RNA samples were extracted in three different experiments; the results are the mean values.
A diluted S. maltophilia D457 overnight culture (3∶100,000) was poured onto agar Mueller Hinton plates. Twenty minutes later, ciprofloxacin Etest strips (AB BioMérieux) were added. For the antagonism assays a square of dried Whatman paper previously soaked with 30 µl of triclosan at a concentration of 1 mg/ml, or in ethanol (control), was placed just below the point of the Etest strip corresponding to the minimal inhibitory concentration (MIC) of ciprofloxacin. Bacteria were grown for 24 h at 37°C.
Experiments were performed in microtitre 96-well plates. Briefly, a diluted (1∶1000) S. maltophilia D457 overnight culture was split into two aliquots; one was supplemented with triclosan to obtain a final concentration of 3 µg/ml, the other was used as control. 198 µl of bacteria, with or without triclosan, were loaded per well, and 2 µl of antibiotic were added. The final antibiotic concentrations were: nalidixic acid 6 µg/ml, ofloxacin 1 µg/ml, norfloxacin 6 µg/ml and ciprofloxacin 1 µg/ml. Bacteria were grown at 37°C for 30 h and the OD595 nm measured every 20 min using a Tecan Infinite 200 plate reader. Growth curves were plotted using Microsoft Excel.
PDB accession code 3P9T. The sequence of the smeT gene and the smeT-smeD intergenic region is deposited at the EMBL nucleotide sequence database under accession number AJ316010. The nucleotide sequence of smeDEF is deposited at the EMBL database under accession number AJ252200.
|
10.1371/journal.pntd.0005964 | The clinical and microbiological characteristics of enteric fever in Cambodia, 2008-2015 | Enteric fever remains a major public health problem in low resource settings and antibiotic resistance is increasing. In Asia, an increasing proportion of infections is caused by Salmonella enterica serovar Paratyphi A, which for a long time was assumed to cause a milder clinical syndrome compared to Salmonella enterica serovar Typhi.
A retrospective chart review study was conducted of 254 unique cases of blood culture confirmed enteric fever who presented at a referral adult hospital in Phnom Penh, Cambodia between 2008 and 2015. Demographic, clinical and laboratory data were collected from clinical charts and antibiotic susceptibility testing was performed. Whole genome sequence analysis was performed on a subset of 121 isolates.
One-hundred-and-ninety unique patients were diagnosed with Salmonella Paratyphi A and 64 with Salmonella Typhi. In the period 2008–2012, Salmonella Paratyphi A comprised 25.5% of 47 enteric fever cases compared to 86.0% of 207 cases during 2013–2015. Presenting symptoms were identical for both serovars but higher median leukocyte counts (6.8 x 109/L vs. 6.3 x 109/L; p = 0.035) and C-reactive protein (CRP) values (47.0 mg/L vs. 36 mg/L; p = 0.034) were observed for Salmonella Typhi infections. All but one of the Salmonella Typhi isolates belonged to haplotype H58 associated with multidrug resistance (MDR) (i.e. resistance to ampicillin, chloramphenicol and co-trimoxazole).;42.9% actually displayed MDR compared to none of the Salmonella Paratyphi A isolates. Decreased ciprofloxacin susceptibility (DCS) was observed in 96.9% (62/64) of Salmonella Typhi isolates versus 11.5% (21/183) of Salmonella Paratyphi A isolates (all but one from 2015). All isolates were susceptible to azithromycin and ceftriaxone.
In Phnom Penh, Cambodia, Salmonella Paratyphi A now causes the majority of enteric fever cases and decreased susceptibility against ciprofloxacin is increasing. Overall, Salmonella Typhi was significantly more associated with MDR and DCS compared to Salmonella Paratyphi A.
| Enteric fever is a bloodstream infection caused by the bacteria Salmonella Typhi or Salmonella Paratyphi A, B, or C. It is common in low resource settings and linked to poor water quality and sanitation. The disease is also endemic in Cambodia and since 2013 there has been a sharp increase in the number of Salmonella Paratyphi A infections. We sought to compare the clinical phenotypes and antibiotic susceptibilities of Salmonella Paratyphi A infections with those of Salmonella Typhi infections in this setting. We retrospectively collected demographic, clinical and laboratory data from clinical charts of 254 patients with a blood culture positive for enteric fever. We also assessed antibiotic susceptibility patterns and sequenced the genome of a subset of isolates. We found that since 2013 the majority of enteric fever cases are caused by Salmonella Paratyphi A which increasingly shows decreased susceptibility to the antibiotic ciprofloxacin, the current first line treatment. In contrast, in Salmonella Typhi a re-emergence of susceptibility for the former first line antibiotics of ampicillin, co-trimoxazole and chloramphenicol was observed. Presenting symptoms of Salmonella Typhi and Salmonella Paratyphi A were identical, minor differences were observed in laboratory parameters.
| Salmonella enterica serovar Typhi (Salmonella Typhi) and Salmonella enterica serovars Paratyphi (Salmonella Paratyphi) A, B, and C are Gram-negative bacteria which can invade the bloodstream and cause typhoid and paratyphoid fever respectively (also jointly known as ‘enteric fever’). They are confined to the human host and are transmitted via the fecal-oral route. Enteric fever poses a serious disease burden in low resource settings where the infection is linked to poor sanitation and limited access to safe drinking water [1]. Although enteric fever has become rare in Western countries it continues to affect international travelers returning from endemic countries [2].
Patients with enteric fever typically present with acute fever and non-specific symptoms. For a long time, Salmonella Paratyphi A was thought to cause milder disease than Salmonella Typhi but several studies have contradicted this [1–4].
For both serovars, antibiotic resistance is increasingly reported and there is now widespread presence of co-resistance against the former first line treatment options of ampicillin, co-trimoxazole and chloramphenicol (known as ‘multidrug resistance’) and decreased susceptibility to ciprofloxacin, the current first line drug [5, 6]. Resistance to ciprofloxacin is also increasingly reported [7, 8]. Worrisome are recent reports on emerging resistance against third-generation cephalosporins and azithromycin, the current alternative treatment options [9, 10].
Although historically the majority of enteric fever cases were caused by Salmonella Typhi, the proportion of Salmonella Paratyphi A infections has been increasing steadily since the turn of the century, in particular on the Asian continent [11].
In 2013, a significant increase in Salmonella Paratyphi A infections was also observed in Cambodia, a country where enteric fever remains one of the most common clinical and blood culture-confirmed diseases [12]. The increase was described in local residents as well as in travelers returning from Cambodia to Europe, New Zealand, Japan and the United States [13–16].
Surprisingly, little is known about the clinical and microbiological characteristics of Salmonella Typhi and Salmonella Paratyphi A infections in Cambodian adults.
This study therefore aims to assess the clinical and microbiological aspects of enteric fever in patients attending an adult hospital in Phnom Penh, Cambodia, during 2008–2015. More specifically it aims to assess differences between infections caused by Salmonella Typhi as compared to Salmonella Paratyphi A.
Sihanouk Hospital Center of HOPE (SHCH) is a 40-bed non-governmental referral hospital for adults in Phnom Penh, Cambodia. Since July 2007, SHCH and the Institute of Tropical Medicine (ITM) in Antwerp, Belgium, have been jointly organising the surveillance of bloodstream infections at this hospital and its associated clinics. For this study, all data collected between 2008–2015 were analyzed. Blood cultures were systematically sampled in all patients presenting at SHCH who were suspected of having sepsis according to the Systemic Inflammatory Response Syndrome (SIRS) criteria [17]. Recently, new definitions and criteria for sepsis have been proposed such as the Sequential [Sepsis-related] Organ Failure Assessment (SOFA) score [18]. Over the 8-year period, 18,927 blood cultures were sampled from mostly, but not exclusively, adults [19]. Of these cultures 1,654 (8,7%) yielded clinically significant organisms.
From all patients whose blood was drawn for culture, basic demographic and clinical data were registered in a surveillance logbook. A medical doctor verified missing data with patients during a routine phone call one week after discharge from the hospital which was part of standard care. In addition, for this study, the available medical charts of all patients with blood culture-confirmed enteric fever were reviewed retrospectively by a second medical doctor for additional symptoms and signs.
Hematology parameters were analyzed using a Sysmex KX-21 and T-1800i analyzer (Sysmex Corporation, Kobe, Japan) and CRP values were measured using a TEMIS Linear Analyzer (Linear Chemicals sl, Montgat, Spain).
Blood cultures were sampled and worked-up as previously described [20]. Isolates biochemically identified as Salmonella spp. were stored at -70°C on porous beads in cryopreservative (Microbank, Pro-Lab Diagnostics, Richmond Hill, Canada) and eventually shipped to the ITM in Belgium. At ITM, the isolates were serotyped using commercial antisera (Sifin, Berlin, Germany) following the White-Kauffmann-Le Minor scheme. A selection of 91 isolates were sent to the Institut Pasteur in Paris for confirmation and whole genome sequencing.
At the ITM, antibiotic susceptibility was determined for all available isolates by disk diffusion on Mueller-Hinton II agar in accordance with the CLSI 2016 guidelines [21]. The following antimicrobial drugs (Neo-Sensitabs, Rosco, Taastrup, Denmark) were tested: ampicillin, sulfamethoxazole-trimethoprim, chloramphenicol, nalidixic acid, pefloxacin, gentamicin, tetracycline, ceftriaxone, ceftazidime, meropenem and ertapenem. Nalidixic acid and pefloxacin served as predictors for ciprofloxacin non-susceptibility.
In addition, for all available isolates, minimal inhibitory concentration (MIC) values for ciprofloxacin and azithromycin were determined by the E-test macro method (bioMérieux, Marcy L'Etoile, France).
Quality control was performed using Escherichia coli (ATCC 25922) and Staphylococcus aureus (ATCC 29213).
Multidrug resistance (MDR) was defined as co-resistance to ampicillin, chloramphenicol and trimethoprim-sulfamethoxazole [22]. For comparison with previously published literature, we used the superseded term ‘decreased ciprofloxacin susceptibility (DCS)’, defined as MIC-values of ≥0.12 mg/L and ≤0.5 mg/L, i.e. currently classified as ‘intermediate susceptibility' but associated with treatment failures or delayed treatment response [21].
Whole genome sequencing was carried out on all 65 Salmonella Typhi isolates and a selection of 26 Salmonella Paratyphi A isolates at the Plateforme de microbiologie mutualisée (P2M) from the Pasteur International Bioresources network (PIBnet, Institut Pasteur, Paris, France). Short-read sequences from 30 previously published Salmonella Paratyphi A genomes were also included [23]. The run accession numbers and related metadata are detailed in S1 Table. Short-read sequences have been deposited to the European Nucleotide Archive (ENA) (http://www.ebi.ac.uk/ena) (accession number PRJEB19906).
The MagNAPure 96 system (Roche Diagnostics, Indianopolis, IN, USA) was used for DNA extraction, libraries were prepared using the Nextera XT kit (Illumina, San Diego, CA, USA) and sequencing was done with the NextSeq 500 system (Illumina). Read alignment, single nucleotide polymorphism (SNP) detection and maximum-likelihood phylogeny were carried out as described previously [23]. Sequence assembly was performed using SPAdes v. 3.6.0 [24].
Salmonella Typhi isolates were categorized as belonging to haplotype H58 based on the presence of the H58 specific single SNP (T at nucleotide 252 on the gene glpA corresponding to STY2513 from GenBank accession no. AL513382, Salmonella Typhi CT18) [25]. Genotyphi (https://github.com/katholt/genotyphi) was also used to classify Salmonella Typhi [26]. Salmonella Paratyphi A isolates were categorized as belonging to clade C5 (the dominant clade in Cambodia) based on the presence of the C5-specific SNP (G to A at position 2 381 607 within the SPA_RS11495 gene) [23]. The presence of antibiotic resistance genes was determined with ResFinder version 2.1 (https://cge.cbs.dtu.dk/services/ResFinder/) [27] and plasmids with PlasmidFinder version 1.3 (https://cge.cbs.dtu.dk/services/PlasmidFinder/) and pMLST version 1.4 (https://cge.cbs.dtu.dk/services/pMLST-1.4/) [27, 28]. The presence of mutations in the quinolone-resistance determining region of the DNA gyrase and topoisomerase IV genes (gyrA, gyrB, parC and parE) was assessed by the visual examination of sequences.
Demographic, clinical and microbiological data were entered encoded into an Excel database that was created for this study (Microsoft, Redmond, WA, USA). The code referring to the patient identity was only known by two medical doctors. Access to the database was restricted to these two medical doctors and patient identifiers were removed prior to analysis.
Only the first isolate and associated clinical data for each unique patient was considered. Isolates recovered from a second blood culture drawn within two weeks after the initial one were considered as duplicates, whereas isolates recovered from a repeat blood culture more than two weeks after the initial one were considered as recurrences (either relapse or repeat infections). Statistical analysis was done with Stata 12 (Stata Corp., College Station, TX, USA). Continuous variables are described by a median and interquartile range (IQR). Comparisons between Salmonella Typhi and Salmonella Paratyphi were performed using a Mann-Whitney U test for continuous values and a Chi square test or Fisher exact test for proportions. A p-value of < 0.05 was considered significant.
The study was conducted according to the principles expressed in the Declaration of Helsinki and involves use of information that was previously collected in the course of routine care. Ethical approval for the Microbiological Surveillance Study was granted by the Institutional Review Board of the ITM, the Ethics Committee of Antwerp University, and the National Ethics Committee for Health Research in Cambodia. This study and approval includes retrospective review of demographic and clinical data which are part of routine clinical history taking as recorded in the clinical chart.
Between 1 January 2008 and 31 December 2015 193 Salmonella Paratyphi A isolates were retrieved from 190 patients and 65 Salmonella Typhi isolates from 64 patients. There were no Salmonella Paratyphi B or C isolates; sixty-two non-typhoidal Salmonella isolates were retrieved from 49 patients. The combined annual proportion of Salmonella Typhi and Salmonella Paratyphi A among all clinically significant organisms varied between 2.8% (2008) and 31.7% (2014).
During 2008–2012, enteric fever was caused mostly by Salmonella Typhi (35 cases) and only 12 cases of Salmonella Paratyphi A infection were identified (Fig 1). In 2013 however, a sharp increase in the number of Salmonella Paratyphi A cases was observed with a total of 72 unique cases. In 2014 and 2015, the absolute annual number of Salmonella Paratyphi A cases decreased, but remained higher than for the period preceding 2013. During this period, the number of Salmonella Typhi cases remained relatively stable.
The majority of Salmonella Paratyphi A (64.7%; 123/190) and Salmonella Typhi infections (59.4%; 38/64) cases occurred during the dry season (months November—April) while there was an overall decreasing trend during the rainy season (months June-October) (Fig 2). Compared to the monthly percentage of total blood cultures sampled and clinically significant organisms found, the monthly combined percentage of Salmonella Typhi and Salmonella Paratyphi A was higher during the hot and dry season (March—May) and lower during the rainy season (June-October).
There were four cases of recurrent infections (37–48 days interval between first and recurrent infection), three with Salmonella Paratyphi A and one with Salmonella Typhi; whole genome sequence data was available for three of the four pairs. SNP analysis of the paired isolates revealed that they differed by only two or three SNPs and the isolate pairs formed discrete clusters within the trees (S1 and S2 Figs).
Available epidemiological, clinical and radiographic findings of all enteric fever patients are listed in Table 1.
Eleven out of 189 (5.8%) Salmonella Paratyphi A patients had a known co-morbidity, i.e. HIV (n = 7), Diabetes Mellitus type 1 or 2 (n = 3) and leukemia (n = 1). Of 63 Salmonella Typhi patients also nine (14.3%) had a known co-morbidity, i.e. HIV (n = 7) and Diabetes Mellitus type 1 (n = 2). All but one HIV patient were on antiretroviral therapy at time of presentation. There were 11 patients known to have hepatitis B-positive serology (8 Salmonella Paratyphi A, 3 Salmonella Typhi). At least five could be classified as inactive chronic carriers and two had signs of chronic liver disease on ultrasound.
Patients with a Salmonella Paratyphi A infection were more likely to be living in Phnom Penh compared to Salmonella Typhi patients (81.5% (154/189) vs. 60.3% (38/63); p = 0.001) but the median duration of illness at presentation was the same (four days).
Eleven (17.5%) patients with typhoid fever and 19 (10.1%) with paratyphoid fever were hospitalized, with no statistically significant difference between the two groups (p = 0.119). Reasons for hospitalization included sepsis, persistent fever despite antibiotic therapy, dizziness due to low blood pressure, suspicion of dengue hemorrhagic fever (thrombocytopenia) and dysregulated diabetes mellitus. There were no deaths nor complications noted.
The most frequently reported symptoms in all enteric fever patients together were fever in 229 patients (99.6%), headache in 138 (62.4%) and abdominal pain in 143 (62,2%). Presence or absence of classic enteric fever signs such as a coated tongue and rose spots were infrequently mentioned in clinical files and therefore not evaluated.
Despite the non-specific symptoms, physicians noted typhoid fever in their differential diagnosis upon admission in 67.7% (136/201) of the cases.
There were no statistically significant differences in individual symptoms between typhoid and paratyphoid fever patients, but patients infected with Salmonella Typhi had a slightly but significantly lower median systolic blood pressure (107 mm Hg vs. 113 mm Hg; p = 0.036).
Treatment was not systematically recorded for all patients as many were lost to follow-up. Various antimicrobial regimens were used, but ceftriaxone (2g I.V., once daily) was given most frequently as empirical treatment and as monotherapy, normally for 10–14 days. In case of de-escalation to oral antibiotics, this concerned mostly ciprofloxacin (500 mg, twice daily) and next amoxicillin/clavulanate (625 mg, three times a day). In case of persistent fever while awaiting blood culture results, amikacin was occasionally added to ceftriaxone.
The laboratory parameters of enteric fever patients on admission are summarized in Table 2. Common laboratory abnormalities for enteric fever patients included moderately risen transaminase levels in 133 patients (70.7%), an elevated CRP in 53 patients (94.6%) and eosinopenia in 49 patients (90.7%). Hematological abnormalities were uncommon; the leukocyte count was normal in 88.1% of all patients. Compared to Salmonella Paratyphi A infected patients, Salmonella Typhi patients had slightly but significantly higher median values for leukocytes (6.8 x 109/L vs. 6.3 x 109/L; p = 0.035) and C-reactive protein (CRP) (47.0 mg/L vs. 36 mg/L; p = 0.034), with more presence of leukocytosis (10.0% vs. 2.2% p = 0.015). Salmonella Paratyphi A infection was associated with a higher monocyte count compared to Salmonella Typhi (0.48 x 109/L vs. 0.33 x 109/L), but this difference did not reach statistical significance (p = 0.069).
Both anaerobic and aerobic blood cultures showed signs of growth after a median of two days (IQR 2–3) for all enteric fever patients. In 221 enteric fever patients a pair of one aerobic bottle and one anaerobic bottle was sampled, and in 180 of those cases (81.4%) both bottles grew. In the other cases (growth in only a single bottle), it was the aerobic bottle which grew in nearly two-thirds (65.9%; 27/41) of pairs.
Reported antibiotic exposure in the two weeks before blood culture sampling was not associated with a difference in the median days to growth for both aerobic bottles and anaerobic bottles.
In total, 183 out of 190 (96.3%) unique Salmonella Paratyphi A isolates and all 64 unique Salmonella Typhi isolates recovered during the study period were available for antibiotic susceptibility testing (Table 3).
For Salmonella Typhi, there was a significant decrease (p = <0.001) in the proportion of isolates that were MDR over the 8-year period (62.9% vs. 17.2%) while decreased susceptibility to ciprofloxacin remained at nearly 100% (96.9%; 62/64) during the entire period. For Salmonella Paratyphi A the emergence of DCS was noted as of 2015 (S1 Table). In this year 19 out of 36 unique isolates (52.8%) showed DCS.
Overall, Salmonella Typhi was significantly more likely to be MDR and more likely to display DCS than was Salmonella Paratyphi A (42.2% vs. 0.0%; p = <0.001 and 96.9% vs. 11.5%; p = <0.001 respectively).
Of note, for both serovars no ciprofloxacin resistance was reported and the presence of nalidixic acid and pefloxacin resistance were excellent predictors of DCS except in case of one isolate with a single gyrB mutation (Table 4 and S1 Table). Furthermore, no resistance against third-generation cephalosporins, carbapenems or azithromycin was found.
All but one of the Salmonella Typhi isolates were confirmed to be of the H58 haplotype (recently reclassified as the 4.3.1 genotype), with only 27 out of 63 (42.9%) unique H58 isolates displaying MDR but 98.4% (62/63) displaying DCS. The only non-H58 isolate belonged to genotype 3.2.1 and was pan-susceptible. All of the Salmonella Paratyphi A isolates with DCS belonged to the C5 clade.
Most frequently observed in both Salmonella Typhi and Salmonella Paratyphi A with DCS was the gyrA mutation leading to serine-to-phenylalanine substitution at codon 83 (Ser83Phe) (Table 4). These isolates showed DCS and resistance to pefloxacin and nalidixic acid. There was one Salmonella Typhi isolate with a double gyrA mutation and one with a gyrB mutation. The latter mutation (serine-to-phenylalanine substitution at codon 464 (Ser464Phe)) was associated with intermediate susceptibility to nalidixic acid and DCS. No mutations in ParC or ParE were observed.
Various resistance genes (blaTEM-1B, catA1, sul1, sul2, dfrA7, tet(B), strAB) were detected in MDR Salmonella Typhi which were associated with the presence of an incHI1 PST6 plasmid (S1 Table).
The present study describes the clinical and microbiological aspects of enteric fever in a large group of patients attending an adult hospital in Cambodia with several interesting results.
First, and in line with some other Asian countries, a clear increase was noted in the proportion of Salmonella Paratyphi A infections. This can largely be explained by a community outbreak which occurred in 2013, but the number of cases has remained high also in succeeding years. A recent genetic study on the Cambodian Salmonella Paratyphi A outbreak isolates showed that these isolates belong to a clade that has been circulating in the South-East Asian region already for decades [23]. Further, no indications were found for significant genetic changes within the Cambodian isolates suggesting that environmental and/or behavioral factors are more likely to play a role.
Patients with paratyphoid fever were significantly more likely than typhoid fever patients to be residents of Phnom Penh, which suggests that exposure to the bacterium is more common in the city. Previous studies from Nepal and Indonesia have linked paratyphoid fever to recent immigration into the capital and consumption of street food [29, 30]. Increased dependency on street food has been linked to urbanization, and Phnom Penh is rapidly expanding.
As part of urban expansion, some of the city’s peri-urban lakes have been filled with sand to reclaim land [31]. These lakes are estimated to receive 80% of the city’s (untreated) waste water and act as a natural sewage treatment through aquatic cultivation of vegetables of which some are consumed raw [32]. Reductions in the size of these lakes could have led to higher concentrations of fecal sludge and bacteria in the remaining water and increased flooding in the city [33].
The majority of enteric fever cases occurred in the dry season. During this season more vegetables are harvested and it is also known for an increased availability of snails and clams (bivalve shellfish), due to low water levels in rivers. Shellfish are known to be able to concentrate micro-organisms from water. They are popular snacks which are dried outside rather than boiled during the dry season. In addition, this season coincides with the two most important festivities of the year, the Chinese and Khmer New Year which are associated with increased migration in and out of the city and longer storage duration of food. Last, high daily temperatures may lead to more indiscriminate intake of water and ice cubes. These factors are currently being explored more in-depth.
Second, based on individual symptoms at presentation, infections caused by Salmonella Typhi vs. Salmonella Paratyphi A were similar and indistinguishable, which is in line with other studies from Asia [3, 4]. For both serovars, the median pulse rates at presentation (106 and 105 beats/minute) were high. This has been noted before in children with enteric fever [34].
No complications or deaths occurred which could be ascribed to a prompt start of antibiotic therapy and an early presentation. The latter can also explain the absence of relative bradycardia and the low rate of diarrhea observed which are typically seen in later stages of the disease [35].
In general, laboratory abnormalities were non-specific and leukocyte counts were normal in 87.4% of all enteric fever patients which was also found by others [36]. Higher leukocyte counts and CRP values were found among Salmonella Typhi-infected patients, suggesting a more severe infection. This is in line with a recent human challenge study, which found that a challenge with Salmonella Paratyphi A in healthy volunteers resulted in a milder disease profile (high rates of afebrile bacteremia) than that observed following typhoid challenge [37].
Some differences in presentation were noted when comparing these results to travelers infected with the same Salmonella Paratyphi A C5 strain returning from Cambodia to France. In the latter study, the majority of patients did have diarrhea (70.6%) and were hospitalized (86%) [14]. This difference may be due to other waterborne or oral-fecal infections travelers frequently contract and/or less financial constrains related to hospital admission [2]. Clinical presentation in the present study also differed from typhoid fever patients in African countries where higher rates of severe complications and mortality are observed [38]. As the same H58 haplotype of Salmonella Typhi is dominant in Asia and in eastern Africa, differences could perhaps be explained by timely access to health care and adequate treatment as well as to host-related factors including underlying co-morbidities like malnutrition.
It has been suggested that isolates of the H58 lineage and MDR strains in general are associated with increased virulence and pathogenicity [39–41]. Therefore the results regarding the clinical presentation and severity of cases as described here, might not be applicable to areas where other lineages dominate.
As a third observation, antibiotic resistance trends were very different for the two serovars. While 42.2% of the unique Salmonella Typhi isolates displayed MDR, none of the Salmonella Paratyphi A isolates did. DCS was present in nearly all Salmonella Typhi isolates, but only emerged in Salmonella Paratyphi A from 2015. The rapid increase of DCS in Salmonella Paratyphi during that year is of concern as ciprofloxacin is the treatment of choice for uncomplicated enteric fever.
Fourth, although all but one Salmonella Typhi isolates were found to belong to the globally dominant H58 haplotype, more than half were not associated with MDR and the proportion of (plasmid-mediated) MDR Salmonella Typhi significantly decreased during the study period. This trend has previously been noted in India, Nepal and neighboring country Vietnam [42–44], but contrasts with a recent study on Salmonella Typhi isolates from rural Cambodia where 89% of the H58 isolates displayed the MDR phenotype [45]. The re-emergence of susceptibility might result from a lack of antibiotic pressure since fluoroquinolones have become the preferred treatment both in community and hospital settings in Phnom Penh.
No resistance against ceftriaxone nor against azithromycin was observed. However, reports on extended spectrum beta-lactamases (ESBL) positive and azithromycin resistant Salmonella spp. isolates are emerging globally including one from the same hospital on Salmonella enterica serovar Choleraesuis [20, 46, 47] underlining the importance of continued microbiological surveillance.
Last, molecular analysis of isolates from three patients with a recurrent infection showed relapse was more likely than re-infection with isolate pairs differing only 2–3 SNPs which can occur during the period of persistence within the human body and suggests relapse rather than re-infection [48]. Relapse is estimated to occur in around 5–10% of enteric fever cases usually two to three weeks after the resolution of fever [5]. In our study, a blood culture confirmed recurrence was witnessed only in 4 out of 254 cases (1.6%). It is likely that other recurrent infections have been missed, partly due to the different medical systems that co-exist in Cambodia in which patients readily switch from one healthcare provider to another, especially if symptoms persist.
This study has several limitations. First, the hospital-based setting precluded generalization to patients whose symptoms were not severe enough to seek medical care in a hospital or clinic. Second, the study concerned mostly adults and therefore findings might not be equally applicable to a pediatric population. Third, the study was retrospective in nature; not all clinical charts were available for review and clinical record keeping was variable among different clinicians. It was not possible to reliably estimate time to defervescence.
Despite these limitations, this is one of the largest and most comprehensive descriptive studies on Salmonella Paratyphi A infections so far which is relevant given the global increase in Salmonella Paratyphi A infections. The data do not represent one single hospital, but several clinics located in different districts of the city. Some of these clinics have reduced rates for the poor and during the study period all blood cultures were provided for free. This helped to overcome some of the bias associated with a hospital based study.
The high proportion of Salmonella Typhi and Paratyphi A recovered from blood cultures indicated that enteric fever is a very frequent disease in Phnom Penh. While efforts are made to increase the microbiological diagnostic capacity in the country, a rapid test for invasive Salmonella infections would be a welcome tool for fast and reliable diagnosis. It could increase knowledge on the burden of disease in the community and could replace the flawed Widal test that is still frequently used.
As the current Salmonella Typhi vaccine provides no to very little protection against Salmonella Paratyphi A, the development of an effective Salmonella Paratyphi A vaccine should be promoted, pending improved water quality and sanitation [49].
Enteric fever is frequent in Phnom Penh and the proportion of cases due to Salmonella Paratyphi A has increased. Studies to investigate risk factors and possible transmission routes are urgently needed to advise public health interventions. No MDR was observed for Salmonella Paratyphi A but DCS increased rapidly. DCS remained highly prevalent in Salmonella Typhi while MDR rates have declined. Ceftriaxone and azithromycin remain highly active in vitro but continued surveillance is imperative to monitor resistance.
|
10.1371/journal.pgen.1000108 | Metabolic Actions of Estrogen Receptor Beta (ERβ) are Mediated by a Negative Cross-Talk with PPARγ | Estrogen receptors (ER) are important regulators of metabolic diseases such as obesity and insulin resistance (IR). While ERα seems to have a protective role in such diseases, the function of ERβ is not clear. To characterize the metabolic function of ERβ, we investigated its molecular interaction with a master regulator of insulin signaling/glucose metabolism, the PPARγ, in vitro and in high-fat diet (HFD)-fed ERβ -/- mice (βERKO) mice. Our in vitro experiments showed that ERβ inhibits ligand-mediated PPARγ-transcriptional activity. That resulted in a blockade of PPARγ-induced adipocytic gene expression and in decreased adipogenesis. Overexpression of nuclear coactivators such as SRC1 and TIF2 prevented the ERβ-mediated inhibition of PPARγ activity. Consistent with the in vitro data, we observed increased PPARγ activity in gonadal fat from HFD-fed βERKO mice. In consonance with enhanced PPARγ activation, HFD-fed βERKO mice showed increased body weight gain and fat mass in the presence of improved insulin sensitivity. To directly demonstrate the role of PPARγ in HFD-fed βERKO mice, PPARγ signaling was disrupted by PPARγ antisense oligonucleotide (ASO). Blockade of adipose PPARγ by ASO reversed the phenotype of βERKO mice with an impairment of insulin sensitization and glucose tolerance. Finally, binding of SRC1 and TIF2 to the PPARγ-regulated adiponectin promoter was enhanced in gonadal fat from βERKO mice indicating that the absence of ERβ in adipose tissue results in exaggerated coactivator binding to a PPARγ target promoter. Collectively, our data provide the first evidence that ERβ-deficiency protects against diet-induced IR and glucose intolerance which involves an augmented PPARγ signaling in adipose tissue. Moreover, our data suggest that the coactivators SRC1 and TIF2 are involved in this interaction. Impairment of insulin and glucose metabolism by ERβ may have significant implications for our understanding of hormone receptor-dependent pathophysiology of metabolic diseases, and may be essential for the development of new ERβ-selective agonists.
| In the present study, we demonstrate for the first time a pro-diabetogenic function of the ERβ. Our experiments indicate that ERβ impairs insulin sensitivity and glucose tolerance in mice challenged with a high fat diet (HFD). Loss of ERβ, studied in ERβ -/- mice (βERKO mice), results in increased body weight gain and fat deposition under HFD-treatment. Conversely, absence of ERβ averted accumulation of triglycerides and preserved regular insulin signaling in liver and skeletal muscle. This observation was associated with improved whole-body insulin sensitivity and glucose tolerance. Increased adipose tissue mass in the presence of improved insulin sensitivity and glucose tolerance is usually observed under chronic stimulation of the nuclear hormone receptor PPARγ. In consonance, we show that activation of PPARγ was markedly induced in gonadal fat from βERKO mice and blockade of adipose PPARγ signaling by antisense oligonucleotide injection reversed the metabolic phenotype. Moreover, our cell culture experiments indicate that ERβ is a negative regulator of ligand-induced PPARγ activity in vitro. Finally, we identify SRC1 and TIF2 as key players in the ERβ-PPARγ interaction. In summary, the present study demonstrates that ERβ impairs insulin and glucose metabolism, which may, at least in part, result from a negative cross-talk with adipose PPARγ.
| The estrogen receptors (ERs) are members of the nuclear hormone receptor family (NHR) which act as eukaryotic ligand-dependent transcription factors. ERs are involved in the regulation of embryonic development, homeostasis and reproduction. Two major estrogen receptors, alpha and beta (ERα and ERβ), convey the physiological signaling of estrogens (17β-estradiol, E2) [1]. Additionally, ERs are activated by specific synthetic ligands such as raloxifene, tamoxifen, the ERβ-specific ligand diarylpropionitrile (DPN), and the ERβ-specific agonist propylpyrazole-triol (PPT), which belong to the group of selective estrogen receptor modulators (SERMS) [2]–[4].
The prevalence of metabolic diseases such as obesity, insulin resistance and type 2 diabetes has increased dramatically during the recent ten years [5]. Gender differences in the pathophysiology of obesity and metabolic disorders are well established [6]–[8]. However, the molecular mechanisms of sexual dimorphism in metabolic diseases are largely unknown. In addition, lack of ER activation has been implicated in postmenopausal impairment of glucose and lipid metabolism, resulting in visceral fat distribution, insulin resistance and increased cardiovascular risk after menopause [9]. In this context the investigation of ER-signaling and its role in metabolic disorders has gained increasing attention [4],[8].
To identify the ER subtype involved in the regulation of metabolic disorders, studies have been carried out in ER-deficient mice. ERα-deficient (αERKO) mice have profound insulin resistance and impaired glucose tolerance [10]–[13]. These studies indicate that ERα has a protective role in metabolic disorders by improving insulin sensitivity and glucose tolerance. The metabolic function of ERβ is not clear. ERβ knockout mice (βERKO) have a similar body weight and equal fat distribution in comparison to wild type littermates. Additionally, βERKO and wild-type (wt) mice exhibit similar insulin and lipid levels [14]. However, previous studies in βERKO mice were only carried out under low fat diet, which may have concealed a phenotype relevant for human obesity normally induced by high-energy/fat diet.
The peroxisome proliferator-activated receptor gamma (PPARγ) belongs to the NHR family and is a major regulator of glucose and lipid metabolism by modulating energy homeostasis in adipose tissue, skeletal muscle and liver [15]–[17]. Glitazones or thiazolidinediones (TZDs) are high-affinity PPARγ agonists, and act as insulin sensitizers. TZDs induce adipogenesis and adipose tissue remodeling followed by an improvement of glucose tolerance [18]. The role of PPARγ in the control of glucose homeostasis expands beyond its primary action in adipose tissue, and involves the regulation of adipocytokine production such as adiponectin, leptin, and resistin [19]–[21]. Consistently, reduced PPARγ activity has important metabolic and cardiovascular pathophysiological consequences leading to insulin resistance, diabetes and end organ damage [15].
The molecular mechanisms underlying PPARγ function are similar to those of ER-signaling. In a basal state, PPARγ, similar to ERs, is bound to corepressor proteins such as nuclear receptor corepressor (NCoR) or silencing mediator of retinoic acid and thyroid hormone receptor (SMRT) [22]. After binding within the ligand binding domain (LBD), PPARγ ligands induce its heterodimerization with retinoid x receptor alpha (RXRα), and its subsequent interaction with co-activators like steroid receptor coactivators (SRCs) followed by binding to PPARγ response elements (PPREs) within target gene promoters [23]. Importantly, PPARγ is sharing a similar pool of cofactors with ERβ which provides a platform for mutual interactions between these two NHRs [23],[24].
To study the crosstalk between ERβ and PPARγ, we investigated the regulation of PPARγ-mediated transcriptional activity by ERβ. Our in-vitro experiments in 3T3-L1 preadipocytes showed that ERβ inhibits ligand-mediated PPARγ-transcriptional activity. That resulted in the blockade of PPARγ-induced adipocytic gene expression and in decreased adipogenesis. Overexpression of nuclear coactivators such as steroid receptor coactivator 1 (SRC1) and transcriptional intermediary factor 2 (TIF2) prevented the ERβ-mediated inhibition of PPARγ activity, whereas the presence of vitamin D receptor (VDR)-interacting protein 205 (DRIP205) or PPARγ coactivator-1alpha (PGC1α) had no effect indicating a role for distinct nuclear coactivators for ERβ-PPARγ interaction in-vitro. High fat diet (HFD)-fed βERKO mice showed increased body weight and fat mass. In contrast, triglyceride content in liver and muscle was decreased in βERKO mice, which was associated with a marked improvement of hepatic and muscular insulin signaling. Compared to wt, βERKO mice demonstrated improved systemic insulin sensitivity and glucose tolerance. In consonance with the metabolic phenotype and with the in-vitro data, βERKO mice exhibited augmented PPARγ signaling in adipose tissue corresponding to increased food efficiency and significantly elevated RQ (respiratory quotient). Blockade of adipose PPARγ signaling in βERKO mice by PPARγ antisense oligonucleotide injection resulted in a reversal of the βERKO phenotype including body weight reduction and impairment of insulin sensitivity.
In summary, the present data demonstrate that ERβ impairs insulin and glucose metabolism which may, at least in part, result from a negative cross-talk with adipose PPARγ.
In order to demonstrate a molecular interaction between PPARγ and ERβ in a metabolically relevant cell system, we first investigated ligand-dependent PPARγ activity in the presence of ERβ in 3T3-L1 preadipocytes. Cells were treated with the PPARγ-agonist pioglitazone (10 µM), with or without additional E2 stimulation, and PPARγ activation was measured using pGal4-hPPARγDEF/pG5TkGL3 luciferase assay [25]. Upon pioglitazone stimulation, 3T3-L1 preadipocytes showed pronounced PPARγ activation (bar 1+2, Figure 1A). This activation was not affected by co-treatment with ligands for ERβ such as E2 (bar 2 vs. 3, Figure 1A) or DPN (data not shown). Overexpression of ERβ led to a marked inhibition of ligand-dependent PPARγ activity (bar 2 vs. 4+6+8, Figure 1A) which was also corroborated in a PPARγ response element (PPRE) luciferase assay ( Figure S1). This inhibition was E2 (bar 4+6+8 vs. 5+7+9, Figure 1A) and DPN independent (data not shown). The inhibitory effect of ERβ seemed to be isoform specific, since ERα overexpression resulted in no inhibition of PPARγ activity (bar 11, Figure 1A and Figure S2). To further explore the regulation of PPARγ by ERβ, we performed additional experiments coexpressing an activation function 1 domain (AF-1) deleted-ERβ construct in 3T3-L1 cells. Overexpression of this truncated form of ERβ which still contains a functional ligand binding domain (LBD) did not reduce PPARγ activity indicating that ERβ AF-1 is necessary for regulation of PPARγ by ERβ (bar 10, Figure 1A). To assure adequate overexpression and function of ERβ in our system, 3T3-L1 preadipocytes were transiently transfected with ERβ followed by Western blot analysis and transactivation assays using ER response elements (ERE)-luciferase system (Figure 1B, C). Both assays confirmed adequate expression and function of ERβ.
While our data implicated a negative regulation of ligand-mediated PPARγ transcription by ERβ, we next investigated the regulation of PPARγ-dependent gene expression during 3T3-L1 preadipocyte differentiation. The preadipocytes were transfected with indicated plasmids and differentiated for 3 days using standard differentiation medium [25]. As the full differentiation procedure requires 7-10 days of treatment, the observed effect on fat droplet accumulation and expression pattern are typical for early phase of adipocyte differentiation. The 3T3-L1 cells transfected with ERβ and differentiated for 3 days showed reduced adipogenesis visualized by fat droplet accumulation in comparison to control cells (Figure 2A). Low levels of ERβ could also be detected in untransfected 3T3-L1 cells and its expression was slightly elevated during differentiation (data not shown) underlining the physiological importance of our findings. Overexpression of the ERα isoform in these cells did not show any inhibitory effect on preadipocyte differentiation (Figure 2A). The adipocyte protein 2 (aP2) gene belongs to the classical PPARγ-regulated genes involved in the early phase of adipogenesis [26]. The expression level of aP2 measured by real-time PCR was significantly elevated in the differentiated control cells (bar 2 vs. 1, Figure 2B). Overexpression of ERβ-but not ERα- in these cells led to a significant reduction of aP2 expression (bar 2 vs. 4 and 6, Figure 2B) indicating that endogenous PPARγ activation in 3T3-L1 cells was inhibited by ERβ.
Furthermore pioglitazone (10 µM) treatment of 3T3-L1 cells overexpressing PPARγ/RXRα showed increased adipogenesis, an effect that was markedly inhibited by coexpression of ERβ (Figure 2C). aP2 expression level was also significantly reduced in cells co-expressing ERβ together with PPARγ/RXRα (bar 2 vs. 3, Figure 2D). These data indicate that ERβ inhibits PPARγ-transcriptional activity resulting in the blockade of PPARγ-induced adipocytic target gene expression and amelioration of adipogenesis.
To investigate ERβ's action on PPARγ in vivo, we studied PPARγ activity and PPARγ target genes in HFD-fed βERKO and wt mice. βERKO mice and their wt littermates were fed HFD containing 60% calories from fat for 12 weeks followed by the analysis of PPARγ-dependent gene expression in gonadal fat tissue. Adipose mRNA expression of PPARγ target genes involved in triglycerides (TG) synthesis such as lipoprotein lipase (Lpl), phosphoenolpyruvate carboxykinase (PEPCK) and CD36 was significantly upregulated in βERKO mice (Figure 3 A–C). Key mediators of insulin and glucose metabolism such as the retinol-binding protein 4 (RBP4) were also regulated in βERKO mice (Figure 3D). Consistently with these findings, adiponectin mRNA expression and adiponectin serum levels were elevated in βERKO mice (Figure 3E, F). No difference of PPARγ target gene regulation between βERKO and wt mice was observed in liver (data not shown).
Positive regulation of a series of adipose PPARγ target genes in βERKO mice suggested a general induction of PPARγ transcription in βERKO mice. To prove this, we performed EMSA assays in gonadal fat from βERKO and wt mice after 12 weeks on HFD. Nuclear fractions isolated from adipose tissues from βERKO mice showed an increased binding/activation of endogenous PPARγ in comparison to wt mice (line 4–7 vs. 1–3, Figure 3G) in the presence of similar PPARγ expression levels, as shown by real-time RT-PCR analysis and Western Blot (Figure 3G). Increased adipose PPARγ target gene expression and PPARγ-DNA binding confirmed an augmented PPARγ signaling in adipose tissue from βERKO mice.
To exclude the possibility that the augmented expression of PPARγ target genes measured in HFD-fed βERKO is the result of increased adipose tissue mass, we performed experiments using ex-vivo fat pads isolated from wt and βERKO mice, treated for 24h with 10 µM pioglitazone or vehicle-control, followed by analysis of PPARγ target gene expression using real-time RT-PCR. In this system augmented ligand-induced PPARγ target gene expression mainly results from enhanced PPARγ transcriptional activity and not from increased fat mass. The expression level of PEPCK and Lpl was significantly increased in both wt and βERKO fat pads under pioglitazone treatment (bar 1 vs. 2 and bar 3 vs. 4 Figure 4 A and B). However, pioglitazone-induced PPARγ target gene expression was markedly elevated in βERKO mice compared to wt mice, indicating an augmented PPARγ signaling in the absence of ERβ (bar 2 vs. 4, Figure 4 A and B).
To further characterize ERβ ligand dependency for its interaction with PPARγ in the mouse model, additional in-vivo studies were performed in estrogen-depleted, ovariectomized wt mice treated with the ERβ-ligand DPN. Analysis of PPARγ target genes (Lpl, PEPCK, CD36 and adiponectin) in gonadal fat isolated from these mice revealed no significant differences in the expression level between vehicle and DPN-treated rodents indicating ligand independency (Figure 4C). These data are consistent with the in-vitro study in 3T3-L1 preadipocytes, where PPARγ activation was not affected by co-treatment with ligands for ERβ such as E2 (bar 2 vs. 3, Figure 1A) or DPN (data not shown).
Given the central role of PPARγ in insulin and glucose metabolism, the metabolic phenotype of βERKO mice was assessed. No difference in fasting/fed blood glucose food intake, and mean arterial blood pressure was observed between βERKO and wt mice under HFD (Table 2). Body weight gain was significantly enhanced in βERKO mice, compared to wt mice (mean BW difference βERKO vs. wt mice after 12 week HFD: 3+/−0.4 g, p<0.05, Figure 5A). Increased body weight in βERKO mice resulted from increased adipose tissue mass. MRI-analysis of body composition demonstrated significantly higher fat mass in βERKO mice compared to wt littermates (Figure 5B), and fat pad weight from gonadal and perirenal depots was increased (Table 1). In contrast, liver weight was significantly reduced in βERKO mice in comparison to wt control littermates (Table 1). Reduced hepatic weight likely resulted from decreased TG-accumulation assessed by H/E-staining of liver tissue sections (Figure 5C), and by TG quantification in dried liver tissue (Figure 5D). In accordance with reduced hepatic TG-content, hepatic insulin signaling was improved. After injection of insulin in the portal vein, liver tissue was dissected and proteins were isolated for Western blot analysis. Insulin-stimulated Akt phosphorylation was enhanced in βERKO mice (Figure 5E and Figure S3). In parallel to decreased TG levels in liver, βERKO mice had decreased muscular TG-accumulation under HFD and improved insulin signaling (Figure 5F, G, and Figure S3). Skeletal muscle and liver are the major insulin responsive tissues, and important sites of glucose metabolism in-vivo. An important mechanism of PPARγ-mediated insulin sensitization involves adipose tissue remodeling and trapping of circulating triglycerides (TG) which protects the liver and skeletal muscle against TG overload. Increased adipose tissue mass in βERKO mice may protect these animals against TG-overload in liver and skeletal muscle resulting in an improvement of hepatic and muscular insulin sensitivity.
Next we investigated insulin and glucose metabolism in βERKO and wt mice. Whole body glucose disposal was assessed using an oral glucose tolerance test (OGTT) (Figure 6A). Following an oral glucose challenge βERKO mice on HFD had moderately but significantly improved glucose tolerance compared to HFD-fed wt mice (Figure 6A, B). In addition insulin sensitivity measured by an insulin tolerance test (ITT) was improved in comparison to wt mice (Figure 6C, D). No difference in fasting and fed blood glucose was observed between βERKO and wt mice under HFD (Table 2). Despite an increased fat mass in βERKO mice, systemic insulin sensitivity and glucose tolerance were significantly improved under HFD when compared to wt-control. To further examine the enhanced weight gain and fat deposition in βERKO mice, we performed indirect calorimetry and monitored food consumption. Food intake did not differ between wt-control and βERKO mice (Table 2). However, deletion of ERβ resulted in a marked increase of food efficiency (ratio of weight gain and food intake, Figure 6E). No significant difference in O2 consumption (Figure 6F), energy expenditure (Table 2), or locomotor activity (Table 2) was detected between βERKO and wt mice. Low RQ values have previously been described for rodents under HFD and in diabetes [27]. Both wt and βERKO mice exhibited low RQ values. βERKO mice had a significantly higher RQ when compared to wt-controls which may be indicative for attenuated fatty acid (FA) oxidation promoting fat accumulation (Figure 6G). These data show that βERKO mice are partially protected against HFD induced insulin resistance. Increased fat mass may likely result from increased food efficiency based on reduced oxidative utilization of fat and increased fat storage. The metabolic phenotype of βERKO mice including increased fat mass, reduced hepatic/muscular TG and improved systemic insulin sensitivity exhibits high similarity to augmented PPARγ activation e.g. under thiazolidinedione (TZD) treatment [28],[29].
To directly demonstrate the role of PPARγ in HFD-fed βERKO mice, PPARγ signaling was disrupted by intraperitoneal (i.p.) injection of PPARγ antisense oligonucleotide (ASO). HFD-fed βERKO mice were injected twice a week for 6 weeks with either PPARγ ASO or control oligonucleotides. PPARγ expression was significantly reduced in liver of ASO-treated βERKO mice, similar to previously reported results in apoB/BATless mice (data not shown) [30]. However, suppression of hepatic PPARγ by ASO injection is unlikely to play an important role in our model, since hepatic PPARγ signaling did not differ between wt and βERKO mice, respectively. More importantly, i.p. application of PPARγ ASO in βERKO mice resulted in 63±4.8% (p<0.05) reduction of PPARγ expression in gonadal adipose tissue compared to βERKO mice injected with control oligonucleotides (Figure 7A). Accordingly, expression of the PPARγ target genes Lpl, PEPCK, CD36, and adiponectin was markedly decreased in adipose tissue from PPARγ ASO-injected βERKO mice, and adipocyte diameters were increased (Figure 7A, G). These data corroborate a relevant reduction of adipose PPARγ signaling by ASO intervention. Body weight gain and gonadal fat accumulation in HFD-fed-βERKO mice were significantly attenuated by PPARγ-ASO injection (Figure 7B, C). Finally, blockade of adipose PPARγ by ASO led to reversal of the improved insulin response observed in βERKO mice, and to an impairment of insulin sensitivity and glucose tolerance (Figure 7D–F). Together these data underline the importance of adipose PPARγ signaling for the metabolic phenotype observed in βERKO mice.
Nuclear coactivators such as SRC1 and TIF2 are important mediators of ERβ and PPARγ-induced transcriptional activation. It has previously been shown that competition of distinct nuclear receptor (NR) for coactivator binding results in a negative cross-talk between NRs [31]. To prove whether common coactivators are involved in ERβ-PPARγ interactions, SRC1, TIF2, DRIP205 or PGC1α were co-expressed together with ERβ and ligand induced PPARγ activation was measured.
Overexpression of SRC1 and TIF2 prevented the ERβ-mediated inhibition of PPARγ activity (Figure 8A, B) whereas the presence of DRIP205 (Figure 8C) and PGC1α (Figure S4) had no effect. To demonstrate that SRC1 and TIF2 are also involved in ERβ-PPARγ interaction in-vivo, we performed ChIP experiments with gonadal fat from HFD-fed βERKO and wt mice. The adiponectin promoter was selected as a PPARγ-target promoter. Binding of SRC1 and TIF2 to the adiponectin promoter was enhanced in gonadal fat from βERKO mice (Figure 8D), indicating that the absence of ERβ in adipose tissue results in exaggerated coactivator binding to a PPARγ target promoter. Together these data suggest that the coactivators SRC1 and TIF2 are involved in the negative regulation of PPARγ by ERβ in vitro and in vivo.
The present study demonstrates that ERβ is a negative regulator of ligand-induced PPARγ activity in-vitro. Consequently, data from βERKO mice suggest that ERβ negatively regulates insulin and glucose metabolism which may, at least in part, result from an impairment of regular adipose tissue function based on a negative cross-talk between ERβ and PPARγ. Loss of ERβ resulted in enhanced body weight gain and fat accumulation in HFD-fed mice. However, absence of ERβ prevented hepatic/ muscular triglyceride overload, preserved regular insulin signaling in liver/ skeletal muscle, and improved whole-body insulin sensitivity and glucose tolerance under HFD. This metabolic phenotype strongly suggested augmented PPARγ signaling in mice lacking ERβ. And indeed, PPARγ target genes and PPARγ-DNA binding were markedly induced in gonadal fat from βERKO mice. Along this line, blockade of adipose PPARγ signaling by PPARγ ASO injection reversed the metabolic changes in βERKO mice.
A mutual signaling cross-talk between ERs and PPARγ has been described previously. PPARγ together with its heterodimeric partner RXRα has been shown to suppress ER-induced target gene expression through competitive binding to an ERE site in the vitellogenin A2 promoter [32]. In accordance with a bidirectional interaction, Wang and colleagues demonstrated that ERs are capable of inhibiting ligand-induced PPARγ activation in two different breast cancer cell lines [33]. In contrast to our results, these authors show that basal and agonist-stimulated PPRE-activity is also blocked by ERα. Transcriptional activity of PPARγ differs markedly depending on the cell system and tissues. The highest level of PPARγ-mediated transcription has been described in adipocytes and adipocytic cell lines, where molecular conditions such as cofactor availability seemed to be optimized [34]. Compared to adipocytes, breast cancer cells exhibit low PPARγ expression and activity reflected by a less than 2-fold induction of PPRE-activity after ligand stimulation [33]. The presence of PPARγ suppression by ERα in breast cancer cells might be a result of weak basal PPARγ transcriptional activity in these cells. In contrast, the pronounced activation of the exogenous PPARγ LBD in 3T3-L1 preadipocytes may require more potent inhibitory stimuli which could not be achieved by ERα overexpression in our system.
Suppression of PPARγ-LBD activation by ERβ did not depend on ERβ ligands which is consistent with previous reports [33]. Also our in vivo studies in estrogen-depleted, ovariectomized wt mice treated with the ERβ-ligand DPN indicate that PPARγ-ERβ interaction is ligand independent. More importantly, overexpression of a truncated form of ERβ containing solely the ERβ-LBD/ AF2 domain did not induce any inhibitory effect on PPARγ suggesting an important role of ERβ's NH2-terminal AF1 domain for ERβ-PPARγ interactions. Consistently, activity of the ER-AF1 domain is usually not dependent on ligand activation [35]. Furthermore, Tremblay and coworkers demonstrated that ERβ-AF1 activation involves ligand-independent recruitment of SRC-1, a cofactor involved in ERβ-PPARγ interactions in our study [36]. These data corroborate our observation that PPARγ suppression by ERβ involves the AF1 domain and ligand-independent interactions with the coactivators SRC1 and TIF2. Repression of PPARγ activity through ERβ was reversed by titration of the p160 coactivators, SRC1 and TIF2, suggesting that the suppressive action of ERβ is a result of p160 coactivator interaction with ERβ thereby preventing the binding of PPARγ to the same coactivators. Similar interactions have been described previously for ER interaction with the thyroid receptor [31].
The present study demonstrates for the first time that ERβ impairs insulin sensitivity and glucose tolerance under HFD implicating pro-diabetogenic actions of this receptor. In consonance, we could recently demonstrate that ERβ has a suppressive role on glucose transporter 4 (GLUT4) expression in skeletal muscle [8],[37]. GLUT4 has been identified as the major mediator of insulin-induced glucose uptake in fat and skeletal muscle. In addition, removal of the E2-ERβ signaling by ovariectomy in ERα-deficient mice improved glucose and insulin metabolism supporting the diabetogenic effect of ERβ [12]. Loss of ERβ resulted in a marked augmentation of adipose PPARγ activity in our model indicating that ERβ mediates its metabolic actions by a negative interaction with PPARγ in adipose tissue. This concept is corroborated by a number of observations. HFD-fed βERKO mice exhibited increased adipose tissue mass in the presence of improved insulin sensitivity and glucose tolerance. These metabolic changes are usually observed under chronic PPARγ stimulation [17]. PPARγ has been identified as an essential regulator of whole-body insulin sensitivity. Two major mechanisms have been described: (1) Adipose PPARγ protects non-adipose tissue against excessive lipid overload and maintains normal organ function and insulin responses (liver, skeletal muscle) by preserving regular adipose tissue function, and (2) Adipose PPARγ guarantees a balanced and adequate production of adipocytokine secretion such as adiponectin from adipose tissue, factors which are important mediators of insulin action in peripheral tissues [38]–[40]. Both processes could be observed in βERKO mice. Further support of this notion comes from clinical actions of anti-diabetic PPARγ agonists (TZD) [28],[29]. Activation of PPARγ by TZDs in diabetic patients resembles the phenotype of βERKO mice including improved insulin sensitization and glucose tolerance in the presence of weight gain. We also observed increased food efficiency and changes in nutrient partitioning reflected by an increased RQ in βERKO mice. Loss of ERβ appears to result in attenuated fatty acid (FA) oxidation which may favor the storage of TGs in adipose tissue and increased fat accumulation, and may provide a possible explanation for the enhanced weight gain. Interestingly, treatment of obese mice with a synthetic PPARγ agonist has been shown to mediate similar changes including an increase in food efficiency and higher RQ values [41]. Finally, blockade of PPARγ signaling in adipose tissue of βERKO mice resulted in a reversal of the metabolic phenotype corroborating the importance of adipose PPARγ in the present model. The observed suppression of hepatic PPARγ activity by ASO injection is unlikely to play a major role since the initial metabolic characterization of untreated βERKO mice under HFD did not reveal any dysregulation of hepatic PPARγ signaling. In summary, the metabolic phenotype of βERKO mice is mediated by an augmented adipose PPARγ action, which implies that in the presence of ERβ, PPARγ activity might be partially suppressed.
The notion, that ERβ-PPARγ crosstalk requires receptor-p160 interaction, was underlined by our observations in WAT from βERKO mice. Binding of SRC1 and TIF2 to the PPARγ-regulated adiponectin promoter in WAT was enhanced in the absence of ERβ. It has recently been demonstrated that p160 coactivators are important regulators of PPARγ transcriptional activity in WAT [42]. In particular, TIF2 has been identified as a nuclear coactivator involved in the adipogenic actions of PPARγ. Future experiments are required to define the functional relevance of TIF2 and SRC1 in our model. So far one may conclude that the metabolic phenotype of HFD-fed βERKO mice is, at least in part, explained by increased adipose PPARγ activity as a result of exaggerated binding of p160 coactivators to PPARγ-regulated target gene promoters. Diabetogenic actions of ERβ are of major significance for the pharmaceutical development of new ERβ-selective agonists intended for use against a multitude of diseases such as rheumatoid arthritis or postmenopausal osteoporosis [43],[44]. Despite the high tissue selectivity of such compounds, and despite the fact that the actions observed in our study were ligand-independent, one has to be aware of the potentially deleterious actions of ERβ on insulin- and glucose metabolism. As a precautionary measure metabolic profiling of new ERβ agonist should be performed.
Collectively, our data provide first evidence that ERβ negatively regulates insulin signaling and glucose metabolism that involves an impairment of regular adipose PPARγ function. Moreover our data suggest that the coactivators SRC1 and TIF2 are involved in this inhibition. In consonance, impairment of insulin and glucose metabolism by ERβ has significant implications for our understanding of hormone receptor-dependent pathophysiology of metabolic diseases, and is essential for the development of new ERβ-selective agonists.
Female estrogen receptor β -/- mice (βERKO) received from J.-A. Gustafsson (Karolinska Institutet, Huddinge, Sweden) and their wt littermates were housed in a temperature controlled (25°C) facility with a 12-h light/dark cycle and genotyped using genomic DNA isolation kit (Invitek) and PCR primers described elsewhere [45]. 4–5 week old mice were fed ad libitum with a high-fat diet (60% kcal from fat, [25]) for 12 weeks. Body weight and food intake were determined throughout the experiment. At start and end of treatment, body composition was determined by nuclear magnetic resonance imaging (Bruker's Minispec MQ10). After 12 weeks' treatment, blood samples were collected from overnight-fasted animals by retroorbital venous puncture under isoflurane anesthesia for analysis of serum adiponectin (mouse-adiponectin ELISA; Linco Research) and glucose (colorimetric glucose test; Cypress Diagnostics). An OGTT using a dose of 2 g/kg body weight (BW) glucose and ITT with intraperitoneally injected 0.5 units/kg BW insulin (Actrapid; Novo Nordisk) were performed. Tail vein blood was used for glucose quantification with a glucometer (Precision Xtra; Abbott). Blood pressure was measured invasively in the abdominal aorta using a solid-state pressure transducer catheter (Micro-Tip 3F; Millar Instruments) under isoflurane anesthesia. Afterwards animals were killed and organs were dissected. For immunohistochemical studies organs were fixed in 4% formalin, embedded in paraffin and stained with Haematoxylin/Eosin (H&E); for RNA, Western blot analysis and measurement of TG content isolated organs were frozen in liquid nitrogen; for EMSA and Chromatin IP assays abdominal fat was stored in ice-cold PBS with proteinase inhibitors (Complete Mini, Roche), and immediately proceeded as described below.
For DPN- treatment, 10 week old female C57BL/6J mice were ovariectomized, and after 1 week recovery set on soy-free diet. Subsequently mice were treated for 21 days with DPN (8 mg/kg) or vehicle administered using subcutaneous pellets (Innovative Research of America). Afterwards animals were killed under isoflurane anesthesia and organs were dissected.
All animal procedures were in accordance with institutional guidelines and were approved.
ASO complementary to murine PPARγ (Gen-BankTM accession number U09138.1), ISIS 141941, 5′-AGTGGTCTTCCATCACGGAG-3′, and ASO control, ISIS 141923, 5′-CCTTCCTGAAGGTTCCTCC-3′ was generously provided by ISIS Pharmaceuticals (Carlsbad, CA, U.S.A.). Both ASO's were injected intraperitoneally twice a week into 6 week-old female βERKO mice (n = 7 per group). Injections were continued over 6 weeks at a dose of 100 mg/kg/week as described previously [30]. At the end of the experiment animals were metabolically phenotyped as described above.
After HFD feeding, βERKO mice and their wt littermates were analyzed for energy expenditure, RQ, and locomotor activity using a custom-made 4-cage calorimetry system (LabMaster; TSE Systems). The instrument consists of a combination of highly sensitive feeding and drinking sensors for automated online measurement. The calorimetry system is an open-circuit system that determines O2 consumption, CO2 production, and RQ. A photobeam-based activity monitoring system detects and records every ambulatory movement, including rearing and climbing movements, in every cage. All the parameters can be measured continuously. Mice (n = 7 per group) were placed in the calorimetry system cages for 24h.
Tissue samples from gonadal fat were prepared from female wt and βERKO mice. Explanted gonadal fat samples were washed 3 times with ice-cold Hanks Balanced Salt Solution (HBSS) and treated for 24h with 10 µM pioglitazone or vehicle in Dulbecco's modified Eagle's medium F2 (DMEM:F12, Invitrogen). Afterwards tissue samples were washed with ice-cold PBS and RNA extraction was performed using trizol (Invitrogen).
3T3-L1 preadipocytes were purchased from the American Type Culture Collection. Preadipocytes were cultured in Dulbecco's modified Eagle's medium with 10% Fetal Bovine Serum (FBS) and 1% Pen-Strep (Invitrogen). For differentiation experiments preadipocytes were grown to confluence and after 12h culture medium was supplemented with methylisobutylxanthine (0.5 mM), dexamethasone (0.25 µM), and insulin (1 µg/ml) in DMEM containing 10% FBS for 72h [25]. Afterwards cells were washed with ice-cold PBS and RNA extraction was performed using trizol (Invitrogen) according to the manufacturer's instructions. For the staining procedure differentiated cells were washed twice with ice-cold PBS, fixed with 4% PFA, and stained for 1h at room temperature with Oil-red-O solution.
Transient transfection and luciferase assays were performed as previously described [25]. Briefly 3T3-L1 cells were plated in 12-well plates and transfected using lipofectamine 2000 and OptiMEM (Invitrogen) with 100 ng pGal4-hPPARγDEF; 400 ng pG5TkGL3, TIF2-pSG5, DRIP205-pSG5 (kindly provided by B. Staels, Institut Pasteur de Lille, France), 5 ng pRL-CMV, a renilla luciferase reporter vector (Promega), hPPARγ2-pSG5 and hRXRα-pCDNA, pSG5 (Stratagene), hSRC1-pSG5 (kindly provided by M. Parker, Institute of Reproductive and Developmental Biology, Imperial College London, United Kingdom), pERE-TkGL3 (kindly provided by P.J. Kushner, Metabolic Research Unit and Diabetes Center, University of California, San Francisco, USA), hERα-pSG5 and ERβ-pSG5 (kindly provided by P. Chambon, Institut Clinique de la Souris, Illkirch Cedex, France), and PGC1α kindly provided by Addgene, USA. Delta AF1-ERβ-pSG5 was cloned from full length ERβ-pSG5. After 3h of transfection cells were washed, left for 12h in serum deprived medium (0.5% FCS, 1% Pen-Strep), and afterwards treated for 24h with 10 µM pioglitazone (Takeda Pharmaceutical Co. Ltd, Japan) or vehicle (DMSO). When treated with E2 or specific ERβ agonist diarylpropionitrile (DPN), cells were cultivated in phenol red free DMEM and coal-striped FCS. Luciferase activity was measured 36 h after transfection using the dual-luciferase reporter assay system (Promega). Transfection experiments were performed in triplicate and repeated at least three times.
Total RNA from cultured preadipocytes, abdominal fat tissue and skeletal muscle was isolated using trizol (Invitrogen) according to the manufacturer's instructions. For real-time PCR analysis RNA samples were DNAse digested (Invitrogen), reverse transcribed using Superscript (Promega), RNasin (Promega), dNTPs (Invitrogen), according to the manufacturer's instructions, and used in quantitative PCR reactions in the presence of a fluorescent dye (Sybrgreen, BioRad). Relative abundance of mRNA was calculated after normalization to 18S ribosomal RNA. Primer sequences are provided in Table S1. For Western blot detection of ERβ cells were grown on Φ10 cm plates and transfected with increasing amount of ERβ plasmid or empty vector control. After 24h cells were harvested and WB analysis was performed as following: cells (and tissues for Akt analysis) were lysed in RIPA buffer (50 mM Tris pH 7.5, 150 mM NaCl, 5 mM MgCl2, 1% Nonidet P-40, 2.5% glycerol, 1 mM EGTA, 50 mM NaF, 1 mM Na3VO4, 10 mM Na4P2O7, 100 µM phenylmethylsulfonyl fluoride with proteinase inhibitors (Complete Mini, Roche). Lysates (tissues (30 µg) and cells (20 µg)) were analyzed by immunoblotting using antibody raised against ERβ (H-150, Santa Cruz), antibody raised against pS473- Akt and total-Akt (Cell Signalling), and secondary horseradish-conjugated antibodies (Amersham). For PPARγ immunoblotting, 20 µg of nuclear fractions used for EMSA were analyzed using antibody raised against PPARγ (E-8, Santa Cruz). For detection, enhanced chemiluminescent substrate kit (Amersham) was used.
Nuclear extracts were prepared by using a nonionic detergent method as described previously [46]. The inputs were normalized for protein contents, as ERβ-deficient mice have increased fat tissue mass. Detection of PPARγ was performed with a [32P] γATP-labeled PPRE oligo (5′-CAAAACTAGGTCAAAGGTCA-3′ 5′- TGACCTTTGACCTAGTTTTG-3′). The DNA binding reactions were performed with 40 µl of binding buffer (20 µg nuclear extracts, 2 µg of poly(dI-dC), 1 µg of bovine serum albumin (BSA), 5 mM dithiothreitol (DTT), 20 mM HEPES, pH 8.4, 60 mM KCl, and 10% glycerol) for 30 min at 37°C. For competition experiments, a cold oligonucleotide probe was used. The reaction products were analyzed via 5% polyacrylamide gel electrophoresis using 12.5 mM Tris, 12.5 mM boric acid, and 0.25 mM EDTA, pH 8.3. Gels were dried and exposed to Amersham TM film (Amersham Pharmacia Biotech) at −80°C using an intensifying screen.
Abdominal fat tissue (gonadal fat) isolated from wt and βERKO mice was washed in ice-cold PBS with proteinase inhibitors (Complete Mini, Roche), cut into small pieces, and incubated for 12h in 1% formaldehyde, PBS and proteinase inhibitors (Complete Mini, Roche) with rotation at 4°C. Formaldehyde was removed by intensive washing in ice-cold PBS and centrifugation. Samples were lysed in RIPA (with proteinase inhibitors, Complete Mini, Roche), sonicated on ice (Sonopuls HD 2070, 4 times 10s, 100%), and centrifuged. Samples from each group were pooled and protein content of clear phase lysates was measured using a Bradford assay (Amersham). For each immunoprecipitation (IP) 1.5 mg of protein was taken. The volume of the samples was kept constant by using dilution buffer (prepared according to Upstate protocol). For preclearance 90 µl of Protein A Sepharose slurry (Amersham) was added, and the samples were rotated for 1h in 4°C. After centrifugation beads were discarded, and 1% of supernatant volume per aliquot was used as an input control. The residual volume was incubated with 6 µg of appropriate antibodies (anti-Pol II (C-18, Santa Cruz), anti-Flag (Sigma), anti-SRC1 (M-20, Santa Cruz), anti-TIF2 (C-20, Santa Cruz)). The antibody-bound proteins were then precipitated using 300 µl Protein A Sepharose slurry (Amersham), washed and further processed according to the Upstate protocol.
Triglyceride-content in skeletal muscle and liver was measured as described previously [47]. Briefly, tissues were homogenized in liquid nitrogen and treated with ice-cold chloroform/methanol/water mixture (2:1:0.8) for 2 min. After centrifugation the aqueous layer was removed and the chloroform layer was decanted. The mixture was incubated at 70°C for chloroform clearance, and the residues were dissolved in isopropanol, and assessed for the triglyceride content using an enzymatic-calorimetric test (Cypress diagnostics) according to the manufacturer's instructions.
Results from real-time PCR of cell lines, transfections, and animal experiments were analyzed by ANOVA followed by multiple comparison testing or with paired/unpaired t tests, as appropriate. Data are expressed as mean±SEM or as indicated. Results were considered to be statistically significant at p<0.05. |
10.1371/journal.pbio.1001151 | Natural Killer Cell Lytic Granule Secretion Occurs through a Pervasive Actin Network at the Immune Synapse | Accumulation of filamentous actin (F-actin) at the immunological synapse (IS) is a prerequisite for the cytotoxic function of natural killer (NK) cells. Subsequent to reorganization of the actin network, lytic granules polarize to the IS where their contents are secreted directly toward a target cell, providing critical access to host defense. There has been limited investigation into the relationship between the actin network and degranulation. Thus, we have evaluated the actin network and secretion using microscopy techniques that provide unprecedented resolution and/or functional insight. We show that the actin network extends throughout the IS and that degranulation occurs in areas where there is actin, albeit in sub-micron relatively hypodense regions. Therefore we propose that granules reach the plasma membrane in clearances in the network that are appropriately sized to minimally accommodate a granule and allow it to interact with the filaments. Our data support a model whereby lytic granules and the actin network are intimately associated during the secretion process and broadly suggest a mechanism for the secretion of large organelles in the context of a cortical actin barrier.
| The immune system's natural killer cells eliminate diseased cells in the body. They do so by secreting toxic molecules directly towards the diseased cells, so causing their death. This process is essential for the host organism to defend itself against infectious diseases. The interface between the natural killer cell and its target—the lytic immunological synapse—forms by close apposition of the surface membranes of the two cells. It is characterized by coordinated rearrangement of proteins to allow lytic granules, which contain the toxic molecules, to fuse with the cell surface at the synapse. Given the large size of the granules, one challenge the natural killer cell faces is how to contend with network of actin filaments just under the cell surface, which potentially could pose a barrier to secretion. The current model proposes large-scale clearing of actin filaments from the center of the immunological synapse to provide granules access to the synaptic membrane. By using very high-resolution imaging techniques, we now demonstrate that actin filaments are present throughout the synapse and that natural killer cells overcome the actin barrier not by wholesale clearing but by making minimally sufficient conduits in the actin network. This suggests a model in which granules access the surface membrane by means of specific and facilitated contact with the actin cytoskeleton.
| Natural killer (NK) cells are lymphocytes of the innate immune system that function in clearance of tumor and virally infected cells [1]. Elimination of susceptible target cells is tightly regulated and follows ligation of germline-encoded activation receptors [2]. As NK cells do not require receptor gene rearrangement, they are constitutively enabled for cytotoxicity. Thus, NK cell activation must be tightly regulated to ensure that healthy cells remain unharmed. Efficient lysis requires the tight adherent formation between the NK cell and the target cell termed the immunologic synapse (IS). The formation of a mature, cytolytic synapse between an NK cell and a target cell occurs in stages that can be thought of as checkpoints in the activation process [3]–[5]. Major cytoskeletal steps that are required in this process include the rearrangement of filamentous actin (F-actin) and the polarization of the microtubule organizing center (MTOC) [6]–. These events culminate in the directed secretion of lytic granule contents at the IS, which is prerequisite for NK cell cytotoxicity.
F-actin accumulation at the synapse is the first major cytoskeletal reorganization event and is critical to subsequent steps and function of the IS [5]. Inhibiting proper F-actin dynamics in NK cells with the actin targeting drugs cytochalasin [6],[9], latrunculin [10], or jasplakinolide [3] inhibits their cytotoxicity. Furthermore, NK cells from patients with Wiskott-Aldrich Syndrome (WAS) who have mutations in the actin regulatory protein, WAS protein (WASp), are poorly cytotoxic [9]. This defect is attributable to improper reorganization of F-actin at the IS. Additionally, the actin nucleator Arp2/3 complex, which is enabled by WASp, is also required for cytotoxicity [10]. Cytochalasin treatment, Arp2/3 complex depletion, or WASp deficiency prevent the normal accumulation of F-actin at the synapse [5],[9],[10].
One question that arises from the creation of a dense polarized network at the IS is how secretion of lytic granules occurs through a potential barrier. The traditional view of granule delivery through the actin network holds that granules reach the synaptic membrane through a void of actin in the center of the network. This model is based on the observation from 3-D confocal microscopy that actin forms a dense peripheral ring around the IS [5],[11]. There is a caveat to the seemingly unobstructed access to the membrane that this “ring” provides: the actin motor protein, myosin IIA, is required for secretion and, more specifically, for granule delivery to the synaptic plasma membrane [12],[13]. These data are at apparent odds with one another as a requirement for myosin IIA for secretion necessitates a requirement for actin. One explanation is that granules are secreted at the periphery of the synapse where the traditional model depicts the location of F-actin. Another explanation is that the center of the synapse actually contains F-actin but does so at a level that has been undetectable by conventional 3-D confocal microscopy.
Here we use microscopy techniques that provide enhanced sensitivity and resolution over those used previously to investigate the NK cell IS. We show that F-actin is present throughout the synapse and that lytic granules likely navigate and are secreted through the filamentous network by accessing minimally sufficiently sized clearances. These data demonstrate a previously unappreciated distribution of F-actin at the NK cell IS and redefine granule access to the synaptic membrane and functional secretion.
Visualization of the synaptic actin network has relied on 3-D reconstructions of confocal slices [5],[11]. Here, we took advantage of the superior resolution of imaging in the XY plane to investigate the polarization and distribution of actin. First, we evaluated GFP-actin expressing NK-92 cells conjugating with the susceptible and adherent cell line, mel1190. These cells conjugated in a manner that afforded us the ability to image the synapse in the XY plane. GFP-actin was polarized toward the contact site (Figure 1A) and surprisingly displayed a diffuse distribution across the synapse (Figure 1B). This distribution was quantitatively analyzed and confirmed using a radial intensity profiling algorithm, which demonstrated that the intensity throughout the contact site was substantially above the background.
To more directly image the cortical region of the NK cell immunologic synapse, we used total internal reflection fluorescence microscopy (TIRFm), which has the benefit of an improved signal to noise ratio over confocal microscopy and is limited to visualization within the first membrane proximal 100 nm [14]. Cells were activated via crosslinking of NKp30, a natural cytotoxicity receptor whose ligand is expressed on tumor cells [15], and CD18, a member of the heterodimeric integrin lymphocyte function-associated antigen-1 (LFA-1). Both integrin receptor and activation receptor activation are critical for polarized secretion of granule contents [16]. This combination of signals resulted in robust activation, which was demonstrated by degranulation as measured by enzymatic activity of granzyme A in the supernatant (Figure S1). TIRFm imaging of activated NK-92 cells demonstrated a distribution of F-actin throughout the synapse (Figure 1C). Quantitative analysis using radial profile plotting confirmed the presence of F-actin throughout the cell contacts (Figure 1E). To ensure that these findings were not particular to the NK-92 cell line, we activated and imaged freshly isolated ex vivo NK cells. Similar to NK-92 cells, synaptic F-actin in ex vivo NK cells was identified throughout the contact (Figure 1D,F). These results demonstrate that the NK cell synapse is defined by an abundant, diffuse F-actin network.
To evaluate the kinetics of actin accumulation at the activated synapse, NK-92 cells expressing GFP-actin were imaged using TIRFm after contacting an activating surface. Actin accumulated quickly, within 5 min, and was sustained over the period of observation (50 min) (Figure S2A, Video S1). There was an initial paucity of actin at the synapse followed by a rapid filling in, as demonstrated by the separation of peak contact area and mean fluorescence intensity (MFI) of GFP-actin in that region (Figure S2B). The decrease in MFI over time was due to photobleaching as separate imaging of fields at 10 and 40 min did not show MFI differences (unpublished data). Importantly, actin was diffusely accumulated prior to timepoints at which granule contents were detected in the supernatant (Figure S1). Thus, actin was present as a potential barrier to lytic granule access to the plasma membrane.
Because there was abundant actin present at the synapse, we wanted to determine if lytic granules might utilize relative clearances in the actin network to access the synaptic membrane. To address this, GFP-actin expressing cells were loaded with LysoTracker Red dye, which enables tracking of lytic granules and definition of their position relative to actin, and followed in real time after activation. Numerous granules were identified in the synaptic actin network using two-color TIRFm. Although some relative hypodensities were apparent in the synaptic actin network (Figure S3A–C), the LysoTracker labeled granules did not necessarily appear in these relative voids of actin (Figure 2A, Video S2). To quantitatively analyze this observation across all synaptic granules in an NK cell, the actin intensity in the region of the synaptic granule was compared to that of the entire synapse by dividing the MFI of the respective intensity values to produce a ratio measurement. This ratio, when compared to minimum and maximum potential ratios, demonstrated that on average granules approached the membrane in areas of actin (Figure S4A,B). Combining measurements of all granules in the synapse over 1 h from 14 cells defined the mean granule ratio value as 1.0 (Figure S4C). Although there was a range of actin intensities present throughout the synapse as measured by the ratios of minimum and maximum intensity values to the MFI, few granules were present in areas of particularly low or high actin content. Thus, the colocalization of lytic granules with mean actin signal suggested that granules access the synapse in close proximity to the actin network.
The MTOC is known to deliver lytic granules to the immunological synapse in NK cells [17]. To investigate the relationship among granules, the MTOC, and the synaptic actin network, we imaged the synapse using both confocal microscopy and TIRFm. The MTOC was present in the plane of the synapse and granules that were also in the plane of the synapse were present at a mean distance of 2.05 µm from the MTOC (Figure 2B,C,E), a distance consistent with granules that are converged to the MTOC [17]. To note, the MTOC was not present in any distinct clearance of F-actin. TIRFm demonstrated more clearly than confocal microscopy the varying density of the synaptic actin network (Figure 2D). These data suggest that the MTOC delivers granules to the synaptic actin network.
Because there was variability in colocalization between synaptic actin and granules (Figure S4), we considered the possibility that an approximated granule might not necessarily be capable of degranulation. Specifically, we reasoned that granules that ultimately degranulate represent a subpopulation of approximation events. Furthermore, we hypothesized that granules capable of degranulation might be those present within focal actin hypodensities. In order to study this directly, we developed a novel degranulation indicator for use in live cells.
Lysosomal-associated membrane protein 1 (LAMP1, CD107a), which is sorted to lytic granules [18], is routinely used to detect cells that have degranulated by its appearance on the cell surface [19],[20]. Although previous investigations used antibody to LAMP1 to visualize degranulation [21], we adopted a cell-intrinsic approach by targeting a reporter fluorophore to the lytic granules. We fused pHluorin, a pH sensitive mutant of GFP that does not fluoresce at acidic pH [22], to the cytoplasmic tail of LAMP1 (Figure S5A,B) and obtained stable expression in NK-92 cells. As expected with localization of the pHluorin-LAMP1 construct to lytic granules, treatment with concanamycin A (which effectively neutralizes lysosomal pH by inhibiting the vacuolar-type H+ ATPase [23]) resulted in a robust increase in green fluorescence as measured by flow cytometry (Figure 3A). Since degranulation is an activation-induced process, we also treated pHluorin-LAMP1 expressing cells with the phorbol ester, PMA, and calcium ionophore, ionomycin, and found a rapid increase in pHluorin fluorescence, consistent with LAMP1 surface upregulation (Figure 3A). To better define pHluorin-LAMP1 localization to acidic granules, LysoTracker Red loaded pHluorin-LAMP1 expressing cells were studied using TIRFm after activation. Individual LysoTracker Red labeled granules could be identified at the synapse and were observed to undergo a shift from red to green fluorescence (Figure 3B, Video S3). This event is consistent with the granule fusing with the synaptic membrane, releasing its contents, and encountering a pH neutral environment. These data are consistent with lytic granule targeting of pHluorin.
We next used pHluorin-LAMP1 expressing NK-92 cells to address whether granule approximation results in degranulation. LysoTracker Red loaded, pHluorin-LAMP1 expressing cells were imaged over time using TIRFm (Figure S6A, Video S4). There were significantly more approximation events than degranulation events (mean = 31 and 8 per cell, respectively) over 1 h (Figure S6B,C). Thus only a subset of granules that approximate the synaptic membrane result in a degranulation event.
To directly investigate where degranulation occurs relative to the synaptic actin network, we stably coexpressed pHluorin-LAMP1 and mCherry-actin in cells and imaged them following activation using two-color TIRFm. Timelapse imaging demonstrated that degranulation events occurred in areas of at least some actin fluorescence, similar to that which was seen with granule approximations (Figure 3C, Video S5). We quantitatively evaluated the actin intensity at degranulation points by dividing the intensity values of the actin signal at the point of degranulation events by that of the entire cell contact (ratio of MFIs), thus generating a normalized, comparable value. Degranulations were identified in regions of actin that had slightly lower signal than the mean actin signal of the cell footprint (ratio = 0.965) (Figures 3D, S7). To provide an additional measure, the ratio of MFIs for points adjacent to the degranulation and the entire cell contact were calculated. When compared to ratio of MFIs for the degranulation point itself, the ratio of MFIs for adjacent points were significantly higher. This indicated that degranulation events were occurring in areas of locally hypodense actin. We additionally calculated the minimum and maximum potential values of the ratio measurement. The degranulation values were between the minimum and maximum values (Figure S7). This indicated that degranulation events did not occur in areas of minimal nor maximal actin but rather that they occurred in close proximity to the actin network.
To further characterize the local actin network at the point of degranulation in consideration of focal hypodense regions, we quantitatively evaluated actin fluorescence in the entire immediate vicinity of degranulation events. Measurements of actin fluorescence were made along sequential pixel radii emanating from the centroid of individual degranulations extending approximately 1 µm outwards (Figure 3E). The fluorescent intensities of actin as well as that of pHluorin were quantified in concentric circles along these radii (Figure 3F). In general, as the pHluorin signal diminished from a degranulation centroid, the mCherry-actin signal increased suggesting that degranulation occurs in a focal actin hypodensity (Figure 3F). To measure multiple events, the change in intensity between consecutive radiating circles was determined and plotted for all observed degranulation events (Figure 3G). The mean value of the intensity change demonstrated reduced actin intensity at each of the innermost four radii compared to the neighboring outer radius, thus reflecting the example in Figure 3F. These values were significantly different from the baseline value of zero, which would have indicated no change in the actin network. This indicates that in moving from the periphery of the region of the degranulation event to its center the actin intensity decreased to a detectable degree. Thus, degranulation tends to occur in locally hypodense areas of the actin network (i.e., in regions with some but relatively less actin). In the presence of an actin barrier at the plasma membrane, therefore, hypodense regions of actin provide a potentially more accessible route to the synaptic membrane.
Since granules are in contact with at least some actin during degranulation, we next investigated the role of actin dynamics in degranulation. We inhibited actin polymerization and dynamics with drugs that prevent F-actin assembly (latrunculin A and cytochalasin D) or disassembly (jasplakinolide). Inhibitor addition at the time of activation almost completely inhibited degranulation (Figure 4A), a result that is consistent with the previously reported requirement for initial actin reorganization in synapse formation and maturation [5]. In order to avoid inhibiting the initial, requisite actin reorganization (which we observed by 5 min—Figure S2), we treated cells with the inhibitors following 10 min of activation (a point at which minimal degranulation had occurred—Figure S1). Addition of the actin inhibitors after 10 min of activation resulted in an approximately 50% decrease in bulk degranulation (Figure 4A). Interestingly, addition of inhibitors at 20 min had only a marginal effect on degranulation. Thapsigargin, which has the net effect of elevating intracellular calcium levels, was used as a positive control and increased degranulation.
To further evaluate the effects of the inhibitors on NK cells, we imaged F-actin at the synapse following inhibitor treatment. Jasplakinolide had no effect on F-actin presence or distribution; latrunculin A completely depleted the actin network; cytochalasin D had variable effects on cells, with some cells appearing unaffected while others showing a relative depletion of some filaments (Figure 4B, and unpublished data). To quantitatively evaluate these effects, cells were analyzed using radial intensity profile measurements. This analysis demonstrated no major depletion in F-actin following jasplakinolide or cytochalasin D treatment and robust depletion of the actin network following latrunculin A treatment (Figure S8A,B). Collectively, these results indicate that F-actin presence and reorganization immediately prior to the start of degranulation but after large-scale actin accumulation has occurred is critical for subsequent granule release.
To determine if the synaptic actin network was dynamic at the time corresponding to degranulation, GFP-actin expressing cells were imaged by TIRFm and evaluated at both early and late timepoints of activation. Subtle but consistent changes in actin intensity were visualized over the course of imaging (Figure S9A, Video S6). Although most peaks and troughs of intensity did not appear to change over time, some shifts were detected as highlighted by surface plot rendering of intensity values (Figure S9B, Video S6). To quantify this, images from 5-min series at early and late timepoints of activation were evaluated for changing fluorescent intensities by plotting line profiles across the synapse. Temporally sequential line profiles were overlaid together on one graph, which demonstrated a constant trend in the fluorescence over time with imperfect alignment (Figure S9C,D). The variability was observed during both timeframes of activation and signified a dynamic state of actin where small changes in intensity were occurring. Variation was consistent at both 10 and 30 min after activation as defined by the standard deviation of mean intensity for each pixel across a linear profile (Figure S9E). We were able to inhibit actin dynamics to a measurable degree by using jasplakinolide to stabilize filaments and latrunculin A to prevent new filament formation (Figure S9F). Thus at the level of TIRFm, the synaptic actin network was dynamic at both early and late timepoints. This is consistent with the requirement for actin function at times after which synaptic actin accumulation has occurred.
The resolution of fluorescence microscopy is diffraction limited to around 200 nm [24]. While we did detect focal hypodensities in the actin network corresponding with regions of degranulation, it was unclear if these represented true openings among the actin filaments. Thus we pursued super-resolution of the synaptic actin network to surpass the limit of diffraction. Stimulated emission depletion (STED) microscopes that utilize continuous wave (CW) fiber lasers can image with spatial resolution below 60 nm and thus provide the opportunity to investigate synaptic structures with superior resolution [25].
Using CW-STED microscopy, we first imaged Citrine-actin expressing NK-92 cells conjugated to mel1190 cells (Figure 5A). Actin was present throughout the contact and had a distribution that was similar to that seen with our confocal microscopy images (with cells activated on glass and with mel1190 cells) and TIRF images (activated on glass). This leads us to conclude that our method of activating and imaging NK cells on glass does not induce a distribution or architecture of the actin network that is distinct from that seen with actual target cells.
We next asked whether there was an activation-dependent change in the synaptic actin architecture. To this end, we imaged F-actin in cells that had been stimulated with antibody to CD18, which does not induce degranulation, or in combination with antibody to NKp30, which robustly induces degranulation (Figure S1). Cells stimulated through CD18 alone had a dense synaptic network while the addition of antibody to NKp30 resulted in a more diffuse architecture with many observable clearances in the network (Figure 5B,C). To quantify this observation, clearances were identified based on measured granule diameters. Granules were separately identified using STED microscopy and had a range of diameters with a mean value of 333±103 nm (Figure S10A). The majority of granules were in the 250–499 nm range of diameters, with a smaller number in the 500–749 nm range. These two ranges, which represent areas that are minimally sufficient in size to allow granules access to the plasma membrane, as well as the larger 750+ nm range, were used to categorize clearance area. The number of these clearances in cells significantly increased when a degranulation signal (anti-NKp30) was incorporated (Figure 5D–F). To note, there were many clearances in the 250–499 nm range, fewer in the 500–749 nm range, and still fewer to none in the large 750+ nm range upon activation (Figure 5D–F). This trend in the frequency of clearances reflects the frequency of granule diameters. Furthermore, the mean clearance size divided by the mean granule size resulted in a value of 1.48±0.48, demonstrating that on average clearances were only slightly larger than granules (Figure S10B). Thus, full activation for degranulation results in an actin network with many access points that are minimally sufficient in size to accommodate granules.
Having defined the presence of clearances in the synaptic actin network upon activation, we next sought to identify granule localization relative to the network. Thus we simultaneously imaged the actin network by STED microscopy and granules by laser scanning confocal microscopy. Granules within activated cells displayed a range of interaction with the actin network and all that were present at the synapse had at least some (Figure 5G). Colocalization of granules with F-actin was first measured as a percent of the granule area that contained an actin signal. There was 58±31% colocalization at the synapse, compared to 11±18% colocalization at a distance of 0.5–1 µm from the synapse, indicating that the granules were extensively contacting the actin network after they were delivered to the synapse. To further quantitatively analyze these interactions, we measured line profiles of intensity values for F-actin and perforin across the granule (Figure 5H). Granules were localized within minimally sized clearances, in slightly larger clearances, or directly atop filaments. In all cases, there was an association of the granule with actin as defined by granule and actin line profiles intersecting and/or overlapping. Thus granules use both minimally sufficiently sized clearances and only slightly larger clearances to gain access to the plasma membrane, and they do so in direct interaction with the actin network.
To obtain unprecedented, nanometer resolution of actin filaments at the synapse, we used platinum replica electron microscopy. In order to expose the inner surface of cells, corresponding to the synapse, for metal coating, cells were “unroofed” by mechanical removal of the bulk of the cell body with the nucleus. Images of platinum replicas of the activated synapse confirmed our earlier light microscopy data that the F-actin network exists throughout the synapse and contains small granule-sized clearances (Figure 6A,B). Filaments within the network, however, were present in varying densities. Consistent with the requirement of WASp for synaptic maturation and previous studies defining the localization of the Arp2/3 complex to the synapse [9],[10], branched arrays of filaments were detected (Figure S11), although the branching frequency appeared lower, as compared to a typical branched network in lamellipodia, and many long filaments were also present.
In order to determine whether the abundance or distribution of clearances changes over time, the filamentous network at the synapse was evaluated after 10 and 30 min of activation. Quantitative assessment of the actin network in multiple cells defined a similar total contact area and total area occupied by filaments (filament density) between the two timepoints (Figure S12A,B). We also measured the individual groups of clearances in the actin network that were appropriately sized to allow lytic granule passage. We were able to detect clearances that would accommodate a granule at both timepoints (Figures 6C–E, S13). The number of clearances at each of the size thresholds was similar at the two activation timepoints tested. The distance of these clearances from the cell centroid, however, was different between the timepoints. After 30 min of activation the clearances were closer to the cell center than at 10 min (Figure S12C), suggesting that a fine-tuning of the actin network occurs and may be a necessary process in the maturation leading to degranulation. Thus ultra-resolution imaging of the synaptic actin network further defined the presence of minimally sufficiently sized clearances that would allow granule passage and demonstrated their changing position over time.
Our TIRF data suggested, and our STED data demonstrated, that granules approximate the synapse in close association with the actin network. To confirm this observation on the nanometer level we used platinum replica electron microscopy to image “unroofed,” membrane-intact NK cells. Granule-sized organelles could be identified on the intracellular face of the actin network in approximation with filaments as well as intercalated within filament clearances (Figure 7A,B). These data are consistent with a model of granule approximation and degranulation whereby granules transit to the membrane through an interaction with the F-actin network and utilize multiple minimally sufficiently sized clearances instead of a single large opening (Figure 7C).
Actin accumulation defines an early stage in the maturation of the NK cell IS and is required for subsequent cytolytic function. Without proper reorganization of the actin cytoskeleton, lytic granules fail to polarize to the synapse and NK cells display inadequate cytotoxicity. Upon polarization, lytic granules require myosin IIA function to approximate the plasma membrane and have their contents directly secreted. The dependence on an actin motor protein for the secretory process suggests a requirement for the actin network itself. We have defined the distribution of actin at the IS using techniques that provide unprecedented sensitivity and resolution. We have done so both by live cells conjugated to target cells as well as a more flexible model system. We have also shown that lytic granules reach the plasma membrane and are secreted in areas of actin. Furthermore, our data suggest that secretion events likely occur in minimally yet sufficiently sized clearances in the actin network.
While our studies relied upon activating an NK cell line using immobilized antibody we were able to obtain physiologically relevant supporting data. Firstly, NK-92 cells activated in this manner released contents of lytic granules (Figure S1). Secondly, the distribution of actin at the IS in NK-92 cells was similar on immobilized antibody and on flat, living target cells even when using enhanced resolution (Figures 1B,C, S2A, 5A,C, 5G, 6A,B). Thirdly, to show direct ex vivo supporting evidence, the F-actin distribution at the IS of freshly isolated human NK cells was consistent with that seen in NK-92 cells (Figure 1). Thus the actin network we imaged using our model activation method was unlikely to be an artifact. A benefit of our model approach, however, was the ability to utilize platinum replica electron microscopy to image the synapse with nanometer resolution, and then correlate these findings with our fluorescence microscopy data.
The F-actin network in secretory cells was first believed to be a barrier to exocytosis [26]. This hypothesis was supported by data from numerous cell types including neuroendocrine cells (rev. in [27],[28]), neurons (rev. in [29]), platelets [30], and goblet cells [31] among others where loss of cortical F-actin correlated with an increase in secretion. In many of these cell types it has become increasingly appreciated that F-actin also serves as a required facilitator of secretion [27]–[29],[32]. In cells of the immune system, there is evidence for both models of secretion. Mast cells require some degree of F-actin present for agonist-induced secretion as robust depletion of F-actin with latrunculin interfered with normal release [33]. Conversely, F-actin is reported to be a barrier to secretion for neutrophil granule secretion [34].
Investigations with F-actin and granule secretion at the IS in cytotoxic lymphocytes have thus far been limited to T cells. In cytotoxic T lymphocytes, it has been suggested that actin is not a barrier to secretion [35],[36]. This includes the observation that actin is cleared from portions of the IS and granules are delivered to the synaptic membrane by the microtubule organizing center [35]. Because of 3-D confocal microscopy depicting F-actin rings at the NK cell IS [5],[11], the delivery of lytic granules is believed to occur in actin-devoid regions as in T cells. However, the reliance for granule secretion on myosin IIA in NK cells has called this model into question [12]. Specifically, that myosin IIA associates with lytic granules and its function is required for approximation to the synaptic membrane [13] strongly suggests that the mechanism for granule secretion is different in NK cells.
Our data support a role for F-actin as a facilitator of secretion rather than a barrier. Inhibiting actin polymerization with cytochalasin D or latrunculin A after activation resulted in diminished secretion of lytic granule contents rather than an increase. This suggested that actin was not a barrier to secretion, but that its dynamics were required. The latter hypothesis is supported by the results showing a similar inhibitory effect when cells were treated with jasplakinolide. Interestingly, the effect of the actin inhibitors was most pronounced at 10 min following activation and less so at 20 min following activation, and thus defined a critical window of actin reorganization that facilitates degranulation (Figure 4). One explanation for this is that reorganization is required for the bulk of granules to approach the membrane and dock, which is accomplished by 20 min. Regulation of actin reorganization as a requirement for secretion has been proposed in other cell types. In muscle cells the Arp2/3 complex and cofilin function in the required actin reorganization for GLUT4 vesicle exocytosis [37]. In chromaffin cells, Cdc42, neural-WASp, and the Arp2/3 complex are proposed to function together at the plasma membrane to facilitate secretion through generation of new filaments [38].
Our work defines an actin network that is more pervasive at the NK cell IS than previously thought. Although this could serve as a potential barrier, we have identified abundant granule-sized clearances that could function as sufficient access points to the plasma membrane. These could provide functionality by allowing granules to pass between filaments and to simultaneously interact with them, whereby myosin IIA could exert force in squeezing granules between filaments or in post-fusion expulsion of granule contents. This latter possibility has been suggested in chromaffin cells where myosin II function was required for appropriate release of catecholamines [39]. Another consideration is that granules may use clearances smaller than their equatorial area by squeezing through adjacent filaments. Although we cannot rule out this possibility, it is nevertheless a potential mechanism that is consistent with our hypothesis. Thus we propose that degranulation events at the NK cell IS represent a coupled interplay between actin filaments and clearances that presents additional opportunities for regulatory steps important to NK cell cytotoxicity.
NK-92 and GFP-actin expressing NK-92 cell lines were a kind gift from K. Campbell and were maintained in Myelocult (StemCell) media supplemented with 100 U/mL penicillin and streptomycin (Gibco) and 100 U/mL IL-2 (Hoffman-La Roche). mCherry-actin, Citrine-actin, and pHluorin-LAMP1 expressing cells were generated by retroviral transduction of NK-92 cells as described [40]. Briefly, 2–4 µg of plasmid DNA were transfected into the Phoenix packaging line using Fugene (Roche) lipofection reagent. Supernatant was harvested on day 2 post-transfection. NK-92 cells, Polybrene (Sigma), and supernatant were mixed and spun in a well of a 6 well plate at 1,000×g for 90 min at 32°C. Following overnight incubation at 32°C, cells were spun down and resuspended in supplemented Myeolocult. Cells were grown for 3 d prior to the introduction of puromycin (2 µg/mL) (InvivoGen) or hygromycin B (150 µg/mL) (Cellgro). mCherry-actin and Citrine-actin expressing cells were sorted for high expression by the University of Pennsylvania Cell Sorting Facility. Ex vivo NK cells were prepared from concentrated whole blood as described [41].
The pHluorin-LAMP1 retroviral plasmid was generated by BioMeans, Inc. by inserting the sequence for pHluorin (a kind gift from G. Miesenböck) between the signal sequence and the transmembrane domain of IL-2Rα linked to the cytoplasmic tail of LAMP1 (a kind gift from M. Marks). A flexible GS linker was added between pHluorin and the transmembrane domain sequences. The entire construct was subsequently cloned into the MIGR1-puromycin vector.
The mCherry-actin retroviral plasmid was generated by PCR amplifying mCherry-actin from a pmCherry plasmid with 5′ BglII and 3′ EcoRI restriction site overhangs. The PCR product was digested and ligated into the pMSCV-Hygromycin plasmid (a kind gift from W. Pear), which had an EcoRI site in the Hygromycin resistance gene sequence eliminated by site-directed mutagenesis. The Citrine-actin retroviral plasmid was generated by amplifying the Citrine sequence from the pRSET-b Citrine plasmid (a kind gift from R. Tsien) with 5′ and 3′ BglII overhangs. The product was digested and ligated into a similarly digested mCherry-actin retroviral plasmid, effectively removing the mCherry sequence and inserting the Citrine sequence. Proper orientation of insert was verified by DNA sequencing by the Children's Hospital of Philadelphia Research Institute sequencing core facility.
Flow cytometry was performed to verify pHluorin-LAMP1 expression. Cells were untreated or treated with phorbol myristate acetate (PMA, 100 ng/mL, Sigma) and Ionomycin (1 µg/mL, Sigma) for 30 min or Concanamycin A (CMA, 100 nM, Sigma) for 90 min and samples were run on a BD FACSCalibur.
Cells were washed and resuspended in supplemented Myelocult prior to use. For imaging of lytic granules, cells were incubated with 100 nM LysoTracker Red DND-99 (Molecular Probes) for 30 min at 37°C, washed once, and resuspended in supplemented Myelocult. ΔT dishes (Bioptechs) were coated with 5 µg/mL anti-NKp30 (Beckman-Coulter) and 5 µg/mL anti-CD18 (Clone IB4) for 1 h at 37°C, washed with PBS, and prewarmed prior to imaging with 1 mL dye free R10 (dye free RPMI 1640 (Gibco), 10% fetal bovine serum (Atlanta Biologicals), 10 mM HEPES (Gibco), 100 U/mL penicillin and streptomycin, 100 µM MEM nonessential amino acids (Gibco), 1 mM sodium pyruvate (CellGro), and 2 mM L-glutamine (Gibco). 4×105 cells were added to the dishes, which were maintained at 37°C with a heated stage and lid (Bioptechs).
For live cell imaging of actin dynamics following inhibitor treatment, cells were activated as above for 10 min before addition of media containing DMSO or jasplakinolide (1 µM, Calbiochem). Following 5 min of incubation, media containing DMSO or latrunculin A (10 µM, Sigma) was added to the dish. After 5 min of further incubation, cells were imaged.
For fixed cell experiments, 1×105 cells were adhered to No. 1 glass coverslips coated with antibody as described above. Samples were fixed and stained with Alexa Fluor 488 phalloidin or 568 phalloidin (Molecular Probes) as described [41]. For experiments with inhibitor treatments, cells were activated on coverslips for 10 min or 20 min, treated with inhibitors for 5 min, and then fixed and stained. Cytochalasin D (Sigma) was used at 10 µM.
Samples were imaged through a 1.49 NA, oil immersion, 60×, APO N TIRFm objective or a 1.45 NA, oil immersion, 100×, PlanApo TIRFm objective (Olympus) when noted. 488 nm (Spectra-Physics) and 561 nm (Cobalt) diode lasers were launched through a two-line combiner of an LMM5 (Spectral Applied Research) into a rear mounted TIRF illuminator (Olympus) on an Olympus IX-81. Lasers were aligned for total internal reflection prior to each experiment. Images were captured using Volocity (PerkinElmer) to control a C9100 EM-CCD camera (Hamamatsu).
Mel1190 cells were plated into ΔT dishes 1 d prior to use. Cells were stained with CellMask Deep Red (Invitrogen) according to manufacturer's instructions just prior to imaging. GFP-actin expressing NK-92 cells were added to the dishes, which were maintained at 37°C, and imaged for up to 1 h. Cells were imaged through a 63×1.4 NA Plan-APOCHROMAT objective (Zeiss) on a Zeiss Observer.Z1 using a C10600 ORCA-R2 camera (Hamamatsu). The microscope was equipped with a CSU10 spinning disk system (Yokogawa). 491 nm (Cobalt) and 655 nm (CrystaLaser) diode lasers were launched through an LMM5 (Spectral Applied Research).
NK-92 cells were activated on antibody coated glass coverslips as described above for 30 min and then fixed and stained with rabbit anti-human Pericentrin (Abcam), Alexa Fluor 488-phalloidin, and anti-Perforin-Alexa Fluor 647 (Biolegend). The secondary antibody to anti-Pericentrin was a goat anti-mouse Pacific Blue (Molecular Probes). Cells were imaged in three dimensions on a spinning disk confocal Olympus DSU IX-81 microscope.
Mel1190 cells were grown in a monolayer overnight on No. 1.5 glass coverslips. Citrine-actin expressing NK-92 cells (106) were resuspended in media and incubated on Mel1190 targets for 30′ at 37°C. Cells were fixed with 2% paraformaldehyde and mounted with ProLong antifade reagent (Invitrogen). Cells were imaged at the plane of the interface between NK and target cells. Separately, for imaging of granules, Citrine-actin expressing NK-92 cells were immobilized to bound antibody as described above, then fixed, permeabilized, and stained with anti-perforin Alexa Fluor 488 (Biolegend).
For visualization of actin and perforin in NK-92 cells, cells were immobilized to bound antibody as described above, then fixed, permeabilized, stained with Alexa Fluor 488 phalloidin and anti-perforin Alexa Fluor 647, and imaged at the plane of the glass or 0.5–1 µm above it.
All samples were mounted with Prolong anti-fade reagent (Invitrogen). Cells were imaged through a 100×1.4 NA HCX APO objective on a Leica TCS STED CW system controlled by Leica AS AF software. Alexa Fluor 488 and Citrine were excited using a 488 nm Argon laser and STED depletion was achieved using a 592 nm continuous wave fiber laser. Alexa Flour 647 was excited using a HeNe 633 laser and imaged using the laser scanning confocal modality of the system. Fluorescence was detected with HyD detectors (Leica).
Cells were washed, resuspended in supplemented Myelocult, and allowed to adhere to 0.15 glass coverslips coated with antibody as described above. For imaging of the actin network alone, samples were prepared following a modified protocol described in [42]. After a period of incubation at 37°C, samples were washed once in PBS, dipped into a 1∶3 dilution of PEM buffer (0.1 M PIPES (Sigma), pH 6.9, 1 mM EGTA (Sigma), 1 mM MgCl2 (Sigma)) in dH20 for 10 s, sonicated in PEM for 1–2 s at a 45° angle to the probe, and incubated with 1% Triton X-100 (Sigma) in PEM for 1–2 min before fixation in 2% glutaraldehyde (Fluka) in PEM. For imaging of granules, cells were “unroofed” by applying a nitrocellulose membrane that had been wet in PEM to the coverslip for 45–60 s, removing it, and then fixing the sample in 2% glutaraldehyde in PEM. All samples were processed for EM as described [43]. Platinum replicas were imaged on a JEM 1011 transmission electron microscopy (JEOL USA) at 100 kV. Images were captured using an ORIUS 835.10W CCD camera (Gatan) and are presented in inverted contrast.
Immulon 4HBX 96 well flat bottom plates (Thermo) were coated with murine IgG (BD), anti-human NKp30, anti-human CD18, or both anti-NKp30 and anti-CD18 at 5 µg/mL in PBS overnight at 4°C. Plates were washed twice with PBS and blocked for 1 h at room temperature with R10 media. 1×105 cells were added to wells and plates were spun at 1,000 rpm for 2 min before incubation at 37°C. For the timecourse degranulation assay, supernatants from spontaneous and activated wells were harvested at indicated times. For inhibitor treatments, media containing inhibitor or vehicle (DMSO) was added at indicated times. Thapsigargin (Calbiochem) was used at 1 µM. Supernatants were harvested after 60 min. For total release, cells were lysed in 0.5% Nonidet P40 (Accurate Chemical and Scientific).
Supernatants were assayed by mixing 20 µL supernatant with 200 µL of a solution containing PBS, 9.8 mM HEPES (Gibco), 196 µM Z-L-Lys-SBzl hydrochloride (BLT, Sigma), and 218 µM 5,5′-dithiobis(2-nitrobenzoic acid) (DTNB, Sigma). Samples were incubated for 30 min at 37°C and absorbance was measured immediately at 405 nm. Percent total release was measured by subtracting the spontaneous release value from activated release values (A–S) and the total release value (T–S), and then dividing (A–S) by (T–S).
Images and timelapse series were analyzed using either Volocity or the FIJI package of ImageJ (http://pacific.mpi-cbg.de). Using Volocity, fluorescently tagged actin footprints were identified using a classifier that identifies objects above a selected standard deviation above the mean intensity (usually 0–1). LysoTracker and pHluorin positive events were similarly identified (4–10 standard deviations above mean intensity), with the additional exclusion of events smaller than 0.05 µm2. For actin fluorescence ratio measurements, the MFI at the location of granule approximation or degranulation was divided by the MFI of the actin footprint. Maximum and minimum potential actin intensities were also measured by dividing the maximum and minimum pixel intensities of the actin footprint by the MFI of the footprint. For the “adjacent points” measurement, the region of interest (ROI) that was identified for the degranulation (pHluorin) signal was moved to four neighboring locations. The actin intensities were measured, averaged, and then divided by the MFI of the actin footprint to generate a single value.
Distance of granules from the MTOC was determined as described [17] by inputting the coordinates of the MTOC centroid and a granule centroid into the Pythagorean equation (a2+b2 = c2: (Xcentroid−Xclearance)2+(Ycentroid−Yclearance)2 = c2, where c is the distance from the MTOC).
FIJI was used to generate radial intensity profiles using the radial profile plugin (http://rsbweb.nih.gov/ij/plugins/radial-profile.html). For a profile of the entire cell footprint an ROI circle was drawn around the cell and data from running the plugin were exported to Excel (Microsoft). For profiles of degranulation events images were first split into red and green images and then a circle with a radius of 8 pixels (1.08 µm) was drawn around each event. Data were generated for 6 radii from the center of each event. To determine the radial intensity change, an outer radial intensity value was subtracted from an inner radial intensity value. Thus a negative value indicates that the outer circle has higher mean intensity than the inner circle.
To measure changes in the actin network over time, sequential images were imported into FIJI, a line was drawn across the center of the cell, and line intensity profile data were generated for each time point at the same location within the cell. Data were exported to Excel and standard deviations calculated. Surface plots were generated using the surface plot function in FIJI.
Images of granules taken using STED microscopy were analyzed in Volocity and diameters measured by drawing lines across the center of the granules. STED microscopy images of the actin network were imported into FIJI and processed before analysis. Background was subtracted using the Rolling Ball Subtraction algorithm with the radius set to 150 pixels and then pixel intensities were squared twice. An ROI was drawn around the interior of the cell and clearances were identified using the default autothreshold with “dark background” unchecked. Clearance areas were sorted, grouped, and counted in Excel based on size. Dividing the number of clearances per cell by the area measured normalized the values. Area cutoffs were implemented by using granule diameter as a reference. The smallest two sizes calculated assume that granules are uniformly spherical and require a clearance that has an area that would accommodate the equatorial area of the granule. The larger size categorizes all clearances that are larger than most granules.
Colocalization analysis was performed in Volocity using the “Find objects” tool to identify granules and a fixed intensity threshold, which was adjusted to generate interfilament or intercellular black space on a per field basis, to identify the actin network. The colocalized area of granule and actin network staining was divided by the area of the granule to yield a value denoting the percent of the granule area colocalized with actin. Five granules from 10 cells over 2 experiments were measured (N = 50 granules). Line profiles of granules and the actin network were performed in FIJI after channels were separated. All STED images were measured in their raw form, but shown after processing in Leica AS AF with the median noise reduction feature.
Images from platinum replica electron microscopy were inverted and linearly contrast enhanced using Photoshop (Adobe) and imported into FIJI for processing and analysis. Background was subtracted from each image using the Rolling Ball Subtraction algorithm with the radius set to 25 pixels. Pixel intensities were subsequently squared. Cells were identified using the default autothreshold and cell contact area and cell centroid were measured with “include holes” checked. ROIs were drawn around the interior of the cell to more accurately identify filaments and to avoid debris. Filaments were identified using the default automatic threshold with “dark background” checked. Clearances in the filamentous network were identified as mentioned above. Distance from the cell centroid was determined by inputting the coordinates of the cell centroid and a clearance centroid into the Pythagorean equation as described above for granule to MTOC distance.
All data were plotted using Prism (Graphpad).
Statistical significance was determined using Prism to perform one- or two-sample, unpaired or paired, two-tailed Student's t tests. Unless otherwise indicated all tests were two-sample, two-tailed, and unpaired. Where noted, n.s. denotes differences that have p values >0.05 and therefore considered not significant.
|
10.1371/journal.pbio.0060324 | Timing Precision in Population Coding of Natural Scenes in the Early Visual System | The timing of spiking activity across neurons is a fundamental aspect of the neural population code. Individual neurons in the retina, thalamus, and cortex can have very precise and repeatable responses but exhibit degraded temporal precision in response to suboptimal stimuli. To investigate the functional implications for neural populations in natural conditions, we recorded in vivo the simultaneous responses, to movies of natural scenes, of multiple thalamic neurons likely converging to a common neuronal target in primary visual cortex. We show that the response of individual neurons is less precise at lower contrast, but that spike timing precision across neurons is relatively insensitive to global changes in visual contrast. Overall, spike timing precision within and across cells is on the order of 10 ms. Since closely timed spikes are more efficient in inducing a spike in downstream cortical neurons, and since fine temporal precision is necessary to represent the more slowly varying natural environment, we argue that preserving relative spike timing at a ∼10-ms resolution is a crucial property of the neural code entering cortex.
| Neurons convey information about the world in the form of trains of action potentials (spikes). These trains are highly repeatable when the same stimulus is presented multiple times, and this temporal precision across repetitions can be as fine as a few milliseconds. It is usually assumed that this time scale also corresponds to the timing precision of several neighboring neurons firing in concert. However, the relative timing of spikes emitted by different neurons in a local population is not necessarily as fine as the temporal precision across repetitions within a single neuron. In the visual system of the brain, the level of contrast in the image entering the retina can affect single-neuron temporal precision, but the effects of contrast on the neural population code are unknown. Here we show that the temporal scale of the population code entering visual cortex is on the order of 10 ms and is largely insensitive to changes in visual contrast. Since closely timed spikes are more efficient in inducing a spike in downstream cortical neurons, and since fine temporal precision is necessary in representing the more slowly varying natural environment, preserving relative spike timing at a ∼10-ms resolution may be a crucial property of the neural code entering cortex.
| The precision of neuronal spike trains is at the center of a fundamental debate in neuroscience as to what aspects of neuronal signaling are important in representing information in the brain. Individual neurons can have extremely precise and repeatable responses to the visual stimuli that strongly drive them (down to 1-ms variability) [1–5], but they exhibit seemingly degraded temporal precision of firing activity in response to suboptimal stimuli [5–10]. In the presence of natural scenes, the activity of individual neurons is sparse [11] and precisely timed across repeated presentations of the visual stimulus, even though natural stimuli tend to vary on a time scale that is several times slower [12]. However, in most natural circumstances, the brain does not have access to multiple repetitions of the same identical stimulus, and, therefore, it is the precision of spiking across neuronal sub-populations on single trials that is ethologically relevant. While synchrony across neurons in the retina and visual cortex has been reported at various time scales, which can depend on the visual stimulus [10,13], the temporal precision of the neural code directly entering primary visual cortex, and its dependence on the stimulus, are still unknown. We used natural visual stimuli to investigate the spike timing precision of populations of geniculate neurons that serve as the direct input to visual cortex. We show that the response of individual neurons is less precise across stimulus repetitions when luminance contrast is reduced. However, this reduction in the precision of spike timing is not observed at the level of the neuronal population. Therefore, spike timing precision in populations of geniculate neurons is relatively insensitive to global changes in visual contrast and remains constant on the order of ∼10 ms. Since closely timed spikes from either a single neuron [14] or several neurons [15] are more likely to induce a spike in the downstream cortical neuron to which they are projecting, and since fine temporal precision is necessary in representing the more slowly varying natural environment [12], preserving the relative timing of spikes at a resolution of ∼10 ms may be a crucial aspect of the neural code entering primary visual cortex.
A short movie of a natural scene recorded from a “cat-cam” [16] was presented repeatedly to anesthetized cats while recording extracellular activity of multiple single units in the lateral geniculate nucleus (LGN) in vivo. To test how spike timing precision in individual cells and cell populations was affected by the properties of the visual stimulus, each group of cells was stimulated with both a high-contrast (HC) version and a low-contrast (LC) version of the movie. Figure 1A shows the firing activity of a typical ON-center geniculate cell in response to a 500-ms section of the movie presented at both high and low contrast. Each line in the raster plot corresponds to a single repeat and shows the spikes generated during one presentation of this movie section. The peri-stimulus time histogram (PSTH) shows the summed response across 64 repeated presentations. The geniculate neurons exhibited a typical pattern of response in which brief intervals of silence lasting 15 ms or more alternated with firing “events” [6], or groups of closely spaced spikes lasting up to 100 ms (median: 30 ms), which consistently occurred at approximately the same time on each movie presentation. A 63% reduction of the luminance contrast in the movie resulted in a 23% reduction in the neuronal firing rate, from 10.5 spikes/s at HC to 8.1 spikes/s at LC on average across cells (see full distribution in Figure S1A), and in a latency increase of 3.4 ms (Figure S1B). The firing rate reduction and latency increase are visible as a decrease in height and a shift in time in the PSTH events at LC in the particular example of Figure 1A.
PSTH events are typically wider in duration at LC than at HC [17] (as in the example rasters and PSTHs shown in Figure 1A, where the average event full width at half-height is 7 ms at HC, 11 ms at LC). Therefore, it has been assumed that suboptimal stimulation leads to a decrease in the temporal precision of the response. However, the widening of the PSTH events from the HC to LC condition could, in principle, arise from several sources, which have different functional implications: either an increase in the inter-spike intervals (ISIs) within the event on each trial (Figure 1B, left), or an increased variability in the timing of the events themselves from trial to trial (Figure 1B, right), or some combination of both factors. Further analyses were performed to test these two possibilities. First, over all cells, the distribution of (within-trial) ISIs was very similar at HC and LC. In Figure 1C, both distributions peak at 2.5 ms and have similar full width at half-height (8 ms at HC, 6 ms at LC). If ISIs were longer at LC than HC, the ISI distribution would be wider at LC than HC, which is the opposite of what we find. The fact that the ISI distribution at LC decays faster than at HC is simply due to the higher firing rate at HC, as is the case for Poisson processes (e.g., see [3]). Second, the timing of events showed significantly more across-trial variability (or “jitter”) at LC (10.9 ± 1.6 ms mean ± standard deviation) than HC (8.2 ± 1.3 ms), with a 2.7-ms difference on average across cells (paired t-test, p < 0.01, n = 45 cells; see Figure 1D). These results suggest, therefore, that overall spike timing variability is primarily due to across-trial variability in the timing of the event as a whole.
Ultimately, to compare within-cell and across-cell variability, correlation analysis is necessary. Therefore, we first re-examined and validated these single-cell precision findings in terms of correlation measures. First, we quantified the average PSTH event duration by measuring the width of the PSTH autocorrelation function, in which the temporal relationship between individual spikes is lost. We also quantified the temporal precision of individual spikes in a spike train by measuring the width of the spike autocorrelation function (i.e., the autocorrelation of the full spike train). As detailed in Figure S2, the width of the PSTH autocorrelation includes two possible sources of spike timing variability (within-trial and across-trial), whereas the temporal width of the spike autocorrelation function corresponds only to the within-trial variability. The PSTH autocorrelation and spike autocorrelation functions are shown in Figure 2A and 2B for four typical LGN cells at HC (top) and LC (bottom), with the temporal widths in each condition indicated. The PSTH autocorrelation functions were significantly wider at LC than at HC on average (Figure 2C; mean ± standard deviation: HC: 10.0 ± 2.5 ms; LC: 13.5 ± 3.1 ms; paired t-test, p = 5 × 10−17, n = 45), whereas spike autocorrelations did not show a significant difference in width between LC and HC (Figure 2D; HC: 8.4 ± 2.8 ms; LC: 8.5 ± 3.8 ms; paired t-test, p = 0.46, n = 45), thus confirming that spike timing variability is primarily due to across-trial event-timing variability (see Figure 1). This result was found not only with natural movies, but also with spatiotemporal white noise visual stimulation (Figure S3), and was not related to the cell X or Y type (Figure S4) nor to the occurrence of LGN bursts (Figure S5). These results indicate that decreasing the overall contrast increases the timing variability of groups of spikes (events), but preserves the relative inter-spike timing precision within each group of spikes (as illustrated in Figure 1B, bottom right), at a time scale on the order of ∼10 ms.
The analyses reported above involved the global level of contrast in the full movie (HC versus LC). To further elucidate the relationship between spike timing precision and contrast, we computed the local contrast experienced by each cell as the visual stimulus unfolds in time. Local values of spatiotemporal contrast ranged from 6% to 50% root-mean-squared (RMS) contrast (see Methods). We classified each firing event as corresponding to one in four levels of local contrast: 6–13%, 13–20%, 20–34%, and 34–50%. For each of these contrast levels, we computed the PSTH autocorrelation and spike autocorrelation of each cell, as above. As shown in Figure 2E and 2F, the results presented above for two global levels of contrast (HC and LC) were confirmed with four levels of local contrast. The width of the PSTH autocorrelation significantly decreased as contrast increased (Figure 2E; paired t-test; left to right pairwise comparisons: p = 1 × 10−6, p = 1 × 10−7, p = 2 × 10−4, n = 45 cells), while spike autocorrelation did not show a significant difference in width across the four different levels of local contrast (Figure 2F; paired t-test, p > 0.05 in all pairwise comparisons, n = 45 cells).
If the relative timing of spikes is preserved at different levels of contrast in single cells, what does it imply across the population? The activity of local groups of cells with neighboring receptive fields can be significantly correlated if the visual input itself has strong spatial and temporal correlations, as is the case with natural scenes [18–20]. Although it had been proposed that retinal and/or LGN neurons could remove these correlations through high-pass filtering achieved by lateral inhibition [21,22], more recent neurophysiological studies suggest that the cells do not de-correlate their inputs [13,23], and thus significant correlations from natural scenes remain present. The strength of pairwise correlation, defined as the area under the cross-correlation function (in the HC condition), decreased with the distance between the receptive fields of two cells (Figure 3A), following the decrease of spatial correlation strength in the visual stimulus (Figure 3A, inset). We focused our analysis on pairs of cells that displayed sufficient cross-correlation in the HC condition (n = 41, see Figure 3A) due to strong correlations in their visual input. Neurons with partially overlapped receptive fields (Figure 3B) typically receive similar visual input and therefore tend to share response events, as is evident in the two typical LGN X ON cells shown in Figure 3B and 3C. Both cells tended to fire during the same events, but there was some degree of timing variability (see also Figure S6).
The increased variability in event timing in the LC condition for individual neurons, reported above (see Figure 2), could coexist with a range of effects across a population of neighboring cells, involving within-trial and across-trial variability in the relative timing of spikes from several cells. Since the above result indicates that within cells, event times are more variable across trials at lower stimulus contrast, we then tested the hypothesis that event timing across cells, both within trial and across trials, is also more variable at lower stimulus contrast. From the perspective of a downstream layer 4 V1 neuron receiving direct thalamic input, incoming spike trains that arrive simultaneously from a pair of LGN neurons may be represented by superimposing both spike trains into a single combined spike train. The ISI distribution for these combined spike trains is largely invariant to changes in contrast and peaks at 2.5 ms for both levels of contrast (Figure 3D; full width at half-height: 6 ms at HC, 4 ms at LC), as was the case for single-cell spike trains (Figure 1C). The across-cell event-time variability can be estimated by merging events from both cells that overlap in time and measuring the variability in the median time of the combined event (Figure 3E; mean ± standard deviation of the event-time variability: 9.9 ± 1.0 ms at HC, 13.0 ± 1.4 ms at LC; n = 41 cell pairs). Across-cell event-time variability ranged between 8 and 19 ms, larger than but still on the same order of magnitude as that for single cells (Figure 1D). Its average value was slightly higher in the LC than in the HC condition (by 3.1 ms on average across cell pairs; p < 0.01, n = 41). However, it should be noted that this measure only indicates how the timing of combined firing events varies across repetitions of an identical stimulus. As argued above, it is the precision of spiking across neurons on single trials that is relevant for the neural population code.
To quantify further the relative precision of spiking across the neuronal population, we computed cross-correlation in pairs of cells. While all pairs under study displayed stimulus-induced correlation, a few pairs also showed correlation on a finer time scale (<1 ms), suggesting that they received common input from the same retinal ganglion cell [15]. Figure 4A and 4B shows the spike and PSTH cross-correlation functions for a pair of cells that shared input from the same retinal afferent (as in 4/41 pairs) and a pair of cells that did not. Importantly, since we are focusing on the neural representation of the visual scene rather than the details of the synaptic connectivity of LGN populations, our measure of spike correlation incorporates “signal” correlations (inherited from correlations present in the visual stimulus) as well as “noise” correlations (arising from other sources such as shared input from a common retinal afferent). Both are integral components of the neural code in natural viewing conditions [24] and, taken together, reflect the relationships between the correlation structure of the visual scene and the functional properties of the local neuronal circuit. Across all pairs of cells under study, the temporal width of spike cross-correlation only showed a small difference between HC and LC (Figure 4C; mean ± standard deviation: 14.7 ± 4.7 ms at HC, 15.7 ± 4.3 ms at LC; t-test, p = 0.05, n = 41 pairs; see also control analyses in Figures S3–S5). Moreover, the PSTH cross-correlation was very similar to spike cross-correlation (Figure 4D). To investigate how spike timing precision of cell pairs is influenced by local visual contrast, we computed the spike and PSTH cross-correlation at different contrast levels: 6–13%, 13–20%, 20–34%, and 34–50%, as done previously for individual cells. The cross-correlation was based only on the events for which the local contrasts in both receptive fields were at the same level. Consistent with the results presented above, there was no trend in the width of spike and PSTH cross-correlations as a function of local contrast (Figure 4E and 4F; some of the pairwise t-tests showed statistical significance, but not as a monotonic decrease in correlation width with increasing contrast). These results indicate that within-trial spike timing precision across cells is invariant to the change in contrast of the natural scene, despite the increased variability in event timings for individual cells across trials with decreasing contrast.
As evident in Figures 2F and 4E, the spike cross-correlation obtained from pairs of neurons was consistently wider than the spike autocorrelation obtained from each individual neuron. The difference in width was 8 ms on average (two-sided Wilcoxon rank sum test, p < 1 × 10−6 for each of four contrast levels, n = 45 cells, n = 41 cell pairs) and was also found on a pair-by-pair basis. In almost all cell pairs, cross-correlation width was significantly greater than the width of the autocorrelation functions of both cells, as shown in Figure 5A (see Figure S8 for the case of pairs lying close to the unity line, i.e., with similar within-cell and across-cell precision). This finding indicates that spike timing precision was coarser in neuronal pairs than in individual neurons. This decrease in precision can be explained by the fact that, in general, the events are not perfectly aligned across both cells, as illustrated in Figure 5B. Even if cells have wider PSTH events at LC than HC, the overall increase in event time variability from HC to LC by 3 ms is small in the face of pairwise variability, which is on average 8 ms greater than single-cell variability.
Another way to compare pairwise variability with contrast-based variability is by computing, for each shared event, the difference in event time between both cells (and its variability) and the difference in event time between the HC and LC condition (and its variability). As shown in Figure 5C, the difference in event times between two cells is more variable (i.e., has a wider distribution) than the difference in event times between HC and LC. The standard deviations of the distributions across all events are 16 ms (HC) and 18 ms (LC) across cells, and only 11 ms across contrast (n = 4,205 events). In other words, the variability in event timing across cells is approximately 1.5 times larger than the variability in event timing across levels of contrast. Therefore, in the face of across-cell variability, the smaller changes in variability due to changes in contrast are negligible. Thus, spike timing precision across most neighboring cells is relatively insensitive to contrast.
In response to movies of natural scenes, spike timing precision across LGN relay cells remained on the order of ∼10 ms, irrespective of contrast. The absolute timing of LGN firing events changed from trial to trial, and more so at low contrast than at high contrast, but the relative timing of spikes occurring in the same trial was insensitive to changes in stimulus contrast—not only within cells but also across correlated neighboring cells. While it is well known that the response properties of single cells are strongly modulated by contrast adaptation, which has effects including slower temporal dynamics and increased gain and selectivity at lower contrast [17,25,26], our results indicate that the temporal precision of the LGN population code is globally maintained in the face of a reduction in contrast.
Interestingly, while in individual neurons PSTH autocorrelations were consistently wider at low contrast than high contrast (Figure 2C and 2E), the width of the PSTH (and spike) cross-correlation was independent of contrast (Figure 4C, 4E, and 4F). This surprising result can be explained by the large variability in event timing across populations of neurons, which is about 1.5 times larger than the variability in event timing caused by changes in contrast (Figure 5C). A related finding is that while in individual neurons, the PSTH autocorrelations were consistently wider than the spike autocorrelations, in cell pairs, PSTH and spike cross-correlation had very similar widths (Figure 4D and 4F). This finding suggests that correlations between cells arose mostly from correlations present in the visual input in our experimental conditions, and that neural “noise” (or across-trial variability arising from intrinsic properties of the system) shows little correlation across neurons (Figure S7). Nevertheless, it should be noted that the presence of weak, pairwise noise correlations does not rule out the possibility of stronger, higher-order correlations at the population level [27–29].
Downstream from the LGN, the influence of stimulus contrast on the timing of spikes across V1 cells has only been recently addressed. Spike timing precision across cells in anesthetized macaque V1 reportedly decreased for low-contrast grating stimuli [10], whereas we found that it was contrast-independent in the cat LGN for natural stimuli. Further, a recent study in cortical slices suggested that the degree of noise correlation between two neighboring cortical cells increased with firing rate [30], unlike the thalamic signal and noise correlations reported here that were invariant to contrast-driven changes in firing rates. These discrepancies may be attributable to differences in the experimental preparations and in the visual stimuli, or they could be explained by specific contrast adaptation mechanisms that occur in cortex but not in precortical areas. For example, in the peripheral auditory system, adaptation to a constant stimulus reduces the firing rate but does not impair spike timing precision [31], in a similar fashion as what we found in the visual thalamus.
Preserving spike timing across cells at a ∼10-ms resolution may be a crucial aspect of the neural population code in natural conditions, given that the representation of spatiotemporally varying natural scenes requires a finer temporal precision than the time scale of the visual stimulus [12]. Furthermore, a 10-ms temporal resolution could facilitate “temporal coding” under the hypothesis that the neural representation of sensory information relies on specific temporal patterns of spikes [32–34]. However, the existence or preservation of specific temporal patterns is beyond the scope of the present study. Downstream from the thalamus, spike timing precision may well vary along the visual pathway. Single-cell studies found that trial-to-trial variability is similar in the LGN and V1 when a V1 cell is presented with its preferred stimulus, but that V1 cells become more variable for suboptimal stimuli [5,8,35]. In the presence of natural movies, which combine optimal and nonoptimal stimulation for each cell, recent studies in primate V1 indicate that some visual information is present in the phase of local field potentials at low frequency (<12 Hz) [36] and that power associated with spiking activity is only informative at frequencies under 12 Hz [37]. Therefore, the relevant time scale in V1 is probably on the order of tens of milliseconds, only slighter longer than what we found in the LGN. The small increase in variability in V1 trial-to-trial spike timing compared to the LGN [35] may be explained by nonlinearities in the spiking mechanism and may coexist with lower variability in V1 membrane potential [38]. It is also possible that temporal precision is higher between cortical cells receiving input from geniculate cells that share a common retinal afferent, in a divergent–convergent pattern of connectivity. In any case, it is difficult to relate these previous results to population coding. The degree of spike timing precision across V1 cells, especially in natural viewing conditions, is not well quantified, and how it would be affected by contrast or other variables is unknown. Further studies are needed to elucidate how the functional architecture of the thalamocortical circuit constrains spike timing precision across cells and how it affects the neural code entering V1.
Preserving synchrony across cells could have a number of functional advantages. Synchronous spikes from several thalamic neurons are reportedly needed to drive cortical cells to threshold [15,39]. Recent studies have suggested that the cortical response is sensitive to the timing of thalamic inputs and that the “window of opportunity” for integration of excitatory inputs at the thalamocortical synapse remains unchanged in the face of adaptation [40]. Therefore, the relative timing of spikes in thalamic neurons could be an important aspect of the population neural code entering primary sensory cortices and could benefit from being insensitive to some properties of the sensory world while maintaining sensitivity to other, presumably more interesting, features.
Single-cell activity was recorded extracellularly in the LGN of anesthetized and paralyzed cats using a seven-electrode system. Four animals were used for a total of ten electrode penetrations. Surgical and experimental procedures were performed in accordance with United States Department of Agriculture guidelines and were approved by the Institutional Animal Care and Use Committee at the State University of New York, State College of Optometry. As described in [41], cats were initially anaesthetized with ketamine (10 mg kg−1 intramuscular) followed by thiopental sodium (20 mg kg−1 intravenous during surgery and at a continuous rate of 1–2 mg kg−1 h−1 intravenous during recording; supplemented as needed). A craniotomy and duratomy were performed to introduce recording electrodes into the LGN (anterior, 5.5; lateral, 10.5). Animals were paralyzed with atracurium besylate (0.6–1 mg kg−1 h−1 intravenous) to minimize eye movements, and were artificially ventilated. Geniculate cells were recorded extracellularly from layer A of LGN with a multielectrode matrix of seven electrodes [42]. The multielectrode array was introduced in the brain with an angle that was precisely adjusted (25–30 degrees antero-posterior, 2–5 degrees lateral-central) to record from iso-retinotopic lines across the depth of the LGN. A glass guide tube with an inner diameter of ∼300 μm at the tip was attached to the shaft probe of the multi-electrode to reduce the inter-electrode distances to approximately 80–300 μm. Layer A of LGN was physiologically identified by performing several electrode penetrations to map the retinotopic organization of the LGN and center the multielectrode array at the retinotopic location selected for this study (5–10 degrees eccentricity). Recorded voltage signals were conventionally amplified, filtered, and passed to a computer running the RASPUTIN software package (Plexon). For each cell, spike waveforms were identified initially during the experiment and were verified carefully off-line by spike-sorting analysis. Cells were classified as X or Y according to their responses to counterphase sinusoidal gratings. Cells were eliminated from this study if they did not have at least 2 Hz mean firing rates in response to all stimulus conditions, or if the maximum amplitude of their spike-triggered average in response to spatiotemporal white noise stimuli was not at least five times greater than the amplitude outside of the receptive field area.
For each cell in the main experiments, visual stimulation consisted of 128–240 repeats of one of two short movies of natural scenes taken from “cat-cam” movies recorded from a small camera mounted on top of a cat's head while roaming in grasslands and forests [16]. As in [17], to improve temporal resolution, movies were interpolated by a factor of two (from 25 to 50 Hz) using commercial software (MotionPerfect, Dynapel Systems Inc.) and then presented at 60 frames per second, i.e., at 1.2× speed. Following interpolation, the intensities of each movie frame were rescaled to have a mean value of 125 (where the full range of intensity values was 0–255). Each movie spanned 48 × 48 pixels at an angular resolution of 0.2 degree per pixel. The first movie (presented to 28 of the cells included in the final analysis) was 750 frames and lasted 12.5 s, while the second movie (presented to the remaining 17 cells) was 600 frames long and lasted 10 s. The stimuli were presented at 60 frames per second with a 120-Hz monitor refresh rate, such that each frame was displayed twice. Each movie was repeated 64–120 times at each of two global levels of luminance contrast: 0.4 (high contrast, or HC) and 0.15 (low contrast, or LC) RMS contrast [17].
In addition to “cat-cam” natural movies, as a control for each cell we also used visual stimulation consisting of spatiotemporal binary white noise shown at high contrast (0.55 RMS contrast) and low contrast (0.20 RMS contrast). The spatial resolution and refresh rate of the white noise stimulus were the same as those of the natural scene movies. Each cell in the reported data was stimulated with the natural scenes movies as well as the white noise stimuli with an equal number of repeats (120 repeats at each level of contrast for 28/45 cells, 64 repeats at each level of contrast for 17/45 cells).
For each cell, the spatiotemporal receptive field was estimated by standard spike-triggered-average techniques based on spatiotemporal white noise stimuli [43,44]. The spatial receptive field was fitted with a difference of two-dimensional Gaussians. The distance between receptive fields was defined as the distance between the centers of the Gaussians. The diameter of each receptive field was estimated as the average length of the major and minor axes of the one–standard deviation ellipse that defines the receptive field center. The overlap between two receptive fields was evaluated as the normalized dot product of the two receptive fields, computed after each receptive field had been normalized so that its dot product with itself was one [45,46].
For each cell at each level of contrast (HC or LC), a single PSTH was computed as the cumulative response of the cell over all 64–120 repeats of the same short movie. Each PSTH was therefore 10 or 12.5 s long, depending on the duration of the stimulus presented to the cell. ISIs were computed as the time intervals between consecutive spikes; in the case of pairs of cells, we merged the spike trains from both cells and computed the ISIs from the combined spike train. Bursts were identified as groups of spikes separated from each other by 4 ms or less, where the first spike is preceded by a period of silence of 100 ms or more [47–49]. The degree of burstiness exhibited by each neuron was defined as the percentage of spikes belonging to a burst.
Previous studies typically define temporal precision of single neurons as the standard deviation of the spike times within an identified event across trials [6–9,31,35]. In this study, we first defined a related measure which is the (temporal) width of the central peak in the PSTH autocorrelation [50]. The width of PSTH events and the width of the PSTH autocorrelation function are directly related, by a factor of √2 in the Gaussian approximation. In computing the PSTH (and its autocorrelation), all spike trains that the cell produced in response to multiple repeats of an identical stimulus were collapsed into one “lumped” spike train (i.e., a PSTH with a 1-ms bin size, of the same duration as a single presentation of the movie, i.e., 10 or 12.5 s). In the PSTH autocorrelation measure, the relative timing of spikes within a given trial or across all trials were confounded. To investigate within-trial temporal precision, we therefore computed a different measure: the width of the central broad peak in the spike autocorrelation, which we defined as the autocorrelation function of the full (several minutes long) spike train without collapsing the trials together [51,52].
Although analysis of single cells was a necessary first step, the primary focus of this study was on spike timing variability across cells. Two definitions of cross-correlation were used: spike cross-correlation [52,53] and PSTH cross-correlation, which is the cross-correlation between two PSTHs. Spike cross-correlation width gives the spike timing variability across cells within each trial. PSTH cross-correlation has a different meaning: it is approximately equivalent to the “shuffled” or “shifted” spike correlation, in which each spike train of one cell is paired with a spike train of the other cell recorded during a different repeat of the same stimulus. The PSTH cross-correlation averages correlations from all possible pairwise combinations of repeats (actually including the non-shuffled one, which is only one in thousands of combinations and therefore has a negligible contribution).
All four types of correlation functions (spike or PSTH, auto- or cross-correlation) were made analogous to Pearson's correlation coefficient by (i) subtracting the product of the average firing rates, and (ii) dividing by a normalization factor (see below), such that correlation could take values between −1 and +1. To determine the existence of a central peak or trough in a correlation function, we found the Gaussian function that best fit the central ±100 ms, in a least-mean-square sense. The standard deviation of this Gaussian provides a measure of the correlation width. In the case of autocorrelation, the height Ai of the best-fitting Gaussian was measured for each cell i and was subsequently set to 1 to normalize the autocorrelation function. In the case of cross-correlation between cells i and j, the best-fitting Gaussian was normalized by a factor of
, where Ai and Aj are the heights of each respective autocorrelation function before normalization. The area under the Gaussian curve after normalization was used to define the strength of the cross-correlation between two neurons.
Inclusion criterion for pairs: A pair of cells was included in the final pairwise analysis if its spike cross-correlation function peaked at a value of 0.065 or higher, an arbitrary threshold below which the cross-correlation function could not be well fitted by a Gaussian function.
For all pairs of pixels corresponding to the receptive field centers of pairs of cells, we measured the correlation function between both time series (i.e., the time series of the intensity values of each pixel across all frames of the movie). Correlation strength was defined as the area under the Gaussian curve that best fit the cross-correlation function. The resulting spatial profile of correlation in the visual input, i.e., the graph of correlation strength as a function of the distance between two pixels, was fitted (in the least-mean-square sense) to an exponential function with a negative exponent, which is the form expected for spatial correlations in a signal with a power spectrum decreasing as 1/f2 with spatial frequency.
Single-cell event analysis: PSTH “events” were first defined in the PSTH at HC as times of firing interspersed with periods of silence lasting at least 20 ms. If the standard deviation of all spike times constituting an event was less than 20 ms, an attempt was made to break up the event into several events, a procedure in which the spikes were fitted to a mixture-of-Gaussians model using the Expectancy Maximization (EM) algorithm for maximum likelihood [54]. PSTH events at LC were then defined by aligning LC spikes to existing HC events if possible, with a preference for an HC event that occurred earlier rather than after the LC spike (since it is known that spikes tend to be more delayed at LC than HC). If no corresponding HC event was found, a new event was created at LC, with a corresponding empty event at HC. The timing of an event on a given repeat was defined as the median time of all spikes composing this event. For each event at a given contrast level, the event time variability was the standard deviation of the timing of the event across repeats. We computed for each cell its average event time variability across all events.
Pairwise event analysis: Starting from the single-cell event analysis above, each event from the first cell was matched to one or several events in the second cell with which it overlapped in time. If several events in one cell could be matched to a single event in the other cell, these events were merged into one. The list of all events that could be matched across the two cells constituted the list of “shared events.” For each shared event at a given contrast level, the event time variability was the standard deviation of the timing of the event across repeats and across both cells. We computed for each cell pair its average event time variability across all events.
Event-by-event analysis of event time difference, within cells and across cells: For all pairs (cell A and cell B), for each pairwise event that existed in the four cases (cell A at HC, cell A at LC, cell B at HC, and cell B at LC), we computed within-cell HC-LC event time difference as the average event time at LC minus the average event time at HC, for each of the two cells (cell A and cell B). In other words, we hold the cell fixed and varied across two contrast levels. We also computed across-cell event time difference at a given contrast level (HC or LC) as the average event time for cell A minus the average event time for cell B. In this case, we hold contrast fixed and varied across two cells. Therefore, each pairwise event yielded four different data points (2 × 2) to compare the distributions of across-cell and within-cell event time difference, as shown in Figure 5C.
For each cell, we computed the local value of spatiotemporal contrast as follows. For each firing event determined as above, we identified the smallest rectangle in the image that encompassed the cell's receptive field (e.g., 3 × 4 pixels) and extracted from the movie the luminance values of these pixels at the six previous frames. Six movie frames at 60 Hz correspond to a duration of ∼100 ms, matching the temporal kernel of the cells. The RMS contrast of this spatiotemporal patch of the movie was computed as the standard deviation over all the corresponding pixel values (e.g., 3 × 4 × 6 values). In the LC movie, local contrast values in the 45 cells ranged from 6–20% RMS contrast. In the HC movie, they ranged from 14–50%. For each cell, each firing event (in either the HC or LC condition) was assigned one in four levels of local contrast: 6–13%, 13–20%, 20–34%, or 34–50%. Correlation analysis was then performed as described above on small sections of data corresponding to individual events. We restricted the cross-correlation analysis to the firing events for which both cells experienced a value of local contrast that fell into the same range (out of the four ranges defined above).
|
10.1371/journal.pbio.1002598 | Neurocomputational mechanisms underlying subjective valuation of effort costs | In everyday life, we have to decide whether it is worth exerting effort to obtain rewards. Effort can be experienced in different domains, with some tasks requiring significant cognitive demand and others being more physically effortful. The motivation to exert effort for reward is highly subjective and varies considerably across the different domains of behaviour. However, very little is known about the computational or neural basis of how different effort costs are subjectively weighed against rewards. Is there a common, domain-general system of brain areas that evaluates all costs and benefits? Here, we used computational modelling and functional magnetic resonance imaging (fMRI) to examine the mechanisms underlying value processing in both the cognitive and physical domains. Participants were trained on two novel tasks that parametrically varied either cognitive or physical effort. During fMRI, participants indicated their preferences between a fixed low-effort/low-reward option and a variable higher-effort/higher-reward offer for each effort domain. Critically, reward devaluation by both cognitive and physical effort was subserved by a common network of areas, including the dorsomedial and dorsolateral prefrontal cortex, the intraparietal sulcus, and the anterior insula. Activity within these domain-general areas also covaried negatively with reward and positively with effort, suggesting an integration of these parameters within these areas. Additionally, the amygdala appeared to play a unique, domain-specific role in processing the value of rewards associated with cognitive effort. These results are the first to reveal the neurocomputational mechanisms underlying subjective cost–benefit valuation across different domains of effort and provide insight into the multidimensional nature of motivation.
| Rewards are rarely obtained without the motivation to exert effort. In humans, effort can be perceived in both the cognitive and physical domains, yet little is known about how the brain evaluates whether it is worth exerting different types of effort in return for rewards. In this study, we used functional magnetic resonance imaging (fMRI) to determine the neural and computational basis of effort processing. We developed two novel tasks that were either cognitively or physically effortful and had participants indicate their preference for a low-effort/low-reward versus a higher-effort/higher-reward version of each. Our results showed distinct patterns of reward devaluation across the different domains of effort. Furthermore, regardless of the type of effort involved, motivation was subserved by a large network of overlapping brain areas across the parieto-prefrontal cortex and insula. However, we also found that the amygdala plays a unique role in motivating cognitively—but not physically—effortful behaviours. These data impact current neuroeconomic theories of value-based decision making by revealing the neurocomputational signatures that underlie the variability in individuals’ motivation to exert different types of effort in return for reward.
| Neuroeconomic theories highlight that a key component of motivation is evaluating whether potential rewards are worth the amount of effort required to obtain them [1, 2]. Behaviours are executed if they have sufficient “subjective value” (SV), which is based on how much a potential reward is discounted—or devalued—by the effort required to obtain that outcome [3]. A characteristic of these cost–benefit valuations is that they are inherently highly subjective and thus vary across individuals [4–6]. Some people are willing to invest a quantum of effort for a reward that others would not. However, not all types of effort are subjectively evaluated in the same manner. Some individuals may be willing to overcome physically demanding challenges but be averse to mental effort, while others might show the opposite profile. Understanding the mechanisms that underlie cost–benefit valuations across different domains of effort is crucial to understanding the variability in people’s motivation [7, 8], but little is known of the neural or computational basis of these mechanisms.
Current theories of value-processing suggest that the computation of SV occurs in a common, domain-general network of brain regions [9]. Single-cell and neuroimaging studies have implicated areas within the basal ganglia and parieto-prefrontal cortices in the computation of SV for rewards that are devalued by costs such as risk, delays, or probability [9, 10]. Separately, research on effort-based decision making has implicated areas including the anterior cingulate cortex (ACC) (area 32’), dorsolateral prefrontal cortex (dlPFC) (areas 8/9), anterior insula (AI), intraparietal cortex (area 7), and several amygdala nuclei [11–16]. However, critical unanswered questions are whether these effort-sensitive areas compute the subjective value of discounted rewards associated with effort costs and whether these areas are differentially sensitive to the nature of those costs.
To date, most research on effort-based decision making has either focused on the cognitive or physical domains in isolation [4, 17–19]. The only previous study to have examined the neural mechanisms associated with different types of effort required participants to perform a cognitively or physically demanding task [20]. Importantly, however, participants in that study were not engaged in the choice of whether it was worthwhile to invest effort for reward. Thus, although this study was useful in examining how the brain motivates the exertion of different effort costs, the neural substrates that underlie the subjective valuation of reward—and the decision of whether to engage in an effortful action—remain unknown. Increasingly, these decision processes are being recognised as a critical component of motivated behaviour, with evidence that aberrant effort-based decision making may be a key element of motivational disorders such as apathy [18, 19, 21].
Here, we used the computation of SV as a key operation to understand cost–benefit decision making across the domains of cognitive and physical effort [9, 22–24]. In contrast to classical accounts, recent research in animals suggests that the mechanisms that underpin cognitive and physical effort discounting might be separable. For example, animal studies of the amygdala have causally linked it to motivation and the devaluation of reward by effort costs [25, 26]. Recently, however, a novel rodent decision-making task showed dissociable effects of amygdala and frontal lesions on cognitive effort–based decisions [27]. Specifically, amygdala and ACC inactivations caused changes to behaviour during a cognitive effort task [27] that were different to those in physical effort tasks [2, 25, 26]. Furthermore, amygdala inactivation influenced individual animals differently, suggesting that the amygdala may play a distinct role in subjectively valuing rewards associated with cognitive effort. Such findings suggest that the computation of SV in the context of effort may not be within a domain-general network of valuation areas, as is often argued [9].
To establish whether the SV of different effort costs are processed within domain-general or domain-specific brain systems, the current study directly examined whether the neural mechanisms underlying subjective reward valuation are sensitive to different types of effort. We first trained participants on two tasks that were closely matched on many properties that are known to influence the valuation of a reward (e.g., probability, duration prior to outcome) [28] but differed in whether cognitive or physical effort was required to obtain rewards. In each, we parametrically varied effort in one domain while holding the demands of the other constant. Then, while being scanned with functional magnetic resonance imaging (fMRI), participants chose between a fixed low-effort/low-reward “baseline” option and a variable higher-effort/higher-reward “offer.”
Central to our paradigm was the use of computational models to calculate the SV of each effort and reward combination relative to the baseline option for individual subjects, which allowed us to calculate subject-specific discounting parameters for each of the cognitive and physical effort tasks. Using model-based fMRI, we then identified regions in which blood oxygen level–dependent (BOLD) activity correlated with these parameters. This revealed that cognitive and physical effort discounting occurred in largely overlapping neural areas, but in addition, the right amygdala contributed uniquely to cognitive effort valuation.
In the cognitive effort task [4], we employed a rapid serial visual presentation (RSVP) paradigm [29], in which participants fixated centrally while monitoring one of two target streams to the left and right of fixation for a target number “7” (Fig 1A). Each target stream was surrounded by three distractor streams. The target stream to be monitored was indicated at the beginning of the trial by a central arrow and, during the trial, participants had to simultaneously monitor the central stream for a number “3,” which would be a cue to switch their attention to the opposite target stream. We parametrically varied the amount of cognitive effort over six levels by increasing the number of times attention had to be switched between streams from one to six. We previously confirmed that this task was able to manipulate perceived cognitive effort while controlling for physical demands and reinforcement rates [4].
In the physical effort task, participants exerted one of six different levels of force on a handheld dynamometer (Fig 1B). The effort levels for each participant were defined as proportions of their individually calibrated maximum voluntary contraction (MVC) (8%, 13%, 18%, 23%, 28%, and 33%), as determined at the beginning of the experiment. The duration of each of the cognitive and physical effort trials was identical (14 s), ensuring that participants’ choices were not due to temporal discounting [30, 31].
Participants were first trained on each of the cognitive and physical effort tasks outside the scanner in counterbalanced order. They undertook an extensive training session of 60 trials for each task to familiarise themselves with the effort associated with each level in each domain and so that we could estimate performance measures for each task (see Materials and Methods). Participants were told that their reimbursement at the end of the study would be contingent on performance and that for each trial that they performed well, they would be awarded one credit, which would be later converted into a monetary amount. This training resulted in participants being rewarded on over 80% of trials, and a repeated-measures ANOVA revealed that, although there was a significant effect of effort (F(1.7, 57.2) = 7.48, p < .005), neither the main effect of domain nor its interaction with effort were significant (p > .05; S1 Fig). Importantly, this indicates that the reinforcement rates did not differ between tasks and ensured that subsequent effort-based decisions in the two domains could not be confounded by participants’ belief that they would be differentially successful at obtaining rewards across the two tasks.
The critical choice phase occurred after the training phase, while participants were being scanned with fMRI (Fig 1C). During this phase, participants made cost–benefit decisions for the cognitive and physical effort tasks separately. On each trial, they were presented with a fixed low-effort/low-reward “baseline” option and a variable high-effort/high-reward “offer.” The baseline option was an opportunity to perform the lowest level of effort for one credit, while the offer presented a higher number of credits (2, 4, 6, 8, or 10 credits) for having to invest a greater amount of effort (levels 2–5). Importantly, by providing participants with the identical range of reward options for both cognitive and physical effort, we could disentangle how cognitive and physical effort differentially devalued the identical rewards. In addition, in order to eliminate the effect of fatigue on participants’ decisions, they were not required to execute their choices within the scanner. Instead, they were instructed that they would be required to perform a random selection of ten of their choices at the conclusion of the experiment and that their remuneration would be based on these randomly selected trials.
Because separate decisions were made for the cognitive and physical tasks, we were able to estimate the extent to which the same amount of reward was devalued within each domain for each participant. An important feature of our design was that we temporally separated the presentation of the offer from that of the response cue. Thus, participants did not know which button corresponded to the baseline or offer until the onset of the response prompt. This ensured that we could examine activity time-locked to a cue from which SV would be processed independently, with activity related to these events not confounded by preparatory motor activity.
Using the modelling parameters derived above, we computed the SV for the effort and reward combinations on every trial and used the difference in value between the SV of the chosen offer and the value of the baseline as a parametric regressor modelled to the onset of the offer cue [37]. Many studies have shown that regions we hypothesised would be engaged by cost–benefit valuations are sensitive to the difference in the SV of two options rather than to the SV of an offer per se [31, 37]. Thus, we fitted the SV difference on each trial within the cognitive domain as a parametric modulator time-locked to the onset of each cognitive offer and performed the corresponding analysis for offers in the physical domain. This allowed us to examine activity covarying with SV for the cognitive and physical domains separately. These parametric modulators were defined based on the discounting parameters estimated for each participant’s choice behaviour. We considered significant those voxels which survived whole brain–level, voxel-wise corrections for multiple comparisons (p < 0.05, corrected for family-wise error [FWE]).
We used model-based fMRI to determine whether shared or separate neurocomputational mechanisms underlie cost–benefit valuation in the cognitive and physical domains. Computational modelling revealed that individuals were differentially sensitive to cognitive and physical effort. Neuroimaging data showed that activity in several areas previously implicated in effort processing covaried with the subjective value of rewards independent of effort domain. This included the dACC, dmPFC, dlPFC, IPS, and anterior insula. Importantly, activity within many of these areas also covaried with absolute reward and effort levels, suggesting an integration of these parameters within these areas. However, in contrast to the view that SV is processed in an entirely domain-general manner, an ROI analysis revealed that the right amygdala appeared to process SV uniquely for rewards associated with cognitive and not physical costs. Importantly, none of these results could be explained by choice difficulty or perceived risk. Together, these data indicate that cost–benefit valuation in the human brain is underpinned mostly by a common, domain-independent mechanism but that the amygdala may play an important role in valuing rewards associated with cognitive effort. These results therefore suggest that the classical view of a domain-general set of brain regions for valuation cannot fully account for the subjective valuation of rewards associated with all effort costs [9].
To our knowledge, no study to date has examined the neural correlates of SV associated with cognitive versus physical effort in a single paradigm. The only study that has addressed the nature of cognitive and physical effort examined the processing of raw magnitudes of effort and reward without considering individuals’ subjective valuations and did not require subjects to make choices about whether the effort was worth exerting to obtain the reward [20]. Such an approach is common in the literature and assumes that rewards have a similar effect across individuals to exert the associated effort [11, 13, 43, 44]. However, preferences vary depending on subject-specific cost–benefit valuations, and SVs potentially afford a more sensitive measure of capturing individual differences in motivation [9, 22, 23, 45]. Furthermore, SV has been proposed as an important entity in understanding apathy in healthy individuals as well as those with clinical disorders of motivation [8, 21, 46]. Defining the neural and computational mechanisms that underlie the choice to exert effort for reward is therefore crucial to understanding the variability in motivated behaviour across individuals. In the present study, by parametrically varying effort across six levels in both domains, we were able to computationally model SVs for individual participants and therefore more closely examine the key computations that underpin choice behaviour and motivation.
Our paradigm had several other advantages. First, the protocol involved manipulating effort in two separate domain-specific tasks, as opposed to requiring participants to exert a combination of both forms of effort to attain specific rewards in each trial [20]. We were therefore able to examine choice behaviour for identical rewards in each domain independently. Second, although many studies have examined the processing of effort and reward, the majority may have been confounded by motor execution for the choices or preparatory activity related to an upcoming effortful exertion. In the design used here, it was possible to investigate activity specifically related to decisions based on SV by temporally separating the choice process from the preparation or execution of the effortful act. Third, by controlling the temporal parameters of both the cognitive and physical effort tasks, it was possible to eliminate delay discounting as an explanation of choice behaviour [5, 47, 48]. Fourth, by using computational modelling approaches, we were able to examine activity that varied with SV. Finally, by ensuring that reinforcement rates were similar for the six levels of effort within and across domains, it was possible to ensure that probability discounting could not have contributed to our findings. Thus, the study reported here isolates the effect of SV on choice and motivation independently of many effects that can confound studies examining effort-based decision making. As such, we can effectively rule out the possibility that several regions that we identified were only related to the energisation of behaviour and not to motivation or the valuation of behaviour [49].
Our model comparisons indicated that individuals valued rewards differently when associated with cognitive and physical effort. This was demonstrated by the winning model, which specified separate discounting functions requiring separate discounting parameters for cognitive and physical effort. This conclusion was also supported by the more general pattern of the computational modelling results, which showed that the models assuming equal reward devaluation across cognitive and physical effort (i.e., those assuming a single discounting parameter) provided poorer fits than those that assumed separate discounting parameters. This finding that different functions best fitted cost–benefit valuations for cognitive and physical effort most likely reflects differential sensitivities to effort in the two domains. Our finding that a parabolic function best accounts for participants’ choice behaviour in the physical effort task is in keeping with previous observations [33]. In contrast, effort discounting in the cognitive domain has been much less studied [6], and it is likely that the specific shape of a discounting function will depend on the specific cognitive faculty being tested (e.g., attention versus working memory). However, the key point for the present study is that, in the tasks that we used, identical rewards were valued distinctly across both domains.
Strikingly, despite rewards being devalued at different rates and in a mathematically distinct manner across the two domains, a largely overlapping network of regions was involved in processing the SV of rewards devalued by both the cognitive and physical effort cost. It is important to note that this finding does not rely on the generalisability of these specific functions to other cognitive or physical effort–based tasks. However, the fact that effort discounting in our task is best described by separate functions does considerably strengthen this result, as it implies that any differences between cognitive and physical effort cannot simply be a matter of scale (e.g., some participants finding one task more effortful than the other). Rather, it suggests a possible difference in the underlying mechanism between the two processes. Furthermore, the separate discounting functions render the imaging results more compelling by showing that the SVs computed from entirely different functions nevertheless engage overlapping brain regions.
Regardless, a question that remains is whether the same pattern of results would be achieved in a cognitive and physical effort task that were best described by the identical discounting function. Exploratory analyses using a single function to model choice across both domains revealed a pattern of domain-general and domain-specific effects that were essentially similar to those of the primary analyses. However, it remains for future studies to verify the conclusions from our study in the case of cognitive and physical effort tasks that are best described by identical discounting functions.
Interestingly, most of the domain-general areas that encoded subjective value also showed a significant negative effect of reward and a significant positive effect of effort. The findings that many domain-general areas that encode SV also encode raw reward and effort levels are not incompatible—indeed, one interpretation is that these regions integrate the reward and effort on offer into a value signal. Although many previous studies have examined the neural basis of processing SV [9], we believe this is one of the first demonstrations that regions of the brain can process a SV formed from costs that devalue rewards at different rates. Furthermore, although some of these domain-general regions may be involved in processing decision difficulty in certain contexts [50], this is not always the case [51], and none of the regions identified in the present study were found to encode choice difficulty across both the cognitive and physical domains.
The key to elucidating the neural basis of cost–benefit decision making will be understanding how this domain-general network learns or forms a valuation of rewards associated with different forms of effort [15, 22]. A central role of the dorsal ACC/dmPFC in value-based decision making and motivation is considered by some to be in signalling the value of a behaviour in comparison to alternatives [52, 53]. The study reported here extends this notion by showing that this region not only processes the SV of an offer but also integrates effort and reward information independent of the nature of the effort cost [50, 52, 54]. In addition, single-unit studies have shown that dACC/dmPFC neurons signal the net value of rewards associated with effort, and the necessity of this region in cost–benefit valuation has been demonstrated by lesion studies that report that inactivation of medial prefrontal cortex impairs an animal’s ability to overcome effort costs [15, 16, 25, 48, 55]. Recently, several human studies have also shown this region to be important in calculating choice value for effortful rewards. Although the majority of these have been in the physical domain [11, 13, 14], a recent investigation reported a similar pattern for cognitive effort [56].
Neurons sensitive to reward information have been identified in the dlPFC [55, 57–59], and the activity of lateral prefrontal areas in humans correlates with predicted SVs that guide decision making [60]. Lateral intraparietal neurons have been found to signal expected value [61], and parietal activity has been reported in tasks requiring value comparisons [62, 63]. Lastly, insular activity is negatively correlated with the SV of rewards associated with higher effort [14, 64], and dopaminergic responses, which play an important role in motivated decision making, exhibit greater variability in the insula with less willingness to expend effort for reward [65]. Our findings extend this body of data by showing that the process of subjective reward valuation occurs independent of the nature of effort costs, and suggest that it is underpinned by activity in a the dACC/dmPFC, dlPFC, IPS, and anterior insula.
Do these regions of domain-independent areas comprise a network for subjective valuation? Tracer studies in macaque monkeys and neuroimaging studies in humans suggest that these domain-independent regions are monosynaptically connected. The upper bank of the dorsal anterior cingulate sulcus is connected to the anterior portions of the insula, several amygdala nuclei, and BA 9/46 in the lateral prefrontal cortex. Similar projections exist between each of these locations and the other domain-independent regions within this putative network [42, 66–69]. In addition to the connectional anatomy, it has been noted that these same domain-independent regions are activated during a variety of different cognitive and motor control tasks [70, 71]. It has been argued that this multiple-demand (MD) network is involved in flexibly controlling the cognitive processes required across a large number of tasks [70].
In this context, our results could be taken as support for the notion that this network is activated independent of the nature of the cost or associated behavioural domain. However, our findings also suggest a more nuanced interpretation of the functional properties of the MD network. In our study, activity in this network was influenced by the value of working and not by the demand alone. Moreover, as highlighted above, these areas contain single neurons that respond to reward valuations, and the BOLD signal in these regions has been shown to scale with subjective reward valuations in studies investigating temporal discounting or probabilistic reward-based decisions. Thus, a more refined account might be that the MD network is crucial for motivating behaviours across different domains of behaviour. Such a notion would explain why these regions are activated during many cognitive and motor tasks in which motivation must be sustained for successful performance [72].
Importantly, we found evidence of domain specificity for cognitive effort valuation, specifically in the right amygdala. The amygdala is known to play an important role in reward valuation, and single-unit recordings have demonstrated that neurons here encode the value associated with individual items [26, 73–75]. Recent evidence points to the amygdala as playing a crucial role in effort-based decision making in rodents, with neurophysiological data showing that the amygdala plays an important role in valuing effort [40, 41]. Recently, some have proposed that the amygdala is sensitive to different types of effort costs [27] and also highlighted the key role for this region in the flexible control of cognitive processes. However, drawing a definitive conclusion, especially in humans, requires comparisons across species and across tasks.
Substantial differences exist between the paradigms used in valuation studies and include differences in reinforcement schedules, training intervals, reward magnitudes, and contrast effects. Furthermore, previous effort-based tasks have not tightly controlled the contributions from each domain to their manipulations of effort, thus making it difficult to compare the relative contributions of the two domains. Indeed, such discrepancies may even underlie varying amygdala involvement in cost–benefit decision-making tasks across cognitive and physical effort. In our study, we designed each of our closely matched tasks to hold all features constant except for the type of effort involved, which was maximised in each domain relative to the other. We were therefore able to provide more direct evidence that the human amygdala may be differentially involved in cognitive over physical effort valuation. Nevertheless, while our result is consistent with the preceding studies noting potentially dissociable roles of the basolateral amygdala for cognitive and physical effort–based decisions, the finding of amygdala domain specificity does deserve replication in future studies and would be even more compelling if it was demonstrable at a whole-brain level.
Interestingly, previous studies have shown that the VS and vmPFC are engaged when processing value [20, 39, 76]. Here, we found no such activity for either cognitive effort, physical effort, or the conjunction. This was the case even after specifically probing these areas with regions of interest defined on the basis of previous studies. A key difference between this study and all previous studies implicating the VS and vmPFC in value processing is that previous tasks required effort to be exerted while participants were being scanned, and most of the effects may have been related to the execution of the effortful task rather than to the choice of whether the effort was worth exerting. This may suggest that the VS and vmPFC process value primarily when value may guide or motivate the execution of a behaviour that will be followed immediately by a rewarding outcome, rather than in the evaluation of whether resources should be allocated to a task at all.
Rewards in real life are rarely obtained without effort. Our model-based fMRI approach revealed that effort discounting in the cognitive and physical domains is underpinned by largely shared neural substrates but that the amygdala uniquely contributes to cognitive effort valuation. Importantly, neither delay nor probability discounting can account for our results. It has been postulated that disorders of diminished motivation—such as apathy and abulia—which are manifest in multiple neurological and psychiatric conditions, may be characterised as a diminished willingness to exert effort for reward [46, 77]. Our findings may therefore help us understand the neural basis for such disorders of motivation by providing an insight into their multidimensional nature and identifying potential neural foci that might be manipulated to modulate motivation [78].
This study was approved by the Central University Research Ethics Committee of the University of Oxford (MSD-IDREC-C1-2014-037). We recruited 38 young, healthy, right-handed participants. All participants had no history of neurological or psychiatric illness and were not taking regular medications. Four participants were excluded: 2 for failing to provide responses on a high proportion of trials while being scanned (over 9%), and a further 2 because of excessive head motion within the scanner (more than 5 mm of translation). The final group of 34 participants (23 females) had a mean age of 24 y (range 19–39).
All participants were behaviourally trained on a cognitively effortful task and a physically effortful task prior to being scanned. These extensive training sessions were aimed at familiarising participants with the effort associated with all levels for both tasks. The training phase for each task began with 18 practice trials (3 per effort level) and was followed by a further 60 trials to reinforce behaviour (10 per effort level). Behavioural analyses of task performance were conducted on the latter 60 trials. After training, we scanned participants while they made economic decisions based on how much effort they would be willing to trade off for varying levels of reward. The order of training in the physical and cognitive effort tasks was counterbalanced across participants.
|
10.1371/journal.pntd.0007316 | Vaccination with a chikungunya virus-like particle vaccine exacerbates disease in aged mice | Chikungunya virus (CHIKV) is a re-emerging pathogen responsible for causing outbreaks of febrile disease accompanied with debilitating joint pain. Symptoms typically persist for two weeks, but more severe and chronic chikungunya illnesses have been reported, especially in the elderly. Currently, there are no licensed vaccines or antivirals against CHIKV available. In this study, we combined a CHIK virus-like particle (VLP) vaccine with different adjuvants to enhance immunogenicity and protection in both, adult and aged mice.
CHIK VLP-based vaccines were tested in 6-8-week-old (adult) and 18-24-month-old (aged) female C57BL/6J mice. Formulations contained CHIK VLP alone or adjuvants: QuilA, R848, or Imject Alum. Mice were vaccinated three times via intramuscular injections. CHIKV-specific antibody responses were characterized by IgG subclass using ELISA, and by microneutralization assays. In addition, CHIKV infections were characterized in vaccinated and non-vaccinated adult mice and compared to aged mice.
In adult mice, CHIKV infection of the right hind foot induced significant swelling, which peaked by day 7 post-infection at approximately 170% of initial size. Viral titers peaked at 2.53 × 1010 CCID50/ml on day 2 post-infection. Mice vaccinated with CHIK VLP-based vaccines developed robust anti-CHIKV-specific IgG antibody responses that were capable of neutralizing CHIKV in vitro. CHIK VLP alone or CHIK plus QuilA administered by IM injections protected 100% of mice against CHIKV. In contrast, the antibody responses elicited by the VLP-based vaccines were attenuated in aged mice, with negligible neutralizing antibody titers detected. Unvaccinated, aged mice were resistant to CHIKV infection, while vaccination with CHIKV VLPs exacerbated disease.
Unadjuvanted CHIK VLP vaccination elicits immune responses that protect 100% of adult mice against CHIKV infection. However, an improved vaccine/adjuvant combination is still necessary to enhance the protective immunity against CHIKV in the aged.
| Chikungunya virus is responsible for outbreaks of febrile illnesses accompanied with debilitating join pain in subtropical and tropical regions of the world. The disease caused by chikungunya virus typically resolves itself within weeks, but may be persistent and more severe in elderly individuals. Currently, there are no licensed vaccines, although a virus-like particle vaccine is currently being tested in Phase II clinical trials. In this study, we formulated chikungunya virus-like particles with adjuvants to skew and enhance the immune responses against chikungunya, and vaccinated adult and aged mice. Our aim was to identify a vaccine formulation that would protect adult and elderly populations. Results showed that the unadjuvanted vaccine was very effective in adult mice, eliciting strong virus-neutralizing antibody titers, and protecting mice against chikungunya infection and disease. In contrast, chikungunya disease was exacerbated in mice vaccinated with the virus-like particle vaccine alone or with QuilA adjuvant. This study highlights the need for an improved vaccine approach to safely and effectively vaccinate the elderly against chikungunya viral infections.
| Chikungunya virus (CHIKV) is a re-emerging pathogen responsible for causing outbreaks of febrile disease accompanied with debilitating joint pain. CHIKV was first discovered in Tanzania in 1952, but outbreaks became more widespread, encompassing countries in Africa, Asia, Europe, and islands of the Pacific and Indian Oceans before emerging in the Americas in 2013 [1, 2]. More recently, in 2016–2017, there has been a resurgence of autochthonous CHIKV transmission in India [3], Pakistan [4], and Italy [5]. The virus is transmitted to humans through the bite of Aedes aegypti and Aedes albopictus mosquitoes. Chikungunya infection results in illness, in which fever and joint polyalrthralgia, are typically reported symptoms [6]. Acute symptoms persist for two weeks, but more chronic arthralgias may persist for months to years in a subset of subjects. More severe and/or chronic chikungunya illnesses were first widely reported in retrospective studies of the chikungunya epidemics of Reunion Island [7, 8] and India [9]. Following infection, patients experience renal, respiratory, hepatic, and cardiovascular system failures. In addition, diseases of the central nervous disease and encephalitis are major areas of concerns [7–9]. People over 60 years of age are at particular risk for severe chikungunya-associated illnesses, with case fatalities reported [8–10]. However, the incidence of CHIKV infection in this population is not remarkable in comparison to other age groups [11–13]. The specific mechanisms that lead to increased severity of CHIKV illness in the elderly are not known, but increased understanding could lead to better treatments and vaccines for this at-risk population.
Vaccinating elderly individuals presents a special challenge since they are more prone to severe illness and vaccine efficacy drops in this population [14]. The age-associated changes in the immune system are collectively termed immunosenescence and include fewer circulating antigen presenting cells and tissue-associated dendritic cells, decreased phagocytosis, decreased toll-like receptor signaling, reduced naïve B and T cells, and chronic basal level of inflammation [15]. Elements of the immune system that remain intact include tissue macrophages and CD8+ T cell-mediated responses [14, 15]. Different vaccine approaches to counter immunosenescence in the aging include the use of higher vaccine doses, booster vaccinations, adjuvants, and vector-based vaccines [15]. Many vaccine delivery platforms are in development for a chikungunya vaccine, including formalin-inactivated viral vaccines, live-attenuated viruses, chimeric alphaviruses, DNA-based vaccines, recombinant subunit vaccines, and virus-like particle [VLP]-based vaccines [16]. The most promising candidates, including a non-adjuvanted CHIK VLP vaccine, are being tested in Phase I and II clinical trials in adults between the ages of 18–60 years of age [17]. Thus, we will continue to have a gap in knowledge regarding 1) CHIKV vaccine efficacy in the elderly and 2) understanding the vaccine characteristics needed to elicit a protective immune response in this population.
In this study, CHIKV virus-like particles were adjuvanated and used to vaccinate adult and aged mice. Adjuvants were chosen for their abilities to not only enhance, but skew immune responses. The goal was to identify a CHIK VLP vaccine formulation that would protect both adult and aged mice populations.
The complete sequence encoding structural proteins (C-E3-E2-6K-E1) of the Chikungunya virus S27 strain [accession #AF369024] was codon-optimized for expression in Spodoptera frugiperda and synthesized by Genewiz [South Plainfield, NJ, USA]. The Bac-to-Bac baculovirus expression system [Thermo Fisher Scientific, Waltham, MA, USA] was subsequently used to generate recombinant baculoviruses expressing CHIKV structural proteins. Briefly, the structural gene sequence was inserted into the pFastBac1 vector, under the control of the Autographa californica multiple nuclear polyhedrosis virus (AcMNPV) polyhedrin for high-level expression in insect cells. The CHIK C-E VLP/pFastBac1 construct was then transformed into DH10Bac E. coli, where C-E genes flanked between mini-Tn7 sites on the pFastBac1 plasmid and the LacZ gene flanked between mini-attTn7 target sites on a AcMNPV bacmid are transposed to generate recombinant bacmid. The presence of C-E genes was verified by polymerase chain reaction (PCR) analysis using primers that hybridize to sites flanking the mini-attTn7 site: pUC/M13 forward 5’-CCCAGTCACGACGTTGTAAAACG-3’ and pUC/M13 reverse 5’-AGCGGATAACAATTTCACACAGG-3’. Baculovirus was generated and passaged in Sf9 S. frugiperda insect cells, maintained in serum-free, SF900 II SFM medium [Thermo Fisher Scientific]. To generate the initial recombinant viruses, 8×105 Sf9 cells per well were seeded onto a 6-well plate and allowed to adhere for 15 min. The cells were then transfected with 1–2 μg bacmids using Cellfectin transfection reagent [Thermo Fisher Scientific]. The cells were observed for cytopathic effect and supernatants were harvested and clarified after 72h post-infection. The P1 virus was then passaged in a 30ml, spinner-flask culture of Sf9 cells at a cell density of 2×106 c/ml, and harvested 72h post-infection to generate P2 virus. For expression, Sf9 cells were cultured in spinner flasks to a density of 2×106 c/ml in a total volume of 250 ml and infected with recombinant baculovirus at a MOI of 1. Cultures were harvested once cell viability was reduced to roughly 80% or 72-96h after infection, and the cells were pelleted at 500×g for 5 min at 4°C. Supernatants were collected and filtered through a 0.22μm pore membrane before sedimentation via ultracentrifugation. CHIK virus-like particles (VLP) were sedimented through a 20% glycerol cushion at 100,000×g for 4h. The sedimented VLP pellets were resuspended in sterile phosphate buffered saline (PBS). Similarly, E1 and E2 genes, designed as transmembrane-truncated versions, were synthesized and cloned into the pFastBac HT vector. The pFastBac HT vector adds an N-terminal 6×His tag and and tobacco etch virus (TEV) proteolytic site to each gene. Recombinant bacmids and baculoviruses were generated as described above and soluble E1 and E2 proteins were expressed in Sf9 spinner flask cultures.
The cultures containing soluble E1 and E2 proteins were harvested and cells were sedimented at 500×g for 5 min at 4°C. Supernatants were collected and filtered through a 0.22μm pore membrane and the proteins were purified by affinity chromatography using Ni-NTA resin [Thermo Fisher Scientific]. Briefly, the clarified cultures were incubated with Ni-NTA resin with shaking for 2.5-3h at room temperature before they were added to the columns. The medium was allowed to flow through and the Ni-NTA resin was washed three times with PBS containing 10mM imidazole. His-tagged proteins were then eluted twice with PBS containing 250mM imidazole. Upon verification of eluted proteins by SDS-PAGE analysis, E1 and E2 proteins were dialyzed and concentrated using Amicon Ultra-15 centrifugal filters [Millipore, Burlington, MA] with a 10 KDa molecular weight cut-off and sterile 10% glycerol in PBS as the exchange buffer. Total protein concentrations for E proteins and VLPs were measured using the Micro BCA protein kit as per manufacturer’s protocol [Pierce, Rockford, IL, USA].
Samples from each step of the purification process were prepared by combining 30μl of samples with 6μl 6×Laemmli buffer with beta-mercaptoethanol [βME] and heating to 100°C for 5 min. Proteins were separated on a Bolt 10% Bis-Tris Plus gel [Thermo Fisher Scientific] at 200V for 30 min and protein bands were stained with PageBlue protein staining solution [Thermo Fisher Scientific] and destained with distilled water.
Samples were prepared by mixing 10μg of total protein in Laemmli buffer with βME, unless otherwise noted. These samples were boiled at 100°C for 5 min and proteins were separated on Bolt 10% Bis-Tris Plus gel as before. Next, the proteins were transferred from the gels onto PVDF membranes using the Trans Blot Turbo apparatus [Bio-Rad, Hercules, CA, USA]. The membranes were blocked for 5–10 min in iBind solution [Novex]. Polyclonal mouse anti-E1 and anti-E2 sera were recovered from mice vaccinated with E1 or E2 proteins in the lab and used to probe for these proteins. Mouse monoclonal antibody against E2 [Clone CHK-48, BEI Resources, Manassas, VA, USA] was also used to probe for E2. Goat anti-mouse conjugated with horseradish peroxidase [Southern Biotech, Birmingham, AL] was used as the secondary antibody. The membrane, antibody, and iBind solutions were loaded into the iBind Western System [Life Technologies, Carlsbad, CA] from which point all steps in the membrane blotting process proceed automatically by sequential lateral flow. Blotting using the iBind system was complete after 2.5 h. Following washing of the membrane twice more with PBS with 0.05% Tween-20 [PBS-T], the membrane was exposed with Clarity Western ECL Substrate [Bio-Rad]. Images were captured using my ECL Imager [Thermo Fisher Scientific].
Vero cells [ATCC, Manassas, VA, USA] were cultured in Dulbecco’s Modification of Eagle’s Medium [DMEM, Mediatech, Manassas, VA, USA] supplemented with 10% fetal bovine serum [FBS], 2mM L-glutamine, 100 U/ml penicillin, and 100μg/ml streptomycin [10%FBS-DMEM] and maintained at 37°C and 5% CO2. C6/36 mosquito cells [ATCC] were cultured at 28°C and 5% CO2 in Eagle’s mimimum essential medium [EMEM, Mediatech] supplemented with 10%FBS, 2mM L-glutamine, 100 U/ml penicillin, and 100μg/ml streptomycin [10%FBS-MEM]. CHIKV LR2006-OPY1 virus was obtained from the World Reference Center for Emerging Viruses and Arboviruses (WRCEVA). Upon receipt, this virus was passaged twice in C6/36 mosquito cells. Virus concentration was determined in Vero cells and reported as the 50% cell culture infectious dose (CCID50) per volume [ml].
Female C57BL/6J mice were obtained from the Jackson Laboratory [Bar Harbor, ME, USA] at 6–8 weeks of age for studies in adult mice. Female C57BL/6J mice were also obtained at 12 months and allowed to age to at least 18 months for studies in aging mice. All procedures in the document were approved by the UGA Institutional Animal Care and Use Committee, # A2015 06-004-Y3-A12. Mice were immunized on days 0, 21, and 42 and blood samples were taken on days 0, 14, and 35 via the submandibular method using 5mm lancets.[18] Vaccines were formulated to contain 30μg [~0.3–0.4μg E2 content] chikungunya VLPs adjuvanted with 20μg QuilA [InvivoGen, San Diego, CA, USA], 10μg R848/Resiquimod [InvivoGen], 1:1 by volume Imject Alum [Thermo Fisher Scientific], or in PBS alone (no adjuvant). Vaccines were delivered via intramuscular injection to the hindlimb quadriceps in a total volume of 50μl or subcutaneous injection to the scruff of the neck in a total volume of 100μl.
Nunc Maxisorp 96-well plates [Thermo Fisher Scientific] were coated overnight at 4°C with 10μg/ml E1, E2, or VLP in PBS. The plates were then washed three times with PBS with 0.05% Tween-20 (PBS-T) and blocked with 200μl 1% bovine serum albumin in PBS (blocking buffer) for 1 hr at room temperature. Serum samples from individual mice were diluted to 1:100 in blocking buffer and added at 100μl/well in duplicate wells. The sera were allowed to react for 2 hr at room temperature. The plates were washed three times with PBS-T and bound sera were reacted with goat anti-mouse IgG-Fc [1:50,000], IgG1 [1:10,000], IgG2c [1:10,000], or IgG3 [1:10,000] antibody conjugated with alkaline phosphatase [Bethyl Laboratories, Montgomery, TX, USA] for 1 hr at room temperature. The plates were washed three more times and allowed to develop for 20 min following the addition of 100 μl para-nitrophenylphosphate substrate [SeraCare, Milford, MA, USA]. The plates were read at a wavelength of 405nm using a BioTek PowerWave XS plate reader with Gen5 version 2.07 software [BioTek, Winooski, VT, USA].
Mouse sera from immunized or naïve mice were heat-inactivated at 56°C for 30 min. Two-fold serial dilutions of the sera were prepared in 10% FBS-DMEM and added to 96-well, cell-culture plates. CHIKV LR2006-OPY1 strain was then added at 200 CCID50 per well, and virus-antibody solutions were incubated together for 1 h at 37°C and 5% CO2. The final serum dilutions ranged from 1:20 to 1:2560. Each plate had two sets of assay controls: one column of wells contained virus only and a second column contained medium only. Vero cells were added at 104 cells/well and plates were incubated for 5 days at 37°C and 5% CO2. The cells were fixed for 20 min with 10% formalin in PBS and stained with crystal violet for 5 min at room temperature. Neutralizing titers were measured and expressed as the reciprocal of the highest serum dilution that inhibited cytopathic effect.
Adult C57BL/6J mice [6-8-week old] were challenged with CHIKV Reunion Island isolate LR2006-OPYI, which is of East Central South Africa lineage [ECSA] as previously described.[19] Mice were observed for 14 days following challenge. Prior to infection, the mice were anesthetized with a 100μl cocktail of 10 mg/kg xylazine plus 100 mg/kg ketamine in via intraperitoneal injection, and initial weight and hind foot measurements were recorded. Foot size was defined as width × breadth (mm2) and measured using a digital micrometer with 0.001mm resolution. The virus was subcutaneously injected into the right hind footpad at 50μl/mouse, while the mice were still under anesthesia. Pilot viral dose challenge experiments were conducted in naïve adult and aging mice to determine optimal challenge conditions. Mice were observed twice daily and weight and foot measurements were recorded once a day. Blood samples were collected on between days 1–5, and at day 14 post-infection. Based on these initial studies, a 105 CCID50 challenge dose of LR2006-OPY1 CHIKV was used to test vaccine efficacy, and blood collections were reduced to 2, 4, 6, and 14 days post-infection. Approximately 40–60 μl of blood was collected from mice on sampling days, except on day 14 when the mice were anesthetized and terminally bled.
A two-step assay was used to measure viral loads in serum samples of mice challenged with CHIKV. C6/36 cells were grown to 100% confluence in T75 culture flasks, detached by scraping, and divided equally into four 96-well plates per T75 flask. The next day, ten-fold serial dilutions (10−1–10−8) of the mouse sera were prepared in 10%FBS-EMEM and used to inoculate 96-well plates of confluent C6/36 at 100μl /well. The cells were allowed to incubate for 3 days at 28°C and 5% CO2. Vero cells were seeded at 2×104 cells/well in a total volume of 100μl 10% FBS-DMEM per well and 25μl of C6/36 culture supernatants were transferred into triplicate wells containing Vero cells. Vero cells were incubated for 4 days at 37°C and 5% CO2. The cells were then fixed with 10% formalin for 20 minutes and stained with 1% crystal violet solution for 5 minutes at RT. Cells with 95% or more cytopathic effect were counted for each dilution and viral loads [CCID50/ml] were calculated using the Spearman-Karber equation.[20]
Mouse TNF-α, IL-6 and IL-1β ELISA MAX kits from BioLegend [San Diego, CA, USA] were used to detect inflammatory cytokines in sera of adult and aged mice, as per manufacturer’s protocol. Briefly, Nunc Maxisorp plates were coated coated overnight at 4°C wit 100 μl/well of anti-mouse TNF-α, IL-6, or IL-1β diluted to 1:200 in carbonate buffer, pH 9.5. After washing once with PBS-T, all subsequent steps were performed at room temperature, with shaking. Plates were blocked with blocking buffer (1% BSA in PBS). TNF-α and IL-6 standards were diluted and used at final concentrations 3.9–500 pg/ml, while IL-1β was used at 15.6–2000 pg/ml in blocking buffer. Pooled sera from adult or aged naïve mice were prepared by mixing 10 individual serum samples together. Pooled sera and standards were added to plates [100 μl/well] and incubated for 2 hr. Plates were then washed four times with PBS-T and incubated with biotinylated detection antibody at 1:200 dilution in blocking buffer for 1 hr. The plates washed 4 times with washing buffer and 1:1000 diluted avidin-HRP was added and incubated for 30 mins. After 5 washes, TMB substrate solution was added (100μl/well) and plates were incubated in the dark for 15 mins. The reaction was stopped with 100μl/well stop solution (2N H2SO4) and plates were read using the PowerWave XS microplate spectrophotometer at a wavelength of 450 nm.
Recombinant mouse TNF-α [Life Technologies] was diluted in 10% FBS-DMEM and mixed with 200 CCID50 CHIKV LR2006-OPY1 virus per well, and virus-antibody solutions were incubated together for 1 h at 37°C and 5% CO2. The final TNF-α dilutions ranged from 5–80 ρg/ml. Each plate had two sets of assay controls: one column of wells contained virus only and a second column contained medium only. Vero cells were added at 104 cells/well and plates were incubated for 5 days at 37°C and 5% CO2. The cells were fixed for 20 min with 10% formalin in PBS and stained with crystal violet for 5 min at room temperature. Wells with 95% or more cytopathic effect were counted for each TNF-α dilution and reported as the percentage of wells with CPE.
GraphPad Prism 7 for Mac OS X software was used to perform statistical analyses. One-way, two-tailed ANOVA, followed by Tukey post-hoc tests were performed for data derived from one time-point. Two-way, two-tailed ANOVA followed by post-hoc tests were performed for data collected over multiple time-points. A p-value of less than 0.05 was considered significant.
All mouse-related experiments were conducted in compliance with the guidelines of the University of Georgia Institutional Animal Care and Use Committee [A2015 06-004-Y3-A12], and in accordance with the National Research Council’s Guide for the Care and Use of Laboratory Animals, The Animal Welfare Act, and the CDC/NIH’s Biosafety in Microbiological and Biomedical Laboratories guide. Management of animal experiments, care, and was conducted by the University of Georgia’s Animal Resources Department that is accredited by the AAALAC.
CHIK VLPs, as well as CHIKV E1 and E2 proteins were produced and purified, as vaccines or reagents to analyze the immune responses elicited against CHIKV antigens. Multiple AcMNPV bacmids were generated to encode CHIKV C-E genes for expression and subsequent VLP assembly and for 6×His-E1ΔTM and 6×His-E2ΔTM. CHIKV gene insertions were verified by PCR analysis [Fig 1A]. Three bacmid clones (c1-c3) were chosen for each construct. The C-E [3765 bp] insert plus flanking sequences (2300 bp) resulted in a 6065 bp band and all three CHIK C-E VLP bacmid clones contained the correct insert as determined by electrophoresis through 1% agarose in tris-acetate-EDTA (TAE) [Fig 1B, top]. The bacmid clones containing E1ΔTM (1266 bp) plus flanking sequences, including upstream HIS-tag region (2430 bp) resulted in a PCR product of 3696 bp and these constructs were also verified by gel electrophoresis [Fig 1B, middle]. The bacmid clones containing E2ΔTM (1122 bp) plus flanking sequences, including upstream HIS-tag region (2430 bp) resulted in a PCR product of 3552 bp [Fig 1B, bottom]. These bacmids were independently used to transfect SF9 cells to successfully produce recombinant baculoviruses Ac-C-E VLP, Ac-6×His-E1ΔTM, and Ac-6×His-E2ΔTM capable of expressing CHIKV VLP and E proteins [Fig 1C–1E].
VLPs self-assembled from C-E proteins and were recovered from Ac-C-E VLP-infected SF9 cell culture medium. These particles were purified from non-particles by ultracentrifugation through a 20% glycerol cushion, as we previously reported [21]. The purified particles were probed using the Chk48 anti-E2 monoclonal antibody and a 50 kDa band representing the CHIKV E2 protein was detected by Western blot [Fig 1C]. Five of six batches of purified VLPs [4/12/17-4/17/17] were pooled and used for vaccination. These lots were compared to a batch produced two years earlier [3/25/15] that were stored at -80°C, demonstrating stability of VLP over time, when frozen. Zika subviral particles (SVP) were not detected using the Chk48 anti-E2 antibody. Like the VLP, HIS-tagged E1 and E2 proteins were secreted into the culture media of Sf9 cells infected with their respective baculoviruses: Ac-6×His-E1ΔTM and Ac-6×His-E2ΔTM. These E1 and E2 proteins were purified via affinity chromatography using NiNTA resin. For each set of purifications, unbound protein flow-through, three separate washes, and two separate elutions were collected and analyzed by SDS-PAGE, with PageBlue protein staining. Untagged proteins were removed by the second and third washes as shown by SDS-Page analysis [Fig 1D and 1E]. Recombinant E1 was successfully recovered in elutions 1 and 2, as demonstrated by the presence of an intense band at approximately the expected size of 48 KDa [Fig 1D]. E1 protein was recovered in the elutions and then pooled, dialyzed, and concentrated. Anti-E1 mouse polyclonal sera reacted strongly with a~48 kDa band by Western blot analysis. Elutions 1 and 2 containing a 44 kDa recombinant E2 protein [Fig 1E] were pooled together, dialyzed, and concentrated. Anti-E2 polyclonal sera recognized the purified, recombinant E2 band.
Adult C57BL/6J mice were vaccinated at weeks 0, 3, and 6 with VLP alone or coupled with one of three different adjuvants: QuilA, R848, and Imject Alum. Preliminary studies identified CHIK VLP, VLP plus QuilA, and VLP plus Alum formulations as the top vaccine candidates in adult mice when delivered intramuscularly based on CHIKV antigen-specific IgG responses [S1A-S1C Fig], neutralizing antibody responses [S1D Fig], and protection against CHIKV-associated arthritis [S1E Fig]. Thus, these top candidates were also assessed in aged mice, following the same regimen. All adult and aged miced vaccinated with CHIK VLP formulations seroconverted after 3 doses as determined by ELISA for anti-VLP total IgG [Fig 2A], and anti-VLP total IgG levels were statistically significant when compared to both, adult and aged PBS control sera. However, anti-VLP total IgG titers were significantly higher in adult mice vaccinated with VLP alone in comparison to aged mice vaccinated with VLP alone (abs 0.9211 to 0.512, p = 0.0439). Morever, adult vaccinated with VLP plus Alum had higher anti-VLP total IgG titers then any of the aged mice vaccinated with VLPs (all p < 0.01). The aged CHIK VLP-vaccinated mice were able to mount anti-VLP IgG1 responses that were comparable to adult CHIK VLP-vaccinated mice [Fig 2B], but significantly lagged in their anti-VLP IgG3 responses [Fig 2D]. Anti-VLP IgG2c responses were insignificant in CHIK VLP-vaccinated adult and aged mice [Fig 2C]. The IgG responses against CHIK E1 protein were robust in all adult mice vaccinated with CHIK VLPs, but were insignificant in aged mice vaccinated with CHIK VLPs [Fig 2E–2H]. For adult mice vaccinated with VLP alone, the anti-E1 IgG subclass responses was dominated by IgG2c [abs 0.805] and IgG3 (abs 0.792), followed by IgG1 (abs 0.621) antibodies. For adult mice vaccinated with VLP plus QuilA, the anti-E1 IgG subclass responses consisted of IgG1 (abs 0.943), followed by IgG3 (abs 0.732), and finally IgG2c (abs 0.547). In contrast, adult mice vaccinated with VLP plus Alum largely consisted of IgG1 antibodies [abs 0.733], less IgG2c [0.354], ad no IgG3. The strongest anti-E2 total IgG responses were elicited in adult mice vaccinated with VLP plus QuilA or VLP plus Alum [Fig 2I]. Vaccination with VLP plus QuilA in adult mice resulted in a strong anti-E2 IgG3 response [abs 0.751, Fig 2L], followed by anti-E2 IgG1 [abs 0.527, Fig 2F], but no significant anti-E2 IgG2c [Fig 2G]. In contrast, the VLP plus Alum vaccination only elicited significant anti-E2 IgG1 (abs 0.324) antibodies. Adult mice vaccinated with VLP alone elicited significant anti-E2 IgG3 (abs 0.861) and IgG2c antibodies (abs 0.313), but no significant anti-E2 IgG1. CHIK VLP vaccinations in aged mice did not elicit any significant anti-E2 IgG responses.
Vaccinations with CHIK VLP-based formulations induced antibodies in adult mice that were able to neutralize CHIKV LR2006-OPY1 [Fig 3, mean titer range 1:691–1:1136]. In adult mice, CHIK VLP only and CHIK VLP plus QuilA formulations elicted significant neutralization titers as compared to control sera. However, in aged mice, the VLP formulations failed to elicit CHIKV-neutralizing antibodies.
Mice were infected with 105CCID50 CHIKV LR2006-OPY1 via subcutaneous injection of the right hind footpad. The infected mice were then monitored for 14 days following CHIKV infection. Blood samples were collected on days 2, 4, 6, and 14. While there was some gradual weight loss in the control adult mice, there was little weight loss observed in the adult mice vaccinated with CHIK VLPs only or CHIK VLPs plus QuilA [Fig 4A and 4D]. Unvaccinated aged mice did not lose any weight. However, aged mice vaccinated with CHIK VLPs or VLPs plus QuilA experienced some gradual weight loss, similar to what was observed in unvaccinated adult mice [Fig 4A and 4D].
We also measured the size of the infected right hindfeet and the uninfected left hindfeet were measured as additional controls. Any foot size deviation beyond 15% of the intitial foot size was considered a significant change. Adult mice vaccinated with PBS experienced significant swelling of the CHIKV-injected right hind foot. Swelling was visible by day 6 post-infection. Peak right foot swelling occurred on day 7, reaching approximately 170% of the initial foot size, before gradually returning to normal size by day 12 post-infection [Fig 4B and 4E]. Immunizations with VLP alone [Fig 4B] or VLP plus QuilA [Fig 4E] in adult mice offered complete protection throughout the complete time-course of the experiment. No measurable inflammation of the left hind feet were observed in any of the adult mice [Fig 4C and 4F]. In contrast to adult mice, naïve aged mice were resistant to CHIKV-mediated arthritis of the injected right hind foot [Fig 4B and 4E]. On the other hand, CHIK VLP-vaccinated old mice were more susceptible to CHIKV infection than naïve old mice. Aged mice vaccinated with VLP had pronounced foot swelling with one peak at day 2 post-infection [125% of initial foot size, Fig 4B], and then later on for a sustained period between 7–12 days post-infection (approximately 130% of initial foot size). There was significant right hind foot swelling in the aged group vaccinated with VLP plus QuilA on days 1–4 post-infection and again on days 6–9 [Fig 4E]. At peak swelling, the right foot size was 156% of its original size. There was little or no change in the size of left hind feet of challenged old mice change size over time [Fig 4C and 4F].
Infectious virus was recovered from adult mice vaccinated with PBS and challenged with CHIK LR2006-OPY1 at day 2 [2.53 × 1010 CCID50/ml], day 4 [2.77 × 105 CCID50/.ml], and day 6 (2.52 × 104 CCID50/.ml) [Fig 5]. Peak viral titers were observed on day 2 [Fig 5] and virus was completely cleared by day 14 post-infection. Virus was not recovered in adult mice vaccinated intramuscularly with VLP or VLP plus QuilA [Fig 5A and 5B, respectively] at 2, 4, or 6 days post-infection. Viral infection in unvaccinated old mice produced significantly lower viral titers on day 2 [5.31×103 CCID50/ml], with infection peaking on day 4 [5.08×104 CCID50/ml], before decreasing on day 6 [507 CCID50/ml], and eventual clearance by day 14 [Fig 5]. In contrast, old mice vaccinated with VLP developed higher viremia [4.2×107 CCID50/ml, Fig 5A] than unvaccinated old mice as measured on day 2 post-infection. However, viral loads in old, VLP-vaccinated mice decreased to similar levels as unvaccinated mice on days 4 and 6 post-infection. Old mice vaccinated with VLP plus QuilA also had higher viremia on day 2 post-infection (3.15×109 CCID50/ml) than unvaccinated mice. The viral loads in old mice vaccinated with VLP plus QuilA decreased on day 4, but peaked again on day 6 post-infection [3.15×109 CCID50/ml, Fig 5B].
The resistance to CHIKV infection observed in naïve, aged mice may be associated with chronic low-grade inflammation that accompanies aging [22]. To test this theory, sera, collected from naïve adult mice and aged healthy mice, were assayed for the presence of TNF-α, IL-6, and IL-1β [Fig 6]. TNF-α, IL-6, and IL-1β were below levels of detection in healthy, naïve, adult mice. Pooled sera from groups of aged mice that appeared otherwise healthy had significantly elevated basal levels of TNF-α at 5.971 ± 3.82 pg/ml in comparison with adult mice [Fig 6A]. Aged mice also had significantly elevated basal levels of IL-6 at 15.17 ± 3.766 pg/ml in comparison to adult mice [Fig 6B]. Serum cytokine levels of IL-1β were not significantly altered in aged mice versus adult mice [Fig 6C]. However, two groups of pooled sera from healthy, aged mice had high IL-1β levels of 15.4 and 56.9 pg/ml. Overall, naïve aged mice have elevated basal levels of inflammatory cytokines in comparison to naive adult mice. Moreover, the presence of TNF-α inhibited infection of Vero cells infected with CHIKV LR2006-OPY1 as measured by reduced CPE in the monolayers with as little as 5 pg/ml [Fig 6D], and an IC50 of 14.49 ± 2.99 as determined by nonlinear regression using GraphPad Prism software.
Two of the most promising vaccines against CHIKV infection, a CHIK VLP vaccine and a measles-vectored vaccine expressing CHIK VLPs, have cleared Phase I trials with positive outcomes. These vaccines 1] are overall safe and tolerable, and 2] elicit CHIKV neutralization titers in adults 18–50 years of age. These two candidates are now being tested in healthy adults of 18–60 years of age in five Caribbean island nations. Based on results from Phase I trials, at least two immunizations with the live-vectored or CHIK VLPs are needed for 100% seronconversion and induction of neutralizing antibodies in healthy adult subjects [23, 24]. The goals of our study were to improve these vaccines by adding adjuvants, and also show that these formulations would be efficacious in an aged model of infection. R848/resiquimod was used as a TLR7/8 agonist to induce a Th1 biased response [25]. Imject Alum induces primarily a Th2 biased response [26] and QuilA was selected because it enhances T-dependent and T-independent immune responses [27]. In adult mice, CHIK VLP administered alone or with QuilA administered by IM elicited strong antibody responses that were neutralizing in vitro, and these vaccines protected 100% of mice against CHIKV challenge, as determined by lack of swelling of the injected foot, lack of weight loss, and lack of viral titers.
Antibodies play a critical role in the clearance of CHIKV infections by both neutralizing virus infection and enhancing the clearance of virally infected cells [28, 29]. Potent neutralizing antibodies, composed of multiple subclasses, are directed against the E2 protein [30]. Early induction of anti-CHIKV anti-E2 human IgG3 subclass antibodies successfully clear CHIKV infections and lead to faster recovery. In adult mice, the most effective vaccine candidates were CHIK VLPs with no adjuvant or VLP plus QuilA, both administered by IM injections. Both of these vaccine formulations elicited antibodies directed against both the E1 and E2 proteins. A combination of robust IgG1, IgG2c, and IgG3 anti-E1 responses were detected [Fig 2F–2H]. In contrast, IgG3 antibodies were the predominant anti-E2 response elicited by these vaccines, with lower titers of pro-inflammatory IgG2c elicited by VLP alone or IgG1 by VLP vaccines formulated with QuilA [Fig 2J–2L]. Mouse IgG3, which is not a homolog of human IgG3, is induced independently of T-cell help and appears shortly after vaccination [31]. This anti-CHIKV IgG3 response may enhance clearance of CHIKV infected cells in adult mice, since CD4 knockout mice are still able to control and clear CHIK virus as effectively as wild-type mice [32]. In contrast, B cell deficient mice (μMT) remain persistently infected [33]. Mouse IgG3 binds FcγRI, but is thought to function primarily through activation of complement [34]. These anti-E1 and anti-E2 responses were absent in aged mice vaccinated with CHIK VLPs or VLP plus QuilA. The sera from aged mice vaccinated did react with whole VLP preparations by ELISA, with similar levels of anti-VLP IgG1 as the adult mice. Perhaps the presence of these antibodies that did not bind specifically to soluble E1 or E2, contributed to CHIKV disease enhancement in these animals. A phenomenon coined antibody-dependendent enhancement has been described for dengue and other viruses whereby non-neutralizing antibodies facilitate entry of antibody-bound virions via FcγR [35]. Antibody-dependent enhancement has also been observed in vitro with another alphavirus, Ross River Virus [36]. A more recent study shows that convalescent sera from CHIKV-infected subjects mediates enhanced binding, but not enhanced replication of CHIKV in primary human monocytes and B cells in vitro via FcγRs [37]. In contrast, increased chikungunya viral replication is observed in Raw 264.7 mouse macrophages in the presence of mouse anti-CHIKV IgG in vitro. Mice infected with CHIKV and then treated with subneutralizing levels of anti-CHIKV IgG also develop higher levels of viremia and disease as measured by foot swelling [37].
Another unexpected feature of the aged mice was their resistance to chikungunya viral infection and disease [Fig 5]. We speculate that this resistance may be associated with elevated levels of inflammatory cytokines present in aged mice, but not in adult mice. Chronic, low-level inflammation associated with aging has been coined “inflammaging”. The elevation of some cytokines, such as IL-6, are associated with longevity, while elevation of other pro-inflammatory cytokines, such as TNF-α, are associated with higher rates of mortality in humans [22]. The role of inflammaging is complicated because while enhanced inflammation leads to disease, mortality, and poor vaccine outcomes, inflammation can help the immune system resist virally-induced disease [22]. Healthy aged mice with detectable levels of inflammatory markers TNF-α and IL-6 [Fig 6A and 6B] were resistant to CHIKV-associated disease [Fig 4B & 4E, Fig 5]. Furthermore, addition of exogenous TNF-α to Vero cells inhibited CHIKVinfection in vitro, based on reduction in CPE. Future studies testing the effects of exogenous TNF-alpha on CHIK viral infections in vivo would help to corroborate these results.
Basal inflammation in the naïve, aged mice may have been helped these mice resist CHIKV infection, but it may also have contributed to the poor immune responses to vaccination with CHIK VLPs. Reduction of vaccine efficacy due to inflammation has been observed in elderly people vaccinated with standard vaccines against influenza virus [38] and hepatitis B [39]. Furthermore, elevated plasma levels of TNF-α correlate with lower antibody titers generated in post-menopausal women following vaccination with influenza vaccine. Thus, it is possible that the basal levels of TNF-α observed in age mice, but not in young mice, contributed to the decrease antibody responses observed after vaccination with CHIK VLP-based vaccines. A recent publication suggests that this problem may be circumvented by pre-treament of elderly patients with anti-inflammatory drugs, such as Losmapimod, a small molecule p38 mitogen-activated protein kinase inhibitor [40]. In addition, given that elderly people are already in a pro-inflammatory state, using an adjuvant to enhance inflammatory responses to a vaccine may not be the most beneficial approach for CHIKV.
In contrast to our observations in CHIKV-infected aged mice, a study by Uhrlaub et al. [41] found that CHIKV infections resulted in more severe infections in aged mice. There are a few key differences between our studies. The ages of the adult and aged mice were similar, but we used female mice, while Uhrlaub et al. used male mice. They also used a different CHIKV strain: SL15649. This strain resulted in a different disease progression in adult male mice than what we observed in adult female mice: foot-swelling was biphasic with two peaks on days 3 and then on day 8 with SL15649 infection, while we observed one main peak on day 7 with LR2006-OPY1. While the progression of CHIKV-associated foot swelling we observed is comparable to what has been previously published by Gardner et al [19] and Metz et al [42] using CHIKV LR2006 OPY1 strain, biphasic foot-swelling has also been in observed with this same strain at 106 CCID50 in female mice [43] and at 103 focus-forming units in female and male mice [44]. Thus, the use of different strains cannot be solely responsible for the difference in results. While not previously anticipated in our study, the stark difference in results may have to do with gender in aged mice. A longitudinal, retrospective study on prognostic factors of inhospital deaths in elderly patients by L. Godaert et al. [45] found that the male sex was an independent predictor of inhospital deaths due to CHIKV infection. Thus, perhaps in aged mice, the male sex may also predispose them to more severe disease. In addition study by Uhrlaub et al. also showed that CD4+ T cells and neutralizing antibody responses elicited by CHIKV infections were significantly decreased in aged, male mice compared to the adult, male mice. Thus overall, even in the male mice, the VLP vaccine formulations would likely also elicit poor protective responses as compared to those in adult mice.
While our study suggests that naïve, aged, female mice are resistant to CHIKV infections, infections in aged people are much more complicated. Comorbidities may increase the risk of developing more severe CHIKV disease upon infection. CHIKV infection may be complicated by pre-exisiting comorbidities or may exacerbate chronic renal, respiratory, cardiac diseases [9, 46]. A recent systematic meta-analysis of 11 different studies showed that hypertension, diabetes, cardiac disease, and asthma were the most frequent comorbidities associated with patients infected with CHIKV [47]. Furthermore, hypertension and diabetes had a 4-5-fold higher prevalence in patients over 50 years of age, and patients with diabetes at higher risk for severe CHIKV disease [47]. Thus, perhaps CHIKV infections in diabetic mouse or non-human primate models may provide a better understanding of CHIKV infection, spread, and virus-induced disease pathology observed in severe CHIKV-induced disease. In addition, it would be important to investigate these preconditions in both female and male animals. These models could then be used for preclinical testing of vaccines and antivirals against CHIKV.
In summary, CHIK VLPs alone elicit strong neutralizing antibody titers that protect 100% of mice from CHIKV infection and disease. However, more research is needed to identify a vaccine that will protect the elderly and people with chronic conditions, such as hypertension and diabetes against CHIKV disease. If the current measles-vectored an CHIK VLP-based vaccines make it through all phases of clinical trials, licensing of these vaccines should be limited to the aged groups in which they were tested in, until further research can be conducted. It is possible that a completely different set of vaccine formulations or regimens may need to be developed for the elderly.These would need to be thoroughly vetted in preclinical studies to ensure that older people are not put at further risk for severe CHIKV infections. Finally, relevant models of severe CHIKV disease in adults and aged animals are needed to evaluate these vaccines.
|
10.1371/journal.pbio.0050043 | In Vivo Reinsertion of Excised Episomes by the V(D)J Recombinase: A Potential
Threat to Genomic Stability | It has long been thought that signal joints, the byproducts of V(D)J recombination, are
not involved in the dynamics of the rearrangement process. Evidence has now started to
accumulate that this is not the case, and that signal joints play unsuspected roles in
events that might compromise genomic integrity. Here we show both ex vivo and in vivo that
the episomal circles excised during the normal process of receptor gene rearrangement may
be reintegrated into the genome through trans-V(D)J recombination
occurring between the episomal signal joint and an immunoglobulin/T-cell receptor target.
We further demonstrate that cryptic recombination sites involved in T-cell acute
lymphoblastic leukemia–associated chromosomal translocations constitute hotspots
of insertion. Eventually, the identification of two in vivo cases associating episomal
reintegration and chromosomal translocation suggests that reintegration events are linked
to genomic instability. Altogether, our data suggest that V(D)J-mediated reintegration of
episomal circles, an event likely eluding classical cytogenetic screenings, might
represent an additional potent source of genomic instability and lymphoid cancer.
| Lymphoid cells recognize billions of pathogens as a result of gene rearrangements that
generate pathogen-specific B- and T-cell receptors. This genetic reshuffling, called
V(D)J recombination, occasionally misfires and damages genomic integrity. When such
aberrations dysregulate proto-oncogenes, cancer ensues. It has become increasingly clear
that multiple oncogenes acting in different cellular pathways can cooperate to cause
cancer. Nevertheless, in the case of T-cell acute lymphoblastic leukemia, about a third
of cases display oncogene activation in the absence of identified aberration, suggesting
the presence of additional mechanisms of chromosomal alteration. In the hunt for such
mechanisms, episomal circles (DNA segments that are excised during V(D)J recombination)
have recently drawn attention. Moreover, signal joints, short sequences formed after
gene rearrangements, once considered harmless, now appear to take part in events that
might compromise genomic integrity. Using ex vivo recombination assays and genetically
modified mice, we demonstrate that episomal circles may be reintegrated into the genome
through recombination occurring between the episomal signal joints and a T-cell receptor
target. Furthermore, we show that cryptic recombination sites located in the vicinity of
oncogenes constitute hotspots of episomal insertion. Altogether, our results suggest
that reintegration of excised episomal circles constitute a potential source of genomic
instability and cancer in leukemia and lymphoma.
| V(D)J recombination is a unique mechanism of somatic recombination aimed to provide a large
antigen receptor repertoire in T and B cells (for review, see [1] and references therein). During
this process, the variable (V) diversity (D) and joining (J) gene segments present within
the immunoglobulin (IG) and T-cell receptor (TCR) loci, are assembled to form a complete VDJ
exon encoding the variable region of the IG/TCR (Figure 1A). The recombination requires the presence of
specific motifs flanking all the V, D, and J gene segments (12– and
23–recombination signal sequences [RSSs]), and allowing the
recruitment, binding, and proper positioning of the products of the recombination activating
genes 1 and 2 (RAG-1/2). Recent data suggest that in vivo, RAG-1/2 proteins
initiate the rearrangement by performing a first single-strand nick at the exact border
between a 12-RSS and its adjacent coding gene segment [2]. This leads to the capture of a 23-RSS, the
formation of a 12/23 synaptic complex in which the two DNA/protein structures are held in
close juxtaposition, and the generation of another nick at the captured 23-RSS. Within this
complex, RAG-1/2 catalyzes a trans-esterification reaction in which each
liberated hydroxyl group attacks the opposite DNA strand [3]; this generates four broken ends held in a
postcleavage synaptic complex: two blunt RSSs or signal ends (SEs), and two covalently
sealed hairpin coding ends (CEs). The broken ends are then efficiently repaired by the
nonhomologous end-joining (NHEJ) pathway [4–6]; on the one hand, hairpins present at the CEs are resolved through the
Artemis endonuclease activity; when opened at bases off the apex, hairpin opening generates
overhanging flaps, which, if filled in by a DNA polymerase activity, form palindromic (P)
stretches. Nontemplated (N) nucleotides may be added de novo by the terminal
deoxynucleotidyl transferase (TdT), and/or nucleotides may be deleted from the CEs. Ligation
of the processed CEs forms a highly diversified coding joint (CJ). By contrast, the two SEs
present in the synaptic complex undergo only limited processing (some N addition and rare
nucleotide deletion) before joining into signal joints (SJs). While CJs give rise to the
functional recombination products, the SJs are merely the byproduct of V(D)J recombination.
SJs have until recently been assumed to be “harmless” and irrelevant in
the dynamics of the V(D)J recombination process, but evidence now starts to accumulate that
this is not the case, and that they play unsuspected roles in events which might compromise
genomic integrity [7,8]. SJs are indeed constituted of
two functional RSSs fused back to back, each of which therefore potentially capable of
further V(D)J rearrangement in presence of RAG-1/2. The issue of SJ reactivity was initially
addressed ex vivo by the use of integrated minilocus and transient extrachromosomal
recombination substrates containing germline gene segments flanked by their RSSs, and
undergoing rearrangement in culture [9–11].
Both integrative and extrachromosomal experiments indicated that, following a first
rearrangement by inversion, the SJ produced was indeed reactive, and could engage into
further cycles of rearrangement with RSS partners in cis (similar to Figure 1C and 1D). In vivo and ex vivo observations have revealed that
the products resulting from such secondary SJ rearrangements consist of one new SJ and one
hybrid RSS/coding-segment junction (hybrid joint [HJ]), albeit with the
molecular features of a CJ (i.e., with N nucleotide insertion, and extensive nucleotide
deletion and P nucleotide addition at both the RSS and coding segment sides; Figure 1D) [8–10,12]. This
junction, which we refer to as a “pseudo-hybrid” joint (ΨHJ), is
thereby morphologically distinguishable from CJs, SJs, and to a large extent from
“genuine” HJs [13–18].
ΨHJs constitute therefore specific signatures of such ongoing SJ rearrangement
events. Interestingly, recent in vivo data suggest that IGK/IGL rearrangement hierarchy and
isotypic exclusion might in part be achieved by ongoing SJ recombination [12]. Thus, SJ reactivity might
have also evolved as part of the dynamics of the V(D)J rearrangement process. Eventually,
the pathological counterpart of this possible physiological extension of the V(D)J
recombination capability has also been shown to occur in cases of oncogenic chromosomal
translocation, in which ongoing rearrangement of the resulting chromosomal SJ (CSJ)
constitutes the source of oncogene activation [8].
In the normal process of V(D)J recombination, the large majority of SJs produced is however
not retained on the chromosome, but excised on episomal circles (ECs; Figure 1A). Because ex vivo RAG binding (or rebinding) also
efficiently takes place on episomal SJs (ESJs), leading to SJ recleavage and, at least in
vitro, to RAG transposition [7], we reasoned that ongoing trans-V(D)J recombination
might also and concurrently occur (Figure
1B). This might result in the same type of insertion of the whole
circle into the genome as previously observed in vivo for RAG-mediated transposition
[19], with the
important difference that it would in this case employ trans-V(D)J
recombination [20–23], a
process potentially more efficient than RAG transposition [15,24–28].
Both mechanisms might obviously lead to similar genomic instability events, including
oncogenic activation/deregulation. In this report, we investigated V(D)J-mediated ESJ
insertion as an additional potent source of genomic instability and oncogenic deregulation
in lymphoid cells.
To investigate the possibility that excised episomes can reintegrate the genome through
ongoing recombination of the SJ, we first assessed the ability of an ESJ to undergo
trans-V(D)J recombination in an ex vivo
trans-recombination substrate assay [20] (Figure 2A). Three different human ESJs and two standard
RSSs were cloned in separate extrachromosomal recombination substrates; two genuine SJs
were used as “donor” ESJ plasmids: Jδ1/Dδ3 and
Ki/Jκ3 [12]; furthermore, an artificial Dβ1Δ ESJ was generated
by mutagenesis deletion of the 12-bp Dβ1 coding sequence located between the
Dβ1 5′ and 3′ RSSs; the human Jβ2.7 and
VκA27 gene segments were used as 12-RSS “target” substrates.
NIH3T3 fibroblasts were cotransfected with donor and target plasmids either with or
without RAG-1/2 and TdT expression vectors. Bulk plasmid DNA was recovered after 48 h of
culture, and junctions resulting from trans-recombination between the ESJ
and the target RSS were amplified in a single round PCR. Primer combinations were designed
to detect the 2 expected integration breakpoints complying with a 12/23 synapsis:
combination (3 + 2) was used to detect putative SJs, and combination (1
+ 4) was used to detect putative ΨHJs (Figure 2A). PCR products were revealed by an
IRD800-labeled primer extension (PE) assay allowing a precise to-the-base resolution of
the amplified species (see Materials and Methods).
In absence of the RAGs, faint and nonrecurrent PCR products scattered at various sizes
were obtained for both primer combinations (e.g., Figure 2B, ΨHJ T2, ~160 bp; additional bands are
also present outside the visualized parts of the gels). To assess the identity of such
junctions, double-nested secondary PCR were performed, and the amplification products were
cloned and sequenced. Sequence analysis confirmed the occurrence of junctions that were
not generated by RAG-mediated recombination, but rather due to random breaks joined
through NHEJs and/or homologous recombination pathways (not shown). Such junctions,
collectively referred to as “break/repair,” have been previously
described, and represent the RAG-independent recombination background of the
trans-V(D)J assay [29].
In presence of RAGs, however, more intense and recurrent PCR products of the expected
sizes were obtained for both PCR combinations (typical examples are illustrated for the
Dβ1Δ × Jβ2.7 combination in Figure 2B). SJ primer combination (3 + 2)
usually displayed one major band and few minor species around that position, as expected
from a standard SJ, which generally presents limited nucleotide processing. ΨHJ
primer combination (1 + 4), on the other hand, displayed a pattern of intense
bands around the expected size, representing the typical spectra of largely processed
junctions. Sequence analysis of cloned products from double-nested secondary PCR confirmed
the presence of the two specific breakpoints expected from V(D)J-mediated ongoing ESJ
recombination in most cases (Figures 3
and 4; also, see below). Thus, and as
anticipated from previous studies describing ongoing recombination of SJ with targets in
cis, our results demonstrate that ESJs are also capable of ongoing
efficient RAG-mediated recombination with RSS targets in trans in the
context of a 12/23 synapsis. However, as the ESJ is formed by a functional 12-RSS and a
functional 23-RSS, both potentially able to bind the RAGs, we next wondered if this
particular structure might allow to bypass the 12/23 rule for synapsis and give rise to
additional recombination products that we would fail to detect with the two primer
combinations used above. Double-nested PCR with the two complementary primer combinations
(1 + 3) and (2 + 4) (Figure
2A) corresponding to a 12/12 synapsis were thus performed on the same bulk DNA.
Such combinations, however, gave rise to only weak amplification products. Cloning and
sequencing confirmed in most cases the occurrence of the symmetrical 12/12 SJ (1
+ 3) and ΨHJ (2 + 4) (Figure S1A). This suggests that although a fraction of
the trans-recombination can occur in violation of the 12/23 rule (as in
normal V(D)J recombination), this represents a minor population, and the large majority of
the RAG-mediated recombinants are detected with primer combinations (3 + 2) and
(1 + 4), in accordance with a 12/23 synapsis between the ESJ and its RSS target.
Accordingly, the use of ESJ donors made of two 12-RSS in the context of a 12-RSS target
did not give rise to any specific amplified signal, while each of the two RSS from an ESJ
donor made of two 23-RSS gave rise to efficient recombination in the context of the same
12-RSS target (Figure
S1B). In addition, the use of ESJ donor constructs made of one functional and one
nonamerless RSS through deletional mutagenesis showed that while deletion of the nonamer
from the reactive 23-RSS completely abolished recombination, deletion of the nonamer from
the bystander 12-RSS did not modify the overall efficiency (Figure S1C).
Altogether, our data clearly indicate that trans-V(D)J recombination of
ESJ obeys the 12/23 rule and is not dependent on alternative mechanisms such as ESJ
opening by nick–nick (see below for a potential role of nick–nick in
the generation of some junctions).
We next analyzed in details the sequences issued from recombination of the ESJ (Figure 3 and 4). SJs obtained were morphologically undistinguishable
from standard SJs, with the presence of some N insertion and limited nucleotide deletion
(Figure 3; 3T3). Likewise, detailed
analysis of the ΨHJ revealed that in most sequences, end processing similar to
that of a standard CJ (nucleotide deletion, N and P nucleotide addition) occurred at both
the SE (the bystander 12-RSS of the ESJ) and the CE sides of the joint (Figure 4; 3T3). This strongly suggests
that once engaged in synapsis, and despite its ability to bind the RAGs, the bystander RSS
of the ESJ mostly behaves as a coding segment, and undergoes therefore the hairpin
formation step, followed by hairpin resolution and further processing (Figure S2, left
panel). Nevertheless, the presence of sequences containing a full-size ESJ bystander RSS
(e.g., Jδ1; Figure 4)
together with the ambiguity in the assignment of P nucleotides in presence of TdT,
suggested possible RAG binding on both ESJ RSSs, and involvement of alternative pathways
of RAG-mediated recombination. To confirm the involvement of trans-V(D)J
recombination and to test the potential contribution of alternative pathways, the ESJ
assay was repeated in the GUETEL Artemis-deficient cell line [30]. During standard V(D)J
recombination, failure to resolve hairpin CEs in absence of the Artemis endonuclease
results in the absence of CJ formation without impeding SJ formation [31–33]. Similarly, the absence of Artemis in the
ESJ assay should prevent ΨHJ formation without hindering the generation of SJs.
Following cotransfection of the GUETEL cell line with the
(Ki/Jκ3)ESJ/VκA27 couple, PCR was performed as above and the
amplification products cloned and sequenced. As expected, SJs were obtained in absence of
Artemis and were undistinguishable from SJs obtained both in the Artemis-proficient 3T3
cells and in the GUETEL cell line complemented with Artemis (GUETEL-A; [30] and Figure 3). Despite the absence of Artemis, weak
amplification products were also obtained for the ΨHJ PCR combination (1
+ 4). However, the sequenced junctions displayed virtual absence of N insertion
and nucleotide processing (Figure 4),
in sharp contrast to the morphology of ΨHJs obtained in Artemis-proficient cells
(3T3 and GUETEL-A). The features of the Artemis−/−
junctions are however strongly reminiscent of HJs generated by the “RAG-mediated
joining” pathway [15–18].
RAG-mediated joining is an NHEJ-independent pathway related to RAG transposition in which
a direct attack of a free 3′ hydroxyl group from the SE into the hairpinned CE
bypasses the hairpin resolution step (illustrated in Figure S2, right
panel); this usually results in the generation of a class of HJs displaying a full-size
RSS joined to a coding sequence with limited processing (depending on the position of the
attack in the hairpin). The presence of such junctions in the absence of Artemis suggests
that both trans-V(D)J recombination and RAG-mediated joining concurrently
occur to generate ΨHJs and HJs, respectively. However, the virtual absence of such
junctions in Artemis-proficient cells (3T3, Artemis-complemented) strongly suggests that
RAG-mediated joining constitutes a minor recombination pathway compared to
trans-V(D)J recombination. Thus, the junctions with a full-size ESJ
bystander RSS initially observed in Artemis-proficient cells are Artemis dependent, and
are either ΨHJs in which processing is limited due to RAG-binding on the bystander
RSS, or HJs generated through “RSS swapping” (Figure S2, middle
panel).
Altogether, our results indicate therefore that standard trans-V(D)J
recombination is the major pathway of RAG-mediated ESJs ongoing rearrangement.
How efficient is trans-V(D)J recombination of ESJs? In the case of
recombination between two standard coding-segment RSSs, the frequency of
trans-V(D)J recombination has been previously shown to be reduced
compared to cis-V(D)J recombination events [20,21]. In the present case of recombination
between an ESJ and its RSS target, the presence of the second RSS in the SJ could
structurally impede—or on the contrary stimulate—RAG fixation and/or
activity on the reactive one, and modify the overall recombination efficiency. We
therefore compared in the ex vivo assay the efficiency of trans-V(D)J
recombination of an ESJ and a RSS target with the efficiency of
trans-V(D)J recombination between two standard coding-segment RSSs (Figure 5A). To do so, we tested two pairs
of plasmids: (Jδ1/Dδ3) ESJ × Jβ2.7 versus
Dδ3 × Jβ2.7, and (Dβ1ΔESJ ×
Jβ2.7 versus Dβ1 × Jβ2.7. Following transfection and
harvesting carried out as above, semiquantitative primary PCR amplification of the
breakpoints was performed, and serial dilutions were revealed by PE. As illustrated in
Figure 5B for the
(Jδ1/Dδ3) ESJ × Jβ2.7 versus Dδ3
× Jβ2.7 couples, no significant difference could be seen in the
formation rate of SJs in the presence of an ESJ. Similarly, results showed comparable
rates in the formation of a ΨHJ relative to a CJ. We conclude that ESJs are at
least as efficient as standard RSSs to undergo trans-V(D)J recombination.
Altogether, this suggests that the presence of the bystander RSS does not impede the
recombination process, whether structurally (through steric constraints) or functionally
(through nick–nick activity). From the mechanistic point of view, this data
predicts therefore that in vivo, V(D)J-mediated reintegration of ESJs (Figure 1B) should not be different from
V(D)J-mediated translocation (Figure
1C); most important, both processes should use the same RSS targets with the same
efficiency.
V(D)J-mediated translocations have been shown to occur not only between authentic RSSs
from distinct IG/TCR loci, but also between authentic RSSs and fortuitous sequences in the
genome resembling a RSS (cryptic RSSs) [34]. The mistargeting of the RAGs towards
cryptic RSSs located in the vicinity of a silent proto-oncogene is a recurrent source of
genomic instability and oncogenesis. Erroneous targeting of cryptic sites located near the
LMO2 and TAL2 proto-oncogenes in t(11;14)(p13;q11) and
t(7;9)(q34;q32) translocations, respectively, represent prototypical examples of such
oncogenic translocations in T-cell acute lymphoblastic leukemia (T-ALL) [8,29,35–37].
Our data above suggest that in vivo, such cryptic sites might provide efficient targets
for ESJ reintegration. To further define the potential oncogenic properties of episomal
reintegration, we next investigated in our ex vivo assay the capacity of ESJs to target
oncogenic cryptic RSS. The human LMO2 and TAL2 cryptic
RSSs and flanking sequences were cloned in a recombination substrate plasmid (Figure 6A) and assayed in parallel to the
Jβ2.7 segment as a target for the (Jδ1/Dδ3) ESJ, using the
PCR/PE assay described above. Our results show a similar considerable high rate of
V(D)J-mediated recombination of the ESJ with the LMO2 and
TAL2 cryptic RSSs than with the Jβ2.7 authentic RSS (Figure 6A). To further estimate the
likelihood in vivo of SJ insertion in such cryptic RSS, the
LMO2 versus the HPRT intron 1 region were assayed as
competitive targets for the (Jδ1/Dδ3 ESJ (Figure 6B). The HPRT intron 1 region was
chosen as a competitor because it contains a well-described cryptic RSS classically
involved in illegitimate V(D)J-mediated deletion of exons 2–3 in vivo
[38], and has
also been identified as a site of ESJ reintegration in vivo ([39]; see Discussion) as well as a region of RAG transposition in vivo
[19]. While
recombination with the LMO2 cryptic RSS constituted a hotspot of
integration, no recombination could be observed in the HPRT region,
neither specifically at the cryptic RSS, nor in the surrounding sequences (Figure 6B). Similar results were obtained
when using a target plasmid containing a second copy of the HPRT fragment
in a head-to-tail orientation, and forming a cruciform structure (not shown). Altogether,
these data strongly suggest that cryptic sites such as LMO2 with much
higher recombinogenic potential than the HPRT intron 1 cryptic RSS might
also be targeted in vivo by ESJ insertion and could potentially lead to oncogenic
activation.
We next investigated the physiological relevance of the reintegration of excised episomes
through ongoing SJ recombination. As a first approach, we sought to assess if
trans-chromosomal ΨHJ breakpoints were present and detectable
in vivo, and if so, to estimate their formation rate. To do so, we designed primer
combinations allowing the amplification of trans-TCR ΨHJ in mouse
thymocyte DNA (Table 1). Additional
combinations designed to amplify trans-TCR CJs and
trans-TCR SJs were also performed as reference. To detect such relatively
rare trans-TCR recombination events, we used a sensitive fluctuation PCR
assay allowing the detection of less than one recombination event in a million cells (see
Materials and Methods). Positive PCR-amplification
replicates were obtained for all trans-TCR combinations in the broad
range of ~1 in 1,000,000 cells to 1 in 10,000 cells. Sequence analysis of the
PCR-amplification products was carried out to assess the identity of the junctions, and
revealed the presence of ΨHJs (Figure S3), CJs, and SJs (not shown). Importantly, the
ΨHJs obtained in vivo displayed the same molecular features as observed in the ex
vivo assay, with extensive nucleotide processing on both sides of the junctions, and N/P
addition. These results clearly demonstrate that ΨHJ breakpoints are indeed
generated in vivo. Furthermore, such junctions are readily detectable in mouse thymocytes
at a range similar to that of equivalent trans-CSJs and CJs. In line with
our ex vivo results, this indicates that in vivo, SJs undergo efficient ongoing
cis- and/or trans-rearrangement with RSS targets in
the genome with surprisingly high frequency. This suggests further that in presence of
RAGs, CSJs and/or ESJs are indeed very recombinogenic structures.
Although we demonstrated above that trans-chromosomal ΨHJs are
readily detectable in vivo, such breakpoints do not represent exclusive signatures of ESJ
reintegration, as they may also derive from ongoing recombination of
trans-CSJs with neighboring gene segments in cis
(compare Figure 1B and 1D). To ensure that the ΨHJs
observed above did not exclusively derive from CSJs issued from trans-TCR
translocations, and to estimate the rate of episomal reintegration in vivo, we generated
double mutant mice in which the generation of trans-CSJs is abolished
(Figure 7).
Eβ−/− knockout mice, in which deletion of the
560-bp Eβ core generates an >100-fold reduction in TCRβ
rearrangements [40]
and TCRδ/β translocations (Table 1), were crossed with a Dβ1GFP knockin
mouse, in which the introduction of the GFP in the Dβ1 gene segment and flanking
23-RSS abolishes Dβ1-Jβ1/2 rearrangements (SS, OC, PF, unpublished
data). In the (Dβ1GFP × Eβ−)
double-mutant (DE) mice, all TCRβ chains are consequently produced via a
Vβ-Dβ2-Jβ2 rearrangement from the Dβ1GFP
allele (unpublished data), and all TCR excision circles issued from
Dβ-Jβ rearrangements carry a (Dβ2Jβ2) ESJ (Figure 7A). Because one allele is knocked
out for Eβ, and the other produces a TCRβ chain, translocations to
TCRβ cannot be present in TCRαβ+ cells
from DE mice. In absence of translocation to TCRβ, no
trans-TCRβδ CSJ is produced, and all
Jβ2.7/Jδ1 ΨHJs are consequently issued from episomal
reintegration. As shown in Table 1,
while the rate of Jβ2.7/Jδ1 ΨHJs was decreased over ~40-fold in
total thymocytes from the Eβ−/− mice compared to
WT, the rate of Jβ2.7/Jδ1 ΨHJs was only decreased 1.2-fold in
sorted TCRαβ+ cells from DE mice compared to WT.
Altogether, and in agreement with the ex vivo data, these results clearly demonstrate
that V(D)J-mediated episomal reintegration is indeed occurring in vivo, and at a rate
comparable to that of V(D)J-mediated chromosomal translocation.
During our screen of trans-TCR SJs described above, we noticed the
presence of amplification products larger than expected. Out of a total of 310 PCR
replicates of seven distinct trans-TCRδ/β SJ
combinations, 120 were PCR positive and four were of unexpected larger size (unpublished
data). Sequencing of the cloned products revealed that large amplicons resulted from
cis-V(D)J recombination to a cryptic site in one case, and to a
downstream gene segment in another case; eventually, two of the four cases showed the
insertion of a sequence from the TCRδ locus with features compatible with
RAG-mediated ESJ reintegration. In one of these two cases, a ~900-bp
Jδ1-Dδ2 fragment was inserted into a Jβ2.7/Vδ2 CSJ
target (Figure 8A). The breakpoints
consisted of a perfect SJ on the right arm of the insertion; on the left arm, a junction
compatible with a ΨHJ was found, displaying a deletion of three nucleotides on one
side of the joint, a deletion of one nucleotide on the other side, and one N nucleotide
addition; in the second case, the same ~900-bp Jδ1-Dδ2 fragment was
observed inserted into a Dβ1/Vδ2 CSJ target (Figure 8B). Similarly, one of the breakpoints displayed a
perfect SJ, and the other breakpoint consisted of an SJ with 2 N insertions. Considering
the structure of the target, this last junction was ambiguous to assign as an SJ or a
ΨHJ, and could have occurred through nick–nick, V(D)J, and/or
RAG-mediated joining. Nevertheless, all junctions complied with the 12/23 rule and with
features of RAG-mediated breaks, and it is therefore very likely that both cases represent
in vivo examples of RAG-mediated reintegration of a Jδ1-Dδ2 ESJ. In
the two cases, both ESJ reintegration and translocation events occurred, and two
mechanisms could therefore account for their formation: either the
trans-TCR translocations occurred first, and provided a
trans-TCR SJ target for ESJ reintegration; or, alternatively, ESJ
reintegration in a standard RSS target (e.g., Vδ2) could have occurred first,
and provided a trans-TCR structure and/or breakpoints prone to
V(D)J-mediated translocation. Intriguingly, the actual frequency of this double
translocation/insertion event (2/120 = ~10−2
trans-recombination events) was not compatible with the expected combined
frequency of each independent event: (trans-TCR translocation
~10−4–10−6 × reintegration
~10−4–10−6 =
~10−8–10−12). Furthermore, as the
assay is limited to PCR-amplifiable sizes of the inserted fragment (such as the relatively
short Jδ1-Dδ2 episome), the apparent rate of reintegration is probably
vastly underestimated. This strongly suggests that the two events are linked, either
because CSJs constitute preferential targets for ESJ reintegration, or because ESJ
reintegration generates genomic instability leading to further chromosomal abnormalities;
such abnormalities could consist of chromosomal translocations as detected here, or
potentially more complex events which could not be detected in the present in vivo assay.
Altogether, these data demonstrate that in vivo, SJs located on excised ECs may be
reintegrated into the genome through trans-V(D)J recombination using
standard and cryptic RSS targets, and further suggest that this event is associated with
additional genomic instability.
During B- and T-cell ontogeny, ECs are sequentially excised from the genome of lymphoid
cells as a result of the hierarchically regulated D-to-J, V-to-DJ, and V-to-J rearrangement
events. Such episomes are believed to be nonreplicative, but have nevertheless been shown to
be surprisingly stable structures, persisting until diluted out by cell divisions
[41,42]. Such episomes carry an ESJ comprised of two
functional RSSs, each of which are susceptible to (re-)bind the RAG proteins when
(re-)expressed at the various cell maturation steps until final downregulation
[7,41,43]. On the other hand, trans-V(D)J recombination
occurring between RSSs or cryptic RSSs located on distinct chromosomes is a relatively
common event in developing lymphocytes [22,23,44], which has
been shown to lead to recurrent oncogenic translocations [34]. Here we show both ex vivo and in vivo that
ECs may be reintegrated into the genome through trans-V(D)J recombination
occurring between the ESJ and an IG/TCR or one of the many cryptic RSS targets scattered in
the genome.
We have demonstrated in ex vivo assays that the efficiency of
trans-V(D)J recombination of an ESJ with a RSS target is not
quantitatively different from the trans-V(D)J recombination occurring
between two RSSs. This is somehow unexpected, because its particular structure confers
additional properties to the ESJ. ESJs are efficiently cleaved ex vivo and in vitro by the
nick–nick mechanism, a symmetrical nick occurring 5′ of each RSS
(simultaneously or sequentially) and generating two flush SEs ending with 3′
hydroxyl groups [7]. Remarkably, this process bypasses both the formation of a hairpin
intermediate and the need of synapsis with another RSS. In the context of ongoing SJ
recombination, one could think that efficient nick–nick opening of lone ESJs
upon RAG binding might prevent the occurrence of synapsis with a RSS target, and thus
considerably reduce the overall frequency of trans-V(D)J rearrangement.
Our ex vivo data clearly argue against this assumption, and suggest that
trans-V(D)J rearrangement is largely independent of nick–nick.
This however does not preclude that ESJ opening by nick–nick might occasionally
participate in the synapsis, and the recent in vivo demonstration of synapsis by capture
[2] provides a
plausible two-step scenario of the occurrence of nick–nick within a 12/23
synapse (Figure
S4).
Incidentally, our data provides additional evidence that HJs can be formed ex vivo
through both NHEJ-dependent and NHEJ-independent pathways (Figure S2).
RAG-mediated joining has been initially proposed as an efficient alternative pathway of
NHEJ-independent HJ formation [16–18].
However, conflicting data have been reported on the relevance of this pathway ex vivo and
in vivo [15,25–27,45,46]. In
particular, recent data indicated that at least ex vivo, RAG-mediated joining is seldom
observed in presence of full-length RAGs, even in the absence of the concurrent NHEJ
pathway [15,46]. We find here the presence
of HJ with features of RAG-mediated joining in the absence, but not in the presence of
Artemis, suggesting that at least ex vivo, HJ formation is a mixed process consisting of
efficient RSS “swapping” and inefficient RAG-mediated joining.
Interestingly, it is possible that the frequency of RAG-mediated joining would be favored
in our assay, due to the participation of an ESJ. As mentioned above, nick–nick
opening of the ESJ in the context of a 12/23 synaptic complex would generate two flush SEs
ending with 3′ hydroxyl groups. While the two reactive RSSs might be sequestered
by tight binding in the SE postcleavage complex (Figure S4), this additional free 3′ OH would
provide an available substrate for hairpin attack, at least in absence of concurrent
hairpin resolution by Artemis. Thus, nick–nick opening might participate in both
trans-V(D)J recombination and RAG-mediated joining.
Using WT and DE double-mutant mice, we could estimate that ESJ reintegration in authentic
TCR RSS targets occurs in vivo in the broad range of 1 in 1,000,000 to 1 in 10,000 mouse
thymocytes, depending on the integration site. In agreement with our results, recent data
provide further evidence for the existence of V(D)J-mediated ESJ reintegration in vivo,
and concur with the idea that it constitutes a significant source of genomic instability.
Using a screen for HPRT mutants, Finette and colleagues recently
identified a case in which a Vα/Jα episome was found inserted in the
cryptic 23-RSS from HPRT intron 1 [39]. This finding is all the more remarkable
given that this HPRT cryptic site has been shown to exhibit a very low
recombinogenic potential in functional assays ([47], our unpublished data), and provides direct
evidence that in vivo, human cryptic RSSs constitute efficient targets for ESJ
reintegration. In the same line, using a double selection assay allowing for the recovery
and quantification of excision/reinsertion events in a pre–B-cell line, Reddy et
al. recently identified three cases of V(D)J-mediated reintegration [48]. In full agreement with our
in vivo estimation, they evaluated a rate of one reintegration out of every 100,000 V(D)J
recombinations. Considering the average number of V(D)J recombination per B or T cell,
these data suggest that the daily lymphocyte output in human could be accompanied by as
many as 5,000 ESJ reintegration events.
Authentic RSSs from IG/TCR must without doubt constitute preferential targets for such
reintegration. However, there are an estimated 10 million functional cryptic sites
dispersed throughout the human genome [34,47], some of them displaying much higher recombinogenic potential than
the HPRT intron 1 cryptic RSS in which ESJ reintegration was found
[29,37]. Most important, some of them are already
known to provide recurrent targets for oncogenic (type 1) V(D)J-mediated translocations.
In T-ALL, for example, the mistargeting of the RAGs towards such cryptic RSSs located in
the vicinity of a silent proto-oncogene is a recurrent source of genomic instability and
oncogenesis [34].
We demonstrate in our ex vivo assay that cryptic sites involved in oncogenic
V(D)J-mediated translocations are also hotspots for ESJ recombination; it seems thus
reasonable to think that in vivo, such cryptic RSSs might provide as efficient targets for
ESJ reintegration than for chromosomal translocations. ESJ reintegration could lead to
similar oncogenic activation/deregulation, either through the insertion of active immune
regulatory elements (e.g., the TCRδ enhancer excised during
ΔRec/ΨJα or Vα/Jα rearrangements
[42,49], or the IGK enhancer excised during KDE/Ki
rearrangements [12,50]) or through the disruption
of locus silencing (e.g., the negative regulatory element upstream of
LMO2 [51,52], WA Dik; BN
et al., unpublished data).
If ESJs are at least as efficient as any standard coding-segment RSS to undergo V(D)J
recombination with a RSS target in trans, the two processes are nearly
identical mechanistically, and the RSS/cryptic RSS targets are the same, why then has
V(D)J-mediated oncogenic translocation been largely documented in the literature, but
V(D)J-mediated reintegration never been reported up to now in lymphoid malignancies?
Higher-order spatial genome organization is a contributing factor in the formation of
recurrent translocations [53]. In contrast, increased mobility of the EC might favor interactions
and recombination of ESJs, with targets located in parts of the genome generally
segregated. Oncogenic targets might thus be distinct for reintegration and translocation.
However, due to the multiplicity and size range (from <1 Kb to several Mb) of the
excised episomes, as well as the diversity of insertion targets, reintegration events are
very unlikely to be detected by routine analysis. Particularly, and in contrast to
chromosomal translocations, episomal reintegration is invisible by standard karyotype
analysis [54,55]. Few examples of identified
class-switch recombination– and RAG-mediated episomal reintegration in lymphoid
neoplasia provide proof of principle that episomal reintegration can indeed lead to
cancer, and illustrate the complexity and unlikelihood of identifying such events without
specifically designed screens [55–57].
For example, using complex DNA fiber-FISH and three-color interphase FISH techniques on
samples from follicular lymphoma patients devoid of the hallmark t(14;18)(q32;q21)
translocation, Vaandrager et al. identified two cases in which the BCL2
gene was excised from 18q21 and inserted into the IGH locus at 14q32 [58]. The relatively high
frequency (5%) of such events, discovered relatively recently by Vaandrager et
al. out of an abundantly studied and characterized pathology, illustrate well the extent
to which insertional events might be missed by routine cytogenetic and molecular analysis.
Another possibility which could account for the rarity of cases of episomal reintegration
observed in lymphoid neoplasia is that episomal reinsertions might generate unstable
transitory structures in the genome, leading to additional aberration events and more
complex genomic configuration. This possibility is supported by the unexpected high
frequency of two in vivo cases observed in this study, combining ESJ reintegration and
translocation events. Similar examples of genomic instability following reintegration of
episomal structures have been previously documented. In mouse plasmacytoma, Kovalchuk and
colleagues have shown that class-switch–mediated Eμ/Sμ episomal
reintegration into c-MYC favors t(12;15) translocation [56]. Although the ground of such instability is
not yet clear, an obvious possibility is that a fraction of the recombination events might
lead to incomplete insertion in vivo. Alternatively, the reinserted episomes might in vivo
retain their initial “open” chromatinized structures, and provide
preferential accessible targets for the recombination machinery. The V(D)J-mediated
instability of reinserted ESJs would be especially relevant in neoplastic cells such as
T-ALLs, in which arrest of differentiation at early stages of T-cell development leads to
sustained RAG levels.
Gene profiling studies have shown that a substantial fraction of the T-ALL cases display
oncogenic activation in absence of detectable chromosomal alterations [59], suggesting the presence of
alternative pathways of oncogenesis. Some of them have indeed been recently discovered,
and involve episomal structures [60]. Considering the large number of ESJs produced daily and the
mechanistic similarities between ESJ reintegration and oncogenic translocations, our data
suggest that reintegration of excised ECs by the V(D)J recombinase might also account for
some of these cases, and constitute an additional potent source of genomic instability.
Definitive answer to this open question will however await large-scale screens of human
lymphoid cancer samples with specifically adapted strategies.
Recombination substrates were derived from a series previously described [29], and the constructs are
summarized in Figure
S5. The various gene segments containing the regions to recombine were amplified
from human DNA with appropriate tailed primers and cloned in the recombination substrate
using unique restriction sites (Mlu1, Sac2, Not1). The (Ki/Jκ3) ESJ was cloned
(Not1/BamH1) in the pPCR-scriptAmp vector (Stratagene, http://www.stratagene.com). Some
constructs were flanked by “tag” sequences, which were previously
tested to be devoid of functional cryptic RSSs, and in which specific PCR primers were
designed. NIH 3T3 Swiss mouse fibroblasts, and human Artemis-deficient GUETEL or
Artemis-complemented GUETEL-A cell lines [30] (generously supplied by J.-P. de Villartay)
were cultured in standard conditions (DMEM/10% FCS). Cells (2 ×
106) were transfected with 3 μg of each recombination substrate, 2
μg pEBB-RAG1, 2 μg pEBB-RAG2 expression vectors (generously supplied by C.
Roman and S. Cherry), and 2 μg pCDNA3TdT expression vector (TdT
cDNA from the pTDT expression vector [a generous gift from N. Doyen]
recloned into pcDNA3 [Invitrogen, http://www.invitrogen.com])
using Superfect (Qiagen, http://www.qiagen.com) according to the instructions recommended by the
manufacturer. Transfected cells were transferred to 10 ml DMEM supplemented with
10% FCS and cultured for 48 h. Cells were subsequently trypsinized, and
plasmids recovered by alkaline lysis and phenol-chloroform extraction as previously
described [29]. To
test the possibility that trans-rearrangements could occur through a
first step of homologous recombination between the identical core regions of the
donor/acceptor plasmids, the trans-recombination assay was also performed
with “coreless” excised linear fragments carrying the RSS target, and
gave rise to similar results (Figure S6). Although we cannot formally exclude the
possibility that when present, stretches of homology can facilitate the recombination
process both ex vivo and in vivo, our data suggests that recombination efficiently occurs
through direct trans-V(D)J synapsis as previously assumed for the assay
[20].
Breakpoints were PCR amplified from 1 μl (1/20) harvested bulk DNA with
appropriate primers (summarized in Table S1) in the following conditions: 4 min at
94°C for 25 cycles (30 s at 94°C, 30 s at 64°C, and 30 s at
72°C), and 7 min at 72°C. Secondary double-nested PCRs were performed in
the same conditions, and the amplification products cloned and sequenced as previously
described [29].
Breakpoints were amplified in a single round PCR from 1 μl (1/20) harvested bulk
DNA with appropriate primers (summarized in Table S2) in the same conditions as above, and otherwise
undetectable amplification products were revealed by PE. PE is a sensitive alternative to
Southern blot consisting of several cycles of DNA polymerization extending from a labeled
primer until the end of a matrix DNA fragment (here a PCR product). This generates linear
(nonexponential) accumulation of labeled fragments of specific size. PE assays were
performed on 1.5 μl, 1.0 μl, and 0.75 μl of primary PCR using a
nested IR800-labeled primer in the following conditions: 5 min at 95°C for 20
cycles (30 s at 95°C, 15 s at 60°C and 1 min at 70°C), and 1 min
at 70°C with the EXCEL II kit (Epicentre Biotechnologies, http://epicentre.com),
using an equal amount of all dNTPs and omitting ddNTPs. A portion (1.2 μl) of the
reaction was used for running on a Li-COR 4200 DNA sequencer (http://www.licor.com).
A sequencing reaction was performed in parallel on the appropriate unrecombined purified
plasmid, using the same IR800-labeled primer and in the same reaction conditions (at the
exception of the dNTP/ddNTP mix), and was run on the same gel, providing both a precise
to-the-base size marker and a positive control for the reaction. Semiquantitative
conditions were calibrated on serial dilutions of bulk DNA and of PCR amplifications
(Figure
S7A–S7C). Furthermore, to exclude possible bias due to difference in the efficiency of
the PCR primer combinations used in the semiquantitative experiments, each comparison
using different sets of primers was calibrated (Figure S7D–S7E). All PCR/PE
assays were performed on at least four independent transfections with similar results.
Thymocyte preparation and cell sorting were performed as described previously
[61].
Phycoerythrin-conjugated mAb against TCRβ (H57–597), purchased from BD
PharMingen (http://www.bdbiosciences.com), was used for cell sorting of
TCRβ+ thymocytes. The sorting windows were defined in such
a way that only cells expressing high levels of TCRβ were purified.
The principle of the assay has been previously described [29]. Detection of rare events by sensitive
double-nested PCR gives rise to fluctuation in target amplification, depending on the
presence or not of the event in the aliquot taken from the sample. Junctions were PCR
amplified from multiple replicates of 2 μg DNA (or 50 ng for positive replicates
out of the fluctuation range) isolated either from total thymocytes of WT or
Eβ−/− mice [40], or from
TCRαβ+ sorted thymocytes from WT or DE
double-mutant mice, using appropriate primers (see list in Protocol S1), and in
the following conditions: 4 min at 94°C for 30 cycles (30 s at 94°C, 30 s
at 64°C, and 30 s at 72°C), and 7 min at 72°C. Secondary
double-nested PCRs were performed on 1 μl of each primary PCR replicate in the same
conditions, and amplification products were cloned and sequenced as previously described
[29]. Frequencies
were calculated using Poisson assumption as previously described [62].
|
10.1371/journal.pcbi.1003091 | The Ability of Flux Balance Analysis to Predict Evolution of Central Metabolism Scales with the Initial Distance to the Optimum | The most powerful genome-scale framework to model metabolism, flux balance analysis (FBA), is an evolutionary optimality model. It hypothesizes selection upon a proposed optimality criterion in order to predict the set of internal fluxes that would maximize fitness. Here we present a direct test of the optimality assumption underlying FBA by comparing the central metabolic fluxes predicted by multiple criteria to changes measurable by a 13C-labeling method for experimentally-evolved strains. We considered datasets for three Escherichia coli evolution experiments that varied in their length, consistency of environment, and initial optimality. For ten populations that were evolved for 50,000 generations in glucose minimal medium, we observed modest changes in relative fluxes that led to small, but significant decreases in optimality and increased the distance to the predicted optimal flux distribution. In contrast, seven populations evolved on the poor substrate lactate for 900 generations collectively became more optimal and had flux distributions that moved toward predictions. For three pairs of central metabolic knockouts evolved on glucose for 600–800 generations, there was a balance between cases where optimality and flux patterns moved toward or away from FBA predictions. Despite this variation in predictability of changes in central metabolism, two generalities emerged. First, improved growth largely derived from evolved increases in the rate of substrate use. Second, FBA predictions bore out well for the two experiments initiated with ancestors with relatively sub-optimal yield, whereas those begun already quite optimal tended to move somewhat away from predictions. These findings suggest that the tradeoff between rate and yield is surprisingly modest. The observed positive correlation between rate and yield when adaptation initiated further from the optimum resulted in the ability of FBA to use stoichiometric constraints to predict the evolution of metabolism despite selection for rate.
| The most common method of modeling genome-scale metabolism, flux balance analysis, involves using known stoichiometry to define feasible metabolic states and then choosing between these states by proposing that evolution has selected a metabolic flux that optimizes fitness. But does evolution optimize metabolism, and if so, what component of metabolism equates to fitness? We directly tested the underlying assumption of stoichiometric optimality by comparing predicted flux distributions with changes in fluxes that occurred following experimental evolution. Across three experiments ranging in length from a few hundred to fifty thousand generations, we found that substrate uptake – an input to the model – always increased, but supposed optimality criteria such as yield only increased sometimes. Despite this, there was a clear trend. Highly optimal ancestors evolved slightly lower yield in the course of increasing the overall rate, whereas more sub-optimal strains were able to increase both. These results suggest that flux balance analysis is capable of predicting either the initial metabolic behavior of strains or how they will evolve, but not both.
| Systems biology is beginning to provide insight into how interactions within complex networks give rise to the holistic behavior of biological systems, and how natural selection would shape these systems over the course of adaptation. Some mathematical models are made with the goal of translating known parameters of components of a small system into predictions of their function. This approach has been used to predict behavior ranging from the oscillation of natural or engineered genetic regulatory networks [1] to flow through small metabolic networks [2], [3]. For larger, genome-scale networks there is insufficient information to generate direct predictions in the same manner. Instead, one can ask how the system should behave were it to have already been selected to function optimally given tradeoffs between different selective criteria. One use of mechanistically-explicit optimality models is to consider the possible optimality of current biological phenomena, such as the optimality of the genetic code [4] or of the enzymatic properties of RuBisCO [5]. On the other hand, optimality models can also be used directly to predict phenotypic changes in a system that would occur over the course of adaptation, such as the evolution of virulence [6] or enzyme expression [7].
The most broadly applied metabolic modeling framework, Flux Balance Analysis (FBA), is a constraint-based evolutionary optimality model. It quantitatively predicts flux through a metabolic network that will maximize a given criterion thought to represent prior natural selection [8]. At the heart of FBA is a stoichiometric matrix, which is a mathematically transformed list of mass-balanced biochemical reactions that fully describes the known topology of the metabolic network of a cell (or other system). It is further assumed that the cell is in a metabolic steady-state, such that the sum of fluxes in and out of each internal metabolite are balanced. As additional constraints are considered (e.g., maximal flux values, irreversible reactions, biomass composition), this matrix can then be used to help define and constrain the space of feasible flux distributions in the cell. Within this feasible space, linear programing is subsequently used to solve for an optimality criterion -such as maximal biomass per substrate (see below)- to identify a feasible flux distribution that permits that optimum.
Evolutionary optimality models are powerful tools as they make it possible to build intuition about the forces that shape biological diversity. However, as has been pointed out most famously by Gould and Lewontin, they can also be misleading and can foster the wrong intuitions [9]. Optimality models make three assumptions: 1) selection (and not other processes) is the primary evolutionary force shaping a trait of interest, 2) we can identify the criterion upon which selection is acting, and 3) there are not underlying constraints which prevent a trait from being optimized. Optimality models are constructive for understanding the evolution of traits only to the extent that these assumptions can be evaluated.
FBA provides an excellent framework to generate testable hypotheses as to which selective criteria are appropriate for a given set of conditions [10], [11]. In environments such as batch culture, selection acts directly upon growth rate -as well as lag and survival in stationary phase- but not upon yield [12]. The most common optimality criterion for FBA is commonly referred to as maximizing growth rate [11]. Because this is performed by constraining one (or occasionally multiple) substrate uptake rate (S/time), this criterion is fully equivalent to predicting the maximum yield (i.e., BM/S) under the given, user-supplied substrate uptake rate. Since FBA cannot predict absolute rates of substrate uptake used as the key constraint, the question as to whether adaptation would optimize BM/S during batch culture critically depends upon the correlation between growth rate and yield. There are solid theoretical grounds to expect absolute limits to the maximization of both rate and yield of reactions [13], but it is often unclear how close biological systems are to these constraints.
In addition to maximization of biomass, various other cellular objectives have been suggested as alternative selective criteria. These include optimal energetic (rather than biosynthetic) efficiency whereby generation of ATP per substrate (ATP/S), or the minimization of the sum of fluxes (BM/Σv or ATP/Σv). The latter are based upon the rationale that enzymes are costly, and thus a general relationship between enzyme levels and reaction rates (although actually quite weak for any given enzyme, [14]) would lead to selection to minimize the total burden of enzymes needed. Finally it has been suggested that selection acts simultaneously upon multiple, competing criteria, leading cells to inhabit an optimal tradeoff surface known as a Pareto optimum [15], [16]. This approach constructs a surface on which no single criteria can be further increased without reducing another. It is then assumed that evolution pushes biological systems to exist somewhere on this surface. Data from a variety of experiments suggested that cells operate near to the Pareto optimum defined by BM/S, ATP/S, and minimization of Σv [15].
Tests of the predictive capacity of FBA have differed in two ways depending upon: 1) whether there was known or assumed adaptation to the substrate in question, and 2) whether tests were a direct or indirect comparison of predicted internal fluxes to measured fluxes (Table 1). The majority of these tests have been conducted with Escherichia coli, and have assumed past selection on BM/S. The direct tests of FBA compared predicted to observed flux distributions (Figure 1) by taking advantage of empirical data generated by 13C-labeling techniques [17]. Briefly, this method to assay relative metabolic fluxes takes advantage of the fact that the carbon atoms of the growth substrate are shuffled in different ways by alternative metabolic pathways, and that these rearrangements leave a signature in biomass. Using gas chromatography-mass spectrometry (GC-MS) to determine the 13C-labeling of protein-derived amino acids, it becomes possible to infer the flux splits in the metabolic pathways leading to their synthesis [17]–[23]. Notable amongst these tests was a quantitative assessment of the relative merits of a series of optimality criteria (and constraints) in their ability to predict the intracellular fluxes of E. coli measured in six environments [11]. Data for wild-type cultures indicated that ATP/Σv2, BM/S or ATP/S were more predictive depending upon the growth condition; however, in all cases there was still significant variation between predicted and measured fluxes.
A key advance in the use and testing of FBA came from the realization that the best test of an optimality model is to examine whether there is movement toward predicted optimal phenotypes following adaptation under known experimental conditions (Table 1). In a classic paper, populations of E. coli were adapted to various carbon substrates for 100–700 generations [24]. The authors ran FBA for all pairwise constraints of substrate and oxygen uptake to predict the maximal BM/S within those constraints, and what metabolites might be excreted. Remarkably, adaptation on five out of six substrates conformed to the predictions, remaining on or evolving toward a ‘line of optimality’ representing the optimal oxygen to substrate ratio. For only one of these substrates did the population move away from the predicted optimality. A follow-up study further showed that the genes expressed in evolved lines correspond to the fluxes predicted to be active by FBA [25]. Since flux changes are only sometimes well-correlated with gene expression [26], however, it remains unclear whether FBA can predict the change in internal fluxes. Although indirect, these studies have suggested that FBA might reasonably capture the evolutionary forces acting on cellular physiology and hence would be capable of predicting the outcome of evolution [27].
To our knowledge there have been only two studies in which the internal fluxes have been measured for both ancestral and evolved strains grown in a constant environment with a single growth substrate. Both involved rapid, short-term adaptation (<1,000 generations) of E. coli under conditions where the cultures were kept in continual exponential growth in batch culture by using frequent, large dilutions. Hua et al [28] measured fluxes following adaptation to the poorly-utilized substrate lactate, while Fong et al [20] measured fluxes following adaptation of a series of E. coli strains with knockouts (KOs) deleting individual enzymes of major branches of central carbon metabolism (e.g., glycolysis). Interestingly, both studies found rather divergent changes in flux distribution across replicates, and found that most improvement in growth rate was the result of increases in substrate uptake. These studies were not compared to FBA predictions, however, thus it remains unclear whether the assumed optimality criteria improved, or whether observed intracellular fluxes moved toward those predicted with a genome-scale FBA model.
In terms of using experimental evolution to test optimality, the cultures that have had the greatest time to adapt are those from the E. coli long-term experimental evolution (LTEE) populations that have been evolving in the Lenski laboratory for over 50,000 generations [29], [30]. These twelve replicate populations have evolved in minimal medium with glucose since 1988, experiencing 100-fold daily dilutions that result in a short lag phase, nearly seven consecutive generations in exponential phase, and then stationary phase. The LTEE experiment has enabled an unprecedented examination of genotypic and phenotypic change over an extended period of adaptation [29], [31]. Despite starting with a wild-type strain capable of rapid growth on glucose, all populations have increased dramatically in both growth rate and competitive fitness through adaptation in batch culture [32], [33]. It should be noted however, that batch culture inherently incorporates some non-steady state conditions and that improvements in lag or survival may have had pleiotropic consequences for growth. Despite this, here we ask how well FBA predictions align with the evolved changes in these populations. If FBA is unable to predict adaptation to single-nutrient, seasonal batch culture conditions we will not be able to apply it to most laboratory environments, not to mention the variable habitats experienced in nature.
The goal of the current work was to test whether the central metabolic fluxes of replicate populations of E. coli with known selective history in the laboratory evolved in a manner that is predictable by FBA (Figure 1). We compared the fluxes inferred from 13C labeling to the ranges predicted to permit optimal performance and summarize these changes in three ways: the % optimality possible given the inferred fluxes, the minimal distance in flux space between the inferred fluxes and the optimal space of distributions, and a flux-by-flux comparison to see how each flux changed relative to predictions. Testing the ability of optimality criteria to predict adaptation not only provides insight into the mechanisms of evolution, but also represents a critical test of the central optimality assumption of FBA. The LTEE lines began with an ancestor operating at near-optimal BM/S, but the independent populations evolved to use central metabolism less optimally. This was reflected in both a small, but statistically significant, decrease in the % optimal BM/S, and a corresponding increase in the distance from the observed to optimal flux state. In contrast, the seven lactate-evolved populations evolved to increase BM/S and moved closer to an optimal flux distribution. The three pairs of KOs had mixed results in terms of optimality and flux pattern. Overall these results indicate that evolved increases in growth rate largely resulted from increased substrate uptake. Furthermore, ancestral strains operating far from optimal yield evolved as suggested by FBA, whereas those close to the optimum experienced a modest decrease in optimality and evolved to be further from FBA predicted fluxes than their ancestor.
Prior to measuring internal metabolic fluxes, we first examined key growth parameters for one isolate from the 50,000 generation time-point for each of 10 independent LTEE populations (Table S1). Growth rate increased by 45% on average (Table S1), which is concordant with the 16% increase observed in these lines after 2,000 generations [32], and the 20% increase measured after 20,000 generations [33]. All evolved lines also increased their glucose uptake rates (individually significant for 5 of 10 lines: A+3, A−2, A−4, A−5, A−6; T-test, p<0.05, two-sample, equal variance throughout unless noted otherwise, Table S1), with an average increase of 18%. The cell dry weight per gram of glucose also increased by an average of 20% while max OD600 increased by 68%. This did not come from decreasing their excretion of organic acids, however, as acetate production actually increased by an average of 50%. No other excreted ions were observed above our limit of detection of ∼50 µM (Table S1).
In order to determine whether the improved performance of the LTEE isolates was reflected in changes in the relative use of central metabolic pathways, we used 13C-labeling of protein-derived amino acids [17] to infer several key flux ratios in central carbon metabolism (Figure 2A). Often the goal is to extrapolate from the measured flux ratios to calculate the flux for each reaction in a network [15], [23]. For this study, however, we limit our discussion and analyses to the flux ratios themselves, as these represent the actual number of inferences from the 13C-labeling data and thus each cellular branch-point is given equal weight (Text S1). It should be noted that 13C data for the LTEE isolates were analyzed with a program, FiatFlux [17], which is based on a simplified model of central carbon metabolism. This program was used for the previous study comparing alternate optimality criteria mentioned above [11], as well as for obtaining the flux data about the lactate [28] and KO [20] lines we analyze below. Inferences with this commonly used program are less variable than inferences based on larger models [34].
We uncovered statistically significant, but modest variation in the flux ratios of evolved isolates relative to their ancestor (Figure 2B, Table S2). In terms of the overall pattern, a MANOVA test found that flux ratios changed significantly as a function of population (Pillai's Trace = 3.80, p<0.001, Figure S1). Additionally, ANOVA tests on the flux ratios for individual lines found at least one significantly different isolate (p<0.05) for all ratios except two, and all lines had significant change in at least one flux ratio. A joint linear regression of the populations found 22 fluxes that differed from the ancestor at a p≤0.05. The False Discovery Rate (FDR) metric suggests that 18 more significant changes were found than expected by chance, whereas the more conservative Tukey HSD test finds that 10 flux changes remain significant.
A few patterns emerged in terms of the actual fluxes found to have changed in evolved isolates. First, the most parallel change was that a small, but significant portion of glucose was routed through the Entner-Doudoroff pathway (Figure 2, flux 2). In all but one case this was accompanied by a similar decrease in the proportion of carbon flowing through the pentose-phosphate pathway (flux 3). On the other hand, replicate lines evolved in opposite directions for flux through glycolysis (flux 1), and for the fluxes producing oxaloacetate from phosphoenolpyruvate (fluxes 4). Additionally, in all cases there was no significant change in the lower bound of production of pyruvate from malate via malic enzyme (flux 6) across evolved isolates.
As a first step in testing the validity of different optimality criteria, we asked whether the flux ratios observed in evolved isolates led to increased or decreased performance with regard to each criterion (Figure 1A). The ‘% optimality’ can be calculated by comparing the maximum value of a criterion when the model was constrained with the observed flux ratios and substrate uptake rate to the maximum value of the criterion in the absence of the flux ratio constraints. Note that because this metric simply compares values of given optimality criteria rather than a particular set of flux ratios it is not affected by the existence of alternate optima for some fluxes.
There was a slight (0.8%) but significant drop in the average percent optimal biomass production (BM/S; T-test, p = 0.008), with 9 of the 10 evolved lines decreasing relative to the ancestor (Figure 3A). Turning to alternative optimality criteria, we first found that ATP/S did not change significantly (Figure 3D), though unlike all other measures throughout, the output was not normally distributed (Shapiro-Wilk test of residuals, p = 0.002; for rest see Figures S2 and S3). Correspondingly, significance for changes in this criterion was tested with the non-parametric Mann-Whitney-Wilcoxon Rank Sum Test (p = 0.79). BM/Σv and ATP/Σv behaved qualitatively similarly to BM/S and ATP/S, respectively, but as neither change was significant these results are displayed only in supplementary material (Figure S4). Finally, we calculated the nearest possible flux distribution for each evolved isolate to the Pareto optimum, and found that 9 of 10 isolates were further from an optimal tradeoff between criteria than the ancestor (Figure S5).
In order to test the sensitivity of these findings to assumptions made in using FBA, we compared the effect of changing the values used for O2 limitation, maintenance energy, and the possible change in biomass composition that would result from the documented increase in average cell size [12]. None of these modulations changed the qualitative results and generally the default values outperformed the others (Figures S6 and S7). Therefore, the conclusion that adaptation did not lead to an increase in any optimality criterion for the LTEE populations seems rather robust.
We next examined whether the flux distributions we inferred for the LTEE isolates moved toward (or away) from the flux distribution predicted to result from optimizing each criterion. We calculated the distance to the optimal fluxes for each evolved isolate relative to the distance between the ancestor and optimality (Figure 1B). Because the per-substrate criteria (e.g., BM/S, ATP/S) had many equally-optimal flux distributions, we identified the optimal solution that minimized the Euclidean distance from observed flux ratios. Choosing the FBA solution that is the closest to our empirical flux observations should, if anything, bias in favor of FBA.
Beginning with the overall pattern of fluxes, we quantified the log ratio of evolved to ancestral flux distance to their nearest optimum (Figure 1B). Both BM/S and ATP/S predicted optima in the opposite direction of the evolutionary flux movement, and hence evolved lines ended up significantly farther from optima than the ancestor (Figures 3B,E; S2; BM/S, T-test p = 0.0008; ATP/S, T-test, p = 0.0004). In both cases the movement away from the optimum was primarily driven by changes in the flux of oxaloaceate from phosphoenolpyruvate.
Turning to individual flux ratios, no criterion fared particularly well (Figure 2B, 3C,F). None correctly predicted the observed increased flux through the Entner-Doudoroff pathway, nor the trend of reduced oxaloacetate from phosphoenolpyruvate in evolved lines.
A second data set we considered was the seven populations of E. coli that evolved on the poorly-utilized substrate lactate for ∼900 generations [28]. These populations improved in growth rate and cell dry weight substantially (112% and 50%, respectively) in addition to increasing lactate uptake by 40% [28].
We found that adaptation to growth on lactate led to a significant increase of 8% in the predicted percent optimal BM/S (Figure 4A; T-test, p = 0.02), whereas the % optimal ATP/S decreased significantly (Figure 4D; T-test, p = 0.01) by 7%. The % optimality for BM/Σv and ATP/Σv again qualitatively followed the respective per substrate criteria (Figure S4). Similarly, fluxes moved closer to the state predicted by BM/S by an average of 20% (Figure 4B; T-test, p = 0.005), largely as the result of changes in the predicted and observed flux to acetate (Figure 4C,F). In contrast, they moved away from the state predicted by ATP/S (Figure 4E; T-test, p = 0.0004). Additionally, 6 of the 7 lactate populations evolved to be further from the Pareto optimal surface than their ancestor (Figure S5).
As a third test of whether strains evolve in a manner consistent with FBA predictions, we considered the results from evolution on glucose for KO populations with lesions in central metabolism (see Figure 2A). These data come from two populations each initiated with strains lacking phosphoglucose isomerase (Δpgi), triose-phosphate isomerase (Δtpi) or phosphoenolpyruvate carboxylase (Δppc) and evolved for ∼800, ∼600, and ∼750 generations respectively [20]. Considering the improvement of these populations jointly, they increased in both growth rate and glucose uptake (172% and 157%), had large changes in central metabolic fluxes, but were largely unchanged in dry cell weight (3%). For analyzing changes in their metabolic fluxes, however, we do not present statistical tests of significance given that we only have two observations for each of these three ancestors.
Our analysis of the flux data indicated that, for BM/S, Δpgi, and Δtpi strains got worse while Δppc strains improved their % optimality (Figure 5A). This pattern largely held for ATP/S as well, though Δtpi strains showed essentially no change in % optimality (Figure 5C). The KO data set is the only one in which minimizing Σv led to qualitatively different behavior from the per substrate analyses. Minimizing flux led to increases in the % optimality for Δpgi and Δtpi when using BM/Σv as a criterion (Figure S4).
Evolution pushed strains further away from optima in all cases except Δpgi as predicted by BM/S (Figure 5B,D). Reduced distance to the optima for Δpgi was driven by reduction in the flux from oxaloacetate to phosphoenolpyruvate in evolved lines. Finally, the two Δpgi evolved isolates evolved to be more Pareto optimal, the Δtpi isolates were essentially equivalent to their ancestor, and the Δppc isolates became less Pareto optimal (Figure S5).
Genome-scale metabolism is sufficiently complex that the current state of the art in predictive models uses stoichiometry and other constraints to define the space of possible flux patterns and then suggests a given state that the cell would adopt if selection had maximized a proposed optimality criterion. The application of a mechanistic evolutionary optimality model to propose a solution to an underdetermined physiological problem is elegant and has been adopted broadly. However, there is a paucity of data testing either the central assumption that intracellular fluxes are optimized by a simple criterion, or which criterion best represents the target of selection. Here we present an analysis of metabolic evolution in the Lenski LTEE populations and make the first direct comparison of observed flux evolution to genome-scale FBA predictions.
Our analysis of the evolution of metabolic fluxes during 50,000 generations of adaptation of E. coli on glucose revealed changes in both the absolute and relative fluxes. Concordant with faster growth rates, we observed that all lines had increases in measured glucose uptake. Beyond this, all populations altered the way in which they utilize glucose, with significant changes in flux ratios observed across the network of central carbon metabolism. The most parallel changes in flux distribution were observed in the glycolytic pathways with a universal small, but significant increase in flux through the Entner-Doudoroff pathway, which was nearly always accompanied by a decrease through the pentose phosphate pathway. This is somewhat perplexing, as the Entner-Doudoroff pathway provides less ATP than glycolysis and no important biosynthetic intermediates. The Entner-Doudoroff pathway is shorter than glycolysis, and hence potentially less enzymatically costly. Indeed, what maintains the pathway in E. coli remains an open question, though it has been observed to be upregulated in E. coli during long-term starvation [35].
The major basis of improvement during selection upon growth rate for the LTEE populations –as was observed for the lactate and KO populations– came from increasing substrate uptake. We found that the LTEE populations continued to increase their growth rate over the 30,000 generations since it was last reported [33]. Alternative measures of yield, such as cell dry weight and OD600, also increased despite the slight decrease in efficiency of biomass production by central metabolism. Cell dry weight depends upon both BM/S in terms of carbon, but can also change due to the relative biomass composition of elements such as nitrogen or phosphorus. OD600 is even more indirect, depending upon all of these factors as well as changes in optical properties such as cell size, which is known to have increased in the LTEE [32]. We only measured flux ratios in central carbon metabolism, and thus would have missed significant adaptation that happened in peripheral metabolic pathways. Alternatively, either the bulk composition of biomass itself or the maintenance energy might change. We addressed these latter two factors in additional analyses (Figure S7), but neither of these factors significantly alters results.
Data on the evolution of central metabolism for the LTEE populations, combined with prior observations of flux evolution on lactate or by a series of three KO strains provided the opportunity to test several facets of whether the direction of evolutionary change was consistent with FBA predictions.
Across experimental systems we ascertained which proposed optimality criteria are most often consistent with the observed evolution in central metabolism. On average across five different ancestors, BM/S outperformed the other criteria in terms of either increasing or going unchanged (Figure S8). The most dramatic example was seen for the lactate-evolved populations, for which BM/S increased while ATP/S decreased. The per flux criteria (BM/Σv and ATP/Σv) behaved qualitatively the same as the per substrate criteria in all but two of the cases (Δpgi and Δtpi). BM/Σv outperformed BM/S in these two cases, but, for example, did not significantly improve in the lactate populations. The data also suggest that cultures quite often evolved to be further from their Pareto optimum representing the space of optimal tradeoffs [15], with 19 of 23 populations in total moving further from the Pareto surface than their respective ancestral genotypes. These results suggest that optimal biomass yield –which is the most commonly utilized criterion for FBA– was the best overall stoichiometric proxy for cultures where selection was directly upon growth rate. It will be quite interesting to analyze populations grown in a manner where yield (BM/S) is directly selected.
Overall, approximately half of the flux data were consistent with FBA predictions, and half refuted the common assumption that evolution acts to optimize efficiency; what accounts for this discrepancy? The major factor that appears to account for this difference is the initial degree of optimality for the ancestor of the evolved lines (Figure 6). For the lactate and Δppc populations, which began at approximately 80% and 90% optimality for BM/S, all 9 total replicates increased in BM/S. On the other hand, 13 of 14 populations starting at or above 95% efficiency –LTEE and the other two KOs– decreased in BM/S. A negative correlation holds whether one performs a parametric statistical test (Pearson correlation, p<0.0001) or a non-parametric Spearman correlation coefficient (p<0.0001), though it should be noted that the strength of the correlation is largely driven by the lactate data set.
The finding that selection on optimal efficiency depends on distance to the optimum is both of practical and fundamental interest. The analysis represents the first direct demonstration that FBA can be used to predict changes in intracellular metabolism that result from adaptation on a single carbon source. This positive result comes with the caveat that strains must begin far from the optimum. Systems initially operating at high yield –like the LTEE and the Δpgi strains that both began above 98% optimal– may end up evolving to be further from optimal than they began. In other words, this suggests one can either predict the initial physiological state or the direction of evolution, but not both.
What is perhaps the most remarkable about these findings is that even for cultures with a negative correlation between rate and yield, the tradeoff was quite modest. Small decreases in BM/S were more than made up for by large increases in uptake, leading to a net increase in growth rate despite mild antagonism. Given that there is no direct selection upon yield during batch culture, this perhaps suggests the existence of constraints upon the further improvement of substrate uptake. As long as uptake is held constant then changes in yield would directly translate into changes in growth rate. As such, this would maintain purifying selection upon yield, even over 50,000 generations. On the other hand, the low efficiency ancestors were able to evolve both improved substrate uptake and yield simultaneously.
Although FBA is typically applied as a practical tool to guide experiments –and it has had some remarkable successes, such as correctly predicting a rather unexpected new metabolic pathway in some cancers [36]– it also serves as a quantitative, testable, falsifiable model that connects physiology to evolution. The interplay of optimality models and laboratory adaptation will be critical as the field continues to move toward a fuller understanding of the selection and constraints that act upon biochemical networks.
Escherichia coli B isolates were obtained from the Lenski LTEE experiment [29] after 50,000 generations. Briefly, 12 populations of E. coli were founded with either the arabinose-negative strain REL606 (populations A−1 to A−6) or the arabinose-positive derivative, REL607 (A+1 to A+6). These were evolved in 10 mL of Davis-Mingioli minimal medium with 139 µM glucose (25 mg/L) as a growth substrate in 50 mL flasks since 1988. These lines have been cultured at 37°C while shaking at 120 rpm and have been transferred daily via 1∶100 dilutions (∼6.64 net doublings per day).
The isolates analyzed in the current experiment consisted of the ancestral line, REL606 [29], as well as the ‘A’ clone from 10 of the 12 lines frozen at 50,000 generations that were used in an earlier paper (A−1 = REL11330; A−2 = REL11333; A−4 = REL11336; A−5 = REL11339; A−6 = REL11389; A+1 = 11392; A+2 = REL11342; A+3 = REL11345; A+4 = REL11348, A+5 = REL11367) [30]. The A−2 clone used is from the ‘large’ lineage that has coexisted with a cross-feeding ‘small’ lineage for tens of thousands of generations [37]. The isolate from the citrate-consuming population A−3 (REL11364) was not used because it adapted to citrate consumption in addition to glucose [38]. The A+6 isolate (REL11370) was excluded from analysis because it had inconsistent growth, and gave irregular flux data. This population was previously excluded from a study of growth rate vs. yield at 20,000 generations for similar reasons [13].
Flux measurements were obtained based on the methods of Zamboni et al [17]. Evolved isolates were grown in 150 mL of Davis-Mingioli minimal media with 139 µM glucose without sodium citrate (excluded to ensure that it was not used as a secondary carbon source by any line). In order to obtain information from different parts of central metabolism, 13C-labeling either utilized a 20∶80 ratio of [U-13C]labeled∶unlabeled glucose or 100% [1-13C]glucose (Cambridge Isotope Laboratories, Andover, MA). The ancestral REL606 was grown in 200 mL to obtain sufficient cell material. At mid-log phase (60–80% max OD) all cells were pelleted from the media, hydrolyzed overnight in 6 M HCl, and dried. The dry cell material was then derivatized for an hour at 85°C with 40 µL each of dimethylformamide and N-tert-butyldimethylsilyl-N-methyltrifluoroacetamide with 1% tert-butyldimethyl-chlorosilane. The derivatized cell material was injected into a Shimadzu QP2010 GCMS (Columbia, MD). The injection source was 230°C. The oven was held at 160°C for 1 min, ramped to 310°C at 20°C min−1, and finally held at 310°C for 0.5 min. Flow rate was 1 mL min−1 and split was 10. The column was a 30 m Rxi-1ms (Restek, Bellefonte, PA). Three technical and three biological replicates were run for each isolate.
Data files from the GC-MS were analyzed in FiatFlux [17], as had been used for the lactate [28] and KO [20] populations we also analyzed. The data conversion files were rewritten to load the raw spectra produced by our MS. Following the established protocol, uninformative amino acid fragments were removed. Means for each biological replicate were calculated from the average of three technical replicates. Shapiro-Wilk tests were used to validate the assumption of normally-distributed errors for estimated flux ratios for each strain (Figures S2, S3). Variance in flux ratios was then analyzed with a MANOVA test using the Pillai's Trace metric with flux ratios entered as separate dependent variables (Figure S1). Univariate ANOVA tests were also run to investigate which of the measured flux ratios changed significantly for individual strains.
The flux of oxaloacetate (OAA) from phosphoenolpyruvate (PEP) was further estimated by a Monte-Carlo method to determine the contribution of the glyoxylate shunt. The method follows Waegeman et al, 2011 [23] and uses MATLAB code they kindly provided. In short, average mass distribution vectors and standard deviations were calculated from the measured samples. The ‘normrand’ function was then used to randomly draw from these mean distributions 1000 times. For each draw, a grid search was used to find the flux ratios that best fit the mass distribution vectors. Substantial variation was found for the fraction of labeled CO2 and flux through the glyoxylate shunt, but in all cases there was very strong support for the flux ratio of oxaloacetate from phosphoenolpyruvate that had previously been calculated by FiatFlux.
Uptake and production of cell material were determined in a separate set of experiments. In these experiments glucose concentrations were increased ten-fold to 1.39 mM so that enough of the compounds would be present to measure precisely. A volume of 250 µL of overnight culture was inoculated into 50 mL of media grown in a 250 mL flask at 225 rpm. Growth rate was determined by fitting a logarithmic model to OD600 measurements. A 10 mL sample was removed at early (OD600 of 0.090–0.120) and late (OD600 of 0.275–0.400) log phase. Cells were immediately removed from the media by passage through a 0.2 µM filter. Glucose concentrations were determined in the spent media using a glucose oxidase assay kit (Sigma, Saint Louis, MO). Acetate concentrations were determined by ion chromatography with a Dionex ICS-200 RFIC. The flow rate was 1.5 ml/min and the column temperature was 30°C. Cell dry weight (CDW), was measured as the mass of the pellet from 100 mL of fully-grown culture after overnight lyophilization. Three replicates were assayed for each measurement.
The degree of parallelism between replicates in the evolution of flux ratios was determined by calculating the coefficient of variation in flux ratios. For each flux ratio the standard deviation between evolved replicates was divided by the mean of that flux ratio. This value was then averaged across all flux ratios. Values close to zero indicate a high degree of similarity between evolved lines.
Flux analysis was carried out with a genome-scale model of E. coli metabolism (iAF_1260 [39]). The model incorporates 2382 reactions and 1668 metabolites. Substrate uptake and excretion were constrained to that observed, otherwise the default minimal media environment was used. The lower bound on maintenance energy was left at the default value of 8.9 mmol ATP/g/hr. Oxygen uptake rates were set to those observed for the lactate strains; however these data were not available for the REL or KO strains. In these cases, oxygen uptake for the ancestor was scaled across the previously observed range of 11.5–14.75 mmol/gCDW/hr [11]. Previous work demonstrated that oxygen uptake varies as a function of evolution, but that the ratio of substrate to oxygen usage remained largely constant [24]. Oxygen constraints for evolved lines were therefore set based on evolved glucose uptake rates and the ancestral ratio of oxygen/glucose. Changing the value of ancestral oxygen constraint, or the slope of constraint line had little qualitative effect (Figure S6), so just the results based on an ancestral uptake of 14.75 mmol/g/hr and a slope maintaining the original oxygen/glucose rates are reported in the text. Gene knockouts were simulated by constraining flux through the missing gene to zero.
For all data sets we systematically tested the predictive ability of four different optimality criteria: max biomass per unit substrate (BM/S), max ATP per unit substrate (ATP/S), max biomass per unit flux (BM/Σv) and max ATP per unit flux (ATP/Σv). These criteria relate to the best performers in Schuetz et al 2007 [11] and were defined as in that study. The per-substrate criteria maximized the criterion and then subsequently chose a flux distribution that minimized the difference from the observed isolate ratios. This process always provides a flux distribution with maximal production of ATP (or biomass). The per-flux criteria optimize the ratio of ATP (or biomass) to the sum of the flux. Optimizing this ratio leads to a single optimal flux solution that often produces less than the maximal ATP (or biomass). For ATP criteria, flux to excess ATP use (via maintenance energy) was maximized while constraining the lower limit of biomass production to the ancestral growth rate.
Minimizing the distance between observed and predicted optimal flux distributions was accomplished by minimizing a distance term. Flux ratios can be constrained by adding a row to the S matrix such that: Where Vn is the flux through reaction n and R is the ratio V2/V1. To minimize distance between observed and predicted ratios the equation becomes:Where D represents distance from the observed ratio and is added as two columns to the S matrix (and concomitant rows in the flux vector). Biomass or ATP can be constrained to its maximum value and then the flux distribution that is closest to observed values can be calculated by running linear optimization minimizing D as the objective function.
We first tested whether flux ratios evolve to increase each selective criterion. The optimal value of each criterion was compared against the maximum value of the criterion when the model was constrained to have the experimentally observed flux ratios. Percent optimality, calculated as the constrained criterion divided by optimal criterion, was determined for the ancestor and evolved lines.
For the LTEE lines the constrained flux ratios were serine through glycolysis, pyruvate through Entner-Doudoroff, oxaloacetate from phosphoenolpyruvate, phosphoenolpyruvate from oxaloacetate, and the pyruvate from malate. The ratios were calculated following Fischer and Sauer 2003 [40]; the exact equations used are provided in the supplementary material (Table S3). Each ratio was constrained by adding a row to the S matrix that defined the relationship between relevant fluxes (as described in the first equation of the previous section). The ratio inferred for pyruvate from malate was treated either as an absolute constraint or a lower bound but because all optimality criteria push this value towards 0 the results were equivalent.
To propagate uncertainty in glucose uptake, acetate excretion and flux ratios for the LTEE isolates, separate calculations of properties such as BM/S were made for each of 3 biological replicates, which themselves represented the average of 3 technical replicates. The mean and standard error for optimality metrics was calculated for each strain from the biological replicates.
Flux constraints for lactate and knockout data sets were implemented as upper and lower bounds, because reported flux ratios were relative to substrate uptake rather than other internal fluxes. Lactate adaptation lines were constrained to have flux ratios ±5% of the values reported in Hua et al 2007 [28]. Gene knockout lines were constrained with the flux ratios and errors reported in Fong et al, 2006 [20].
To determine whether strains evolved towards predicted optimal intracellular physiologies we used a standardized metric to ask if evolved lines were closer to an optimal solution than the ancestor. This distance metric was calculated as:where DEO was the distance of the evolved flux ratios from the closest optimal solution, and DAO was the distance of the ancestor from its closest optimal solution. Distances were calculated as Euclidean distance between the flux ratios observed in each data set and those predicted. It should be noted that because optimal flux ratios change with substrate uptake the ancestral and evolved optima were different points. The metric is 0 if the evolved isolate distance has not changed relative to the ancestor, increasingly positive as the evolved strain moves nearer an optimum, and increasingly negative as it moves further away.
A Pareto optimal surface was calculated for each line by constraining the substrate uptake rate and then doing a nested grid search [15]. A grid search across the range of feasible biomass values was executed. At each value of biomass a grid search of ATP yields was carried out and the sum of fluxes was subsequently minimized at every interval. Conservatively, for each isolate we then determined the closest possible position to its optimal surface given the observed constraints. Distance between the isolate and the Pareto optimal surface was calculated from the difference in standardized criteria.
The normality assumption for physiological measurements for the LTEE populations and optimality metrics for all data sets were checked with the Shapiro-Wilk test on the residuals of the linear model fitting the metric against strains. In all but one case the null hypothesis that the distribution was normal could not be rejected at p<0.05. The % optimality for the LTEE lines with ATP/S as the optimality criterion was not normally distributed. Q-Q plots are presented in the supplementary material (Figures S2 and S3).
For the LTEE lines ancestral versus evolved values were compared with two-sided, two sample T-tests assuming equal variance. For the non-normal ATP/S comparison a Mann-Whitney Wilcoxon Rank Sum Test was used instead. For the lactate populations only a single value was available for the ancestor so two-sided, one-sample T-tests were performed testing against the ancestral value as the mean.
|
10.1371/journal.pbio.0050118 | Odorant-Binding Proteins OBP57d and OBP57e Affect Taste Perception and Host-Plant Preference in Drosophila sechellia | Despite its morphological similarity to the other species in the Drosophila melanogaster species complex, D. sechellia has evolved distinct physiological and behavioral adaptations to its host plant Morinda citrifolia, commonly known as Tahitian Noni. The odor of the ripe fruit of M. citrifolia originates from hexanoic and octanoic acid. D. sechellia is attracted to these two fatty acids, whereas the other species in the complex are repelled. Here, using interspecies hybrids between D. melanogaster deficiency mutants and D. sechellia, we showed that the Odorant-binding protein 57e (Obp57e) gene is involved in the behavioral difference between the species. D. melanogaster knock-out flies for Obp57e and Obp57d showed altered behavioral responses to hexanoic acid and octanoic acid. Furthermore, the introduction of Obp57d and Obp57e from D. simulans and D. sechellia shifted the oviposition site preference of D. melanogaster Obp57d/eKO flies to that of the original species, confirming the contribution of these genes to D. sechellia's specialization to M. citrifolia. Our finding of the genes involved in host-plant determination may lead to further understanding of mechanisms underlying taste perception, evolution of plant–herbivore interactions, and speciation.
| Most herbivorous insects specialize on one or a few host plants; understanding the processes and genetics underlying this specialization has broad implications across biology. Drosophila sechellia, a fruit fly endemic to the Seychelles, feeds exclusively on the ripe fruit of Morinda citrifolia, a tropical plant commonly known as Tahitian Noni. Although other fruit flies never approach this fruit because of its toxins, D. sechellia is resistant and is actually attracted by the same toxins. D. sechellia is a close relative of D. melanogaster, an established model species of genetics. By comparing D. melanogaster and D. sechellia, we revealed that two genes encoding odorant-binding proteins, Obp57d and Obp57e, are not only involved in the fruit fly's taste perception, but can also change the behavioral response of the flies to the toxins contained in the fruit. By knowing how an insect's food preference is determined by its genes, we can gain insight into how insect lifestyles evolve and investigate whether such changes can lead to the formation of new species. We can also begin to understand how to manipulate insects' behavior by changing their preference for particular substances.
| Every animal must locate and identify sufficient food to meet its biological requirements. For herbivorous insects, this results in an endless battle with their host plants [1]. For example, some plants develop a chemical defense system that causes toxicity to generalist herbivores [2]. In response, generalist herbivores may then evolve a behavioral system to avoid such toxic plants. If an insect species acquires resistance to a plant toxin, however, it may reap an ecological advantage by gaining exclusive access to the toxic plant and may subsequently evolve as a specialist herbivore with a specific preference towards that plant. Such physiological and behavioral specialization plays an important role in the evolution of divergent ecological interactions between herbivores and their host plants. Nevertheless, it does not necessarily follow that ecological specialization for a particular host plant drives speciation of herbivores itself. Such specialization may not be sufficient to maintain divergence between populations at an early stage of speciation, in the face of potential gene flow via hybridization between evolving populations. The role of ecological specialization in speciation remains, therefore, to be proven [3]. Thus, it is necessary to identify the genes and molecular mechanisms responsible for ecological adaptation if we are to understand whether ecological adaptation can be a cause, or merely a consequence, of speciation [4].
Behavioral adaptation of herbivorous insects to their host plants involves the evolution of the chemosensory system [5–7]. With the recent identification of olfactory and gustatory receptors [8], knowledge of the genetic and molecular mechanisms of insect olfactory and gustatory system markedly progressed. Recent analysis of genomic information from several insect species has also revealed that the number of genes encoding these receptors varies considerably between species, indicating a close relationship between the genomic constitution of chemoreceptor gene families and the species-specific lifestyles of insects [9–11]. Thus, it is likely that the genes responsible for ecological adaptation are to be found among these receptor-encoding and receptor-related genes.
Genetic studies of Drosophila have also contributed to a substantial amount of our knowledge of “speciation genes” [4]. However, these studies have primarily focused on genes that cause reproductive isolation, and genetic analysis of ecological adaptation is relatively rare. This is, in part, due to the surprisingly limited information about Drosophila in the wild, compared with those flies reared in the laboratory as a sophisticated model system of genetics. In fact, we know little about their natural foods in the wild, except for a few species.
Drosophila sechellia is a specialist of Morinda citrifolia, which is commonly known as Tahitian Noni [12]. Although D. sechellia shows a preference for and resistance to the ripe fruit of M. citrifolia, its most closely related species, D. simulans and D. mauritiana, as well as D. melanogaster, are generalists and die upon contact with M. citrifolia, and thus avoid the fruit [13,14]. Because of genetic resources available for D. melanogaster and D. simulans, D. sechellia is an ideal organism with which to explore the genetics of ecological specialization. Analysis of quantitative trait loci (QTL) between D. sechellia and D. simulans has already identified the chromosomal regions responsible for the interspecies difference in resistance to the toxicity of M. citrifolia [15]. However, D. sechellia's preference for M. citrifolia was explained only by the transformation of olfactory sensilla resulting in an increase of the ab3 subtype that responds to the host odorant methyl hexanoate (MH) [16]. These findings successfully describe the present status of D. sechellia's specialization for M. citrifolia, but the evolutionary history, especially how an ancestral population started to use the toxic plant as its host, has been unexplained.
Here, for the first time, we have identified genes involved in D. sechellia evolution. These genes are responsible for the behavioral differences between species in their responses to hexanoic acid (HA) and octanoic acid (OA), the toxins contained in the ripe fruit of M. citrifolia, which give it its characteristic odor. Having identified the genetic factors constituting D. sechellia's adaptation to M. citrifolia, we are now able to discuss more confidently whether host-plant specialization can drive D. sechellia speciation.
We previously reported that the behavioral difference (preference/avoidance) between D. sechellia and D. simulans in response to HA, one of the main components of odor from the ripe fruit of M. citrifolia, is controlled by at least one gene on the second chromosome [17]. Further analysis of the introgression lines between D. sechellia and the D. simulans second chromosome marker strain (net b sd pm) indicated that the behavioral difference is linked to the marker pm, which is on the distal end of the right arm of the second chromosome (I. Higa and Y. Fuyama, unpublished data).
Considering the fact that the overall structure of the second chromosome is conserved between D. simulans and D. melanogaster, we mapped the locus in more detail using a series of D. melanogaster deficiency strains lacking a terminal part of the right arm of the second chromosome. Because D. sechellia's preference for HA is a recessive trait to D. melanogaster's avoidance [17], the interspecies hybrids between D. sechellia and D. melanogaster deficiency strains that lack a region containing the responsible gene(s) were expected to show the D. sechellia–like phenotype, i.e., preference for HA.
Two deficiency strains, Df(2R)exu1 and Df(2R)AA21, showed preference for HA when they were crossed with D. sechellia, defining the responsible locus within a very small chromosomal region, in combination with Df(2R)exu2, which showed avoidance to HA when crossed with D. sechellia (Figure 1A). Because the break points of these deficiency chromosomes had been deduced from cytological observations, we determined the position of these break points precisely by PCR-direct sequencing of genomic DNA from hybrids between D. melanogaster deficiency strains and D. sechellia (Figure 1B). According to the left break point of Df(2R)exu1 and the left break point of Df(2R)exu2, the locus was narrowed down within about 200 kilobases (kb) of the genomic region that contains 24 predicted genes. There is no large deleted region in the Df(2R)AA21 chromosome around this area, which is inconsistent with the result that Df(2R)AA21 also showed preference for HA when crossed with D. sechellia. While examining the marker sequences used in break-point determination of Df(2R)AA21, however, we incidentally found that this chromosome has a small, ten–base pair (bp) deletion in the first exon (open reading frame [ORF]) of the Odorant-binding protein 57e (Obp57e) gene resulting in a frame-shift mutation (Figure 1C). Insect OBP is a protein secreted into the lymph of chemosensory hairs, and it has been shown to play a crucial role in chemosensation [18]. Thus, it seemed likely that Obp57e is a gene responsible for the interspecies difference in response to HA. However, when Obp57e ORF sequences from D. melanogaster, D. simulans, and D. sechellia are compared, there is no D. sechellia–specific alteration except for L11I, which does not affect the result of signal peptide–sequence prediction (Figure 1D). Thus, D. sechellia Obp57e ORF is supposed to be functionally intact, suggesting that the interspecies difference is not in the structure of the gene product, but rather in gene expression.
Quantitative reverse-transcriptase polymerase chain reaction (RT-PCR) analysis revealed that the level of Obp57e transcripts is higher in the legs of D. sechellia than in D. melanogaster and D. simulans (Figure 2). This could be due to an elevated transcription activity in particular cells and/or a widened expression pattern. According to the lacZ reporter experiment, D. melanogaster Obp57e is expressed only in four cells associated with chemosensory hairs on the fourth and fifth segments of each tarsus, the most terminal part of an insect leg [19]. We confirmed that as short as 450 bp of the upstream region of Obp57e completely reproduces the reported expression pattern (Figure 3A–3C). We then cloned the corresponding region from D. simulans and D. sechellia, and introduced it into D. melanogaster with a green fluorescent protein (GFP) reporter gene. The D. simulans sequence successfully reproduced the same expression pattern as observed in D. melanogaster (Figure 3D). However, the D. sechellia sequence failed to drive GFP expression in any parts of the fly body (Figure 3E), indicating that the function of the D. sechellia sequence to promote gene expression is altered. Indeed, when the upstream sequence of Obp57e is compared between species, a 4-bp insertion was found in the D. sechellia Obp57e upstream sequence (Figure 3H). GFP expression was restored by removing the inserted 4-bp nucleotides from the D. sechellia sequence, showing that this 4-bp insertion abolishes the function of the D. sechellia Obp57e promoter sequence in D. melanogaster (Figure 3F and 3G). Nevertheless, the results of GFP reporter experiments are inconsistent with that of quantitative RT-PCR analysis, thus, the exact expression pattern of Obp57e in D. sechellia remains unclarified. Therefore, it is necessary to evaluate using more direct methods whether Obp57e is truly responsible for the interspecies difference in behavioral response to HA.
We generated D. melanogaster knock-out flies for Obp57e, as well as for its neighbor Obp57d, and for both Obp57d and Obp57e, by gene targeting (Figure 4). The ends-out method was employed to achieve precise gene replacement in the gene-dense Obp57d/e region (Figure 4A). To avoid side effects on transcription of surrounding genes, the marker gene (3 kb) was excised by Cre recombinase, leaving only 34 bp of the loxP sequence. Each donor construct was designed such that the ORF was removed exactly from the ATG translation initiation site, but a putative poly-A additional signal was left intact, ensuring the termination of residual transcription that may affect the expression of downstream genes via read-through events (Figure 4B).
The loss of transcripts from the targeted gene was confirmed by quantitative RT-PCR in each knock-out strain (Figure 2). We observed, however, an unexpected interaction between Obp57d and Obp57e in their expression control. The amount of Obp57e transcripts was higher in Obp57dKO flies than in the w1118 control strain. On the other hand, the amount of Obp57d transcripts decreased in the legs of Obp57eKO flies. Because each knock-out strain lacks the intron and the ORF, these regions may contain elements that influence the expression of the other gene.
Each knock-out strain responded to HA differently from the control strain in the trap assay (Figure 5). Obp57dKO and Obp57eKO avoided HA, whereas females of Obp57d/eKO preferred it, suggesting that not only Obp57e, but also Obp57d, is involved in the behavioral difference observed in the screening assay. In fruit flies, host plants are largely determined by the oviposition site preference of adults. Thus, we also examined the oviposition site preference of knock-out flies in response to HA. Indeed, Obp57eKO and Obp57d/eKO seem to prefer lower concentrations of HA than the control flies, although the difference is not statistically significant (Figure 6, Tables 1–4). The direction of behavioral alteration was, however, not the same as that found in the trap assay for Obp57d/eKO. We also examined oviposition site preference in response to OA, the main toxic component in Morinda fruit. Because of its toxicity at high concentrations, the oviposition assay was carried out at concentrations lower than those of HA. Obp57dKO and Obp57eKO preferred higher concentrations of OA. This preference was particularly obvious for Obp57dKO, which was comparable to that of D. sechellia. Contrary to the responses to HA and OA, knock-out strains preferred concentrations of acetic acid and butyric acid similar to those preferred by control flies, showing that the alteration of behavioral responses in these knock-out strains is specific to HA and OA.
Our observation of the behavior of Obp57dKO, Obp57eKO, and Obp57d/eKO revealed that these strains are qualitatively different from each other in their responses to HA and OA. This strongly suggests that Obp57d, as well as Obp57e, is involved in D. sechellia's behavioral adaptation to M. citrifolia. Nevertheless, none of these knock-out strains was identical to D. sechellia in behavior. This is consistent with the results of quantitative RT-PCR analysis in which no knock-out strain exhibited an expression profile identical to that of D. sechellia, proving that this species is not a simple null mutant of Obp57d and/or Obp57e. Rather, D. sechellia seems to be a neomorphic mutant with an altered expression control of these genes.
To examine the functions of Obp57d and Obp57e in D. simulans and D. sechellia, we cloned these genes from D. simulans and D. sechellia and introduced them into the D. melanogaster Obp57d/eKO strain. Because an interaction between the two genes was observed with respect to their expression control, a genomic fragment spanning both Obp57d and Obp57e was used for genetic transformation. The resulting transformant flies showed altered responses to HA and OA in the oviposition site–preference assay (Figure 6; Tables 3 and 4). Obp57d/eKO; simObp57d/e flies avoided HA as D. simulans does. Conversely, Obp57d/eKO; secObp57d/e flies preferred high concentrations of OA as D. sechellia does. These results clearly showed that the Obp57d/e genomic region contains genetic information responsible for, at least in part, the interspecies differences in behavioral responses to HA and OA.
However, these transgenic flies are not complete mimicries of the original species. Although D. simulans avoided OA, as well as HA, the response of Obp57d/eKO; simObp57d/e flies to OA was not significantly different from that of the D. melanogaster control strain (Figure 6; Table 4). The responses of these two transgenic strains in the trap assay were also different from that of the original species (Figure 5). Consistent with the results of the oviposition assay, D. simulans avoided HA and D. sechellia preferred it. Obp57d/eKO; simObp57d/e females, however, did not avoid HA, and both sexes of Obp57d/eKO; secObp57d/e flies did not prefer it. Indeed, the expression profiles of Obp57d and Obp57e were not exactly the same between the transgenic strains and the corresponding original species (Figure 2). Although the genomic fragments seemed to reproduce the native expression better than the GFP reporters, there still remains significant differences in expression profile, particularly between Obp57d/eKO; simObp57d/e and D. simulans. These differences suggest a contribution of additional loci to Obp57d/e expression, and thus to the interspecies differences in behavioral responses to HA and OA.
Nevertheless, the Obp57d/e genomic region from D. simulans and D. sechellia could reproduce, at least in part, the behavioral pattern of the original species in an otherwise D. melanogaster genomic background, proving that a genetic difference in this region is actually involved in interspecies differences in behavioral responses to odorants contained in M. citrifolia. It should be particularly noted that the Obp57d/e region is alone sufficient for the strong avoidance of HA by D. simulans, which is a key trait in the evolution of D. sechellia's adaptation to M. citrifolia, as discussed below.
LUSH (OBP76a), the best studied OBP in D. melanogaster, functions as an adaptor molecule in vaccenyl acetate (VA) recognition by an odorant receptor, OR67d [20]. Mutants lacking LUSH lose their neuronal response to VA; thus, they do not respond to VA behaviorally [18]. However, our Obp57d/eKO flies retained their behavioral responses to HA and OA, suggesting that OBP57d/e do not function as adaptors for HA and OA. Rather, they seem to modulate dose-dependent responses to HA and OA, which might be achieved by other proposed functions of OBP, such as the titration or degradation of ligands [21].
There are qualitative differences in the behavioral responses to HA and OA between Obp57dKO and Obp57eKO flies. These differences might reflect functional diversification between OBP57d and OBP57e. However, the elimination of either Obp57d or Obp57e affected the expression level of the other in these knock-out flies. Obp57d removal by gene targeting increased Obp57e expression level, and Obp57e removal repressed Obp57d expression. Thus, we cannot exclude the possibility that the behavioral differences between the knock-out strains are caused by an altered expression level of either gene. A more operative method such as the Gal4-UAS system could be used to separate promoters from ORFs, thus minimizing the interaction between these two genes in expression control. It would then be possible to examine the molecular functions of OBP57d and OBP57e independently.
The results from our GFP reporter experiments and quantitative RT-PCR analysis are inconsistent. This inconsistency is also a feature of previous studies. Galindo and Smith [19] showed that the reporter constructs with 3 kb of upstream sequence from Obp57d and Obp57e were expressed in four cells in each leg, which matches the results of our GFP reporter experiments. However, using RT-PCR analysis, Takahashi and Takano-Shimizu [22] detected the transcripts not only in tarsi, but also in labella and wings. Together with the results of our quantitative RT-PCR analysis, it is clear that the reporter constructs do not reflect the complete expression pattern of Obp57d/e. The difference could be, at least in part, due to the lack of coding region in the reporter constructs. In fact, the elimination of a coding region of either Obp57d or Obp57e affected the expression level of the other gene in Obp57dKO and Obp57eKO, suggesting the involvement of ORFs and/or an intron in expression control (Figure 2). Furthermore, the introduction of the Obp57d/e genomic region from D. simulans and D. sechellia reproduced the expression of Obp57d/e in the head as well as in the legs, which was not observed in GFP reporter experiments.
Although the Obp57d/e genomic region contains a considerable part of the genetic information that controls Obp57d/e expression, it is still not sufficient to explain all the differences in the expression profile between the species; genetic factors at loci other than Obp57d/e are also likely to be responsible. There are two possibilities for such factors: (1) Trans-acting factors such as a transcription factor that is necessary for Obp57d/e expression, could control expression by determining which type of cell expresses Obp57d/e, or by determining transcription level in particular Obp57d/e-expressing cells. (2) Developmental factors determining the cell fate to become Obp57d/e-expressing cells, could increase/decrease the number of Obp57d/e-expressing cells by transforming cell fate at the expense of other cell types. In fact, ab1 and ab2 sensilla on antennae are transformed into ab3 sensilla in D. sechellia [16]. Such cell-type transformation might have occurred also in Obp57d/e-expressing cells. Genes of these two categories could be identified by, for example, screening of mutants that alter the Obp57d/e > GFP expression pattern.
D. sechellia's adaptation to M. citrifolia consists of genetic changes at many loci. Although there are likely to be additional genetic differences between D. sechellia and D. simulans, the present status of D. sechellia's adaptation to M. citrifolia can be explained by alterations in three classes of genetic factors, as follows.
Factors responsible for resistance to the host-plant toxin OA: genes of this class are mapped to at least five loci scattered over all major chromosome arms [15], suggesting that the alleles at these loci were fixed independently from each other during the course of D. sechellia evolution.
Factors responsible for the olfactory preference for M. citrifolia: D. sechellia can detect Morinda fruit from as far as 150 m away [23]. Consistent with this observation, the number of olfactory sensilla specifically tuned to the host odor MH is increased in D. sechellia [16] (but also note that MH is commonly found in many other plants). In contrast, however, there are no data showing that D. simulans avoids Morinda fruit purely on the basis of olfactory cues; all behavioral assays, including our trap assay, enable flies to come in direct contact with the odor source. There is also no neural response to HA and OA from the antennae of either D. melanogaster or D. sechellia [16]. We therefore assume that the olfactory cues from Morinda fruit are fundamentally attractive to Drosophila, and not repulsive even to D. simulans. D. sechellia has an enhanced preference specifically tuned to the Morinda blend of olfactory cues, in which MH is a functionally major component. Genes responsible for this enhanced preference are thought to function in cell fate determination during neuronal development [16], but the number of genes involved is not yet known.
Factors responsible for the D. simulans' avoidance of Morinda fruit: we found this behavior to be based on gustatory cues, and confirmed that the introduction of the Obp57d/e region from D. simulans made D. melanogaster avoid HA in the same manner as D. simulans (Figure 6), proving that D. simulans' avoidance of HA-containing media as an oviposition site is determined by Obp57d/e. These two genes are physically close to each other and are thus treated as a single locus in the following discussions.
Here, we discuss the order of allele fixation at the loci responsible for D. sechellia's adaptation to M. citrifolia. In particular, we focus on the issue of which mutation was the first to be fixed, because it must have played a key role in D. sechellia's evolution.
We speculate on this with respect to the ecological validity of each phenotype in light of natural selection. We assume that the first mutation arose at a single locus, and examine the resulting phenotype in an ecological context. (1) If the first mutation occurred at a resistance QTL, the resulting phenotype would be partially resistant to M. citrifolia. However, this phenotype is ecologically “silent” because these flies avoid Morinda fruit and may not lay eggs on it (a mutation on the resistance QTL cannot be advantageous unless a fly's behavior is changed). (2) If the first mutation was for the enhanced preference for the host odorant, flies should be attracted to Morinda fruit. This phenotype would conflict with the gustatory avoidance of Morinda fruit. The consequence of conflicting olfactory and gustatory cues is unpredictable, but we hypothesize that, at least in oviposition behavior, gustatory avoidance would override olfactory preference. Olfactory preference for a plant that is not acceptable as an oviposition site is ecologically inconsistent and obviously disadvantageous. (3) If the first mutation was at the Obp57d/e locus, the resulting phenotype would be the loss of gustatory avoidance of M. citrifolia. This seems to be also disadvantageous because flies would die upon contact with Morinda fruit because of their lack of resistance to it. However, there are circumstances that might enable an evolving population to survive and reproduce. The toxicity of Morinda fruit declines as it rots and OA degenerates [13]. Thus, there will be a point at which the toxicity is potentially low enough to be counteracted by the nutritional gain from the fruit. Moreover, because M. citrifolia flowers and fruits throughout the year, newly eclosing adults are likely to mate and reproduce on the same Morinda tree. Such conditions may not be optimal with regard to the quality of nutrients, but could potentially provide a niche with fewer competitors and may result in selection for a resistance to host toxicity. Altogether, among the three traits constituting D. sechellia's adaptation to M. citrifolia, only the loss of avoidance provides an ecologically realistic scenario for specialization without any other phenotypic changes.
The above discussion, of course, does not exclude the possibility of a simultaneous fixation of the alleles responsible for D. sechellia's adaptation to M. citrifolia. Nevertheless, it is parsimonious to assume that if there was a single causative mutation at an early stage of D. sechellia's adaptation to M. citrifolia, it was the mutation at the Obp57d/e locus that led to the loss of avoidance.
D. sechellia, together with D. mauritiana, D. simulans, and D. melanogaster, serves not only as a subject of genetic analysis of reproductive isolation, but also as a good model for genetic analysis of ecological adaptation. There are more than 50 Obp genes in the D. melanogaster genome. Studies of natural variation at these loci will provide insight into the contribution of ecological interactions to the genomic constitution.
The fly strains used were w1118 for D. melanogaster, S357 for D. simulans, and SS86 for D. sechellia [17]. Adult flies were collected immediately after eclosion, and staged for 3 d at 25 °C with a cotton plug soaked with liquid medium (5% yeast extract and 5% sucrose). Staged flies were used for the trap assay, the oviposition site–preference assay, and quantitative RT-PCR analysis.
A 30-ml glass flask containing 20 ml of HA solution in 0.05% Triton-X and a control flask containing the same amount of 0.05% Triton-X were placed in a plastic cage covered with a lid made of wire mesh. Up to 100 staged flies were introduced into the cage and kept in a dark, ventilated chamber overnight at 25 °C. The response index was calculated as RI = (Nh − Nw)/(Nh + Nw), where Nh is the number of flies trapped in HA solution and Nw is that of flies in the control trap.
The PCR primers used are listed in Table 5. To amplify a fragment of about 300–600 bp from genomic DNA extracted from the interspecies hybrids between D. melanogaster deficiency strains and D. sechellia, each primer was designed within the ORF of predicted genes, with the expectation that there is enough conservation of sequences between the two species. PCR products were subjected to direct sequencing with the same primer used for amplification. The deficiency chromosome was considered to cover the position when the sequence derived from D. melanogaster or those from both D. melanogaster and D. sechellia were detected, and it was not considered to cover when only the D. sechellia sequence was detected.
Signal peptide sequence was predicted using SignalP 3.0 [24].
The genomic sequence upstream of Obp57e was PCR amplified with the primer pair 5′-(NotI) GCGGCCGC-GCGGTGGCACCCAAAATCAAT-3′ and 5′-(BamHI) AAAGGATCC-ACTTGCTATATTCCTAGGGAA-3′. PCR products were cloned into pGreenPelican [25], and then introduced into D. melanogaster by the established P element–based transformation method. To remove the inserted 4 bp from the sechellia > GFP construct, the vector was PCR amplified using the KOD-plus enzyme (Toyobo, http://www.toyobo.co.jp/e/) that does not append a T on the ends with the primers 5′-GATTATCCATTATATTGAAATTTAATTGC-3′ and 5′-ACATTTTTAATTGCACACACATTCAG-3′, and self-ligated after phosphorylation. At least five independent transformant lines for each construct were examined for GFP expression.
Disruption of Obp57d and Obp57e was carried out by the ends-out method using the vectors provided by Dr. Sekelsky [26]. A hsp70-white marker gene was excised from pBS-70w with SphI and XhoI and subcloned into the SmaI site of pBSII after blunting to obtain pBSII-70w. The Obp57d upstream region amplified with the primer pair 5′-(EcoRI) AAAGAATTC-TTAATACGAGTATATCCCAGCAAAATCGAT-3′ (P1) and 5′-(BamHI-loxP) GGATCC-ATAACTTCGTATAGCATACATTATACGAAGTTAT-CAAACTAGTTGAAGATATCATAG −3′ and the downstream region amplified with the primer pair 5′-(PstI-loxP) CTGCAG-ATAACTTCGTATAATGTATGCTATACGAAGTTAT-GGACAAGTACTACGATACTGG −3′ and 5′-(NotI) GCGGCCGC-TATGAACACTCGCCGTGGTC-3′ (P2) were subcloned into pP{EndsOut2} with hsp70-white excised from the pBSII-70w with BamHI and PstI. The Obp57e upstream region amplified with the primer pair 5′-(BamHI-loxP) GGATCC-ATAACTTCGTATAGCATACATTATACGAAGTTAT-ACTTGCTATATTCCTAGGGAA −3′ and P1 and the downstream region amplified with the primer pair primers 5′-(PstI-loxP) CTGCAG-ATAACTTCGTATAATGTATGCTATACGAAGTTAT-GCGGCCGAGAAGTATGTTTC-3′ and P2 were subcloned into pP{EndsOut2}, similarly to the case of Obp57d. The Obp57d upstream region and the Obp57e downstream region were used for the Obp57d/e targeting vector. The fly transformation and targeting crosses were carried out as described by Sekelsky (http://rd.plos.org/pbio.0050118). Two, one, and three knock-out lines were obtained for Obp57d, Obp57e, and Obp57d/e, respectively. Each knock-out line was backcrossed to the w1118 control strain for five generations.
Genomic fragments including Obp57d/e were PCR cloned from D. simulans and D. sechellia with the primers P1 and P2, and cloned into the pCaSpeR3 transformation vector. The w1118; Obp57d/eKO strain was transformed with these vectors by the established method. At least three independent transformant lines were obtained for each construct.
RNA was extracted from the legs or heads of 20 staged females using an RNeasy Micro kit (Qiagen, http://www1.qiagen.com). cDNA was made using a SuperScript III First-strand Synthesis System (Invitrogen, http://www.invitrogen.com) with the oligo(dT)20 primer. Quantitative RT-PCR was carried out with the Chromo 4 realtime PCR analysis system (BioRad, http://www.bio-rad.com) using SYBR Premix ExTaq (Takara, http://www.takara-bio.com) with primers 5′-TTATTTTGGAAATTCAATTTAGAACTGCCG-3′ and 5′-TGATTCGGCTATATCTTCGTCTATTCCTTG-3′ for D. melanogaster Obp57d, 5′-TGCGCAAATGTTCTCGCTAACACTT-3′ and 5′-ATTCTCCATCACTTGGTGGGCTTCATA-3′ for D. melanogaster Obp57e, 5′-TTATTTTGGAAATTCAGTTTAGAATTTCCG-3′ and 5′-AATTGCTTCAGCTATATCTTCGTCTATTCC-3′ (P3) for D. simulans Obp57d, 5′-TGCGCAAACGTTCTTGCTTACACTT-3′ and 5′-GGCCATTTCTCCATCACTTGGTTG-3′ (P4) for D. simulans Obp57e, 5′-TTGGAAATTCAGTTTAGAAATTCTGAATGT-3′ and P3 for D. sechellia Obp57d, 5′-TGTGCGCAAATGTTCTTGCTTACACTT-3′ and P4 for D. sechellia Obp57e, and 5′-GCTAAGCTGTCGCACAAATG-3′ and 5′-TGTGCACCAGGAACTTCTTG-3′ for rp49 of all species. Either of a primer pair was designed at an exon boundary to ensure amplification only from spliced transcripts.
Staged females were individually supplied with media (1% yeast extract [Gibco, http://www.invitrogen.com/content.cfm?pageid=11040]) and 0.8% Bacto Agar [Gibco]) containing an acid at four concentrations (0 mM, 10 mM, 20 mM and 30 mM for acetic acid, butyric acid, and HA; and 0 mM, 2.5 mM, 5 mM, and 7.5 mM for OA) simultaneously, and allowed the choice of medium for oviposition in a dark, ventilated box overnight at 25 °C. The number of eggs laid on each medium was scored, and the weighted mean of acid concentration was calculated for each individual. At least 36 individuals were tested for each strain with three replications.
Obp57d/e sequence data have been deposited under the GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers AB232138–AB232143.
|
10.1371/journal.pcbi.1007017 | The nonlinear dynamics and fluctuations of mRNA levels in cell cycle coupled transcription | Gene transcription is a noisy process, and cell division cycle is an important source of gene transcription noise. In this work, we develop a mathematical approach by coupling transcription kinetics with cell division cycles to delineate how they are combined to regulate transcription output and noise. In view of gene dosage, a cell cycle is divided into an early stage S 1 and a late stage S 2. The analytical forms for the mean and the noise of mRNA numbers are given in each stage. The analysis based on these formulas predicts precisely the fold change r* of mRNA numbers from S 1 to S 2 measured in a mouse embryonic stem cell line. When transcription follows similar kinetics in both stages, r* buffers against DNA dosage variation and r* ∈ (1, 2). Numerical simulations suggest that increasing cell cycle durations up-regulates transcription with less noise, whereas rapid stage transitions induce highly noisy transcription. A minimization of the transcription noise is observed when transcription homeostasis is attained by varying a single kinetic rate. When the transcription level scales with cellular volume, either by reducing the transcription burst frequency or by increasing the burst size in S 2, the noise shows only a minor variation over a wide range of cell cycle stage durations. The reduction level in the burst frequency is nearly a constant, whereas the increase in the burst size is conceivably sensitive, when responding to a large random variation of the cell cycle durations and the gene duplication time.
| Gene transcription in single cells is inherently a stochastic process, resulting in a large variability in the number of transcripts and constituting the phenotypic heterogeneity in cell population. Cell division cycle has global effects on transcriptional outputs, and is thought to be an additional source of transcription noise. In this work, we develop a hybrid model to delineate the combined contribution of transcription activities and cell divisions in the variability of transcript counts. By working with the analytical forms of the mean and the noise of mRNA numbers, we show that if the transcription kinetic rates do not change considerably, then the average mRNA level is increased about 1 to 2 folds from earlier to later cell cycle stages. When transcription homeostasis is attained by varying a single kinetic rate between the two cell cycle stages, we find no significant changes in the transcription noise, and the homeostasis nearly minimizes the noise. In our continuous study on the transcript concentration homeostasis that the transcription level scales with the cellular volume, we find only minor variations of the noise if the homeostasis is maintained either by reducing the transcription burst frequency or by increasing the burst size in late cell cycle phase, in the face of a large cell cycle stage duration variation. The reduction in the burst frequency is relative robust, while the increase in the burst size is conceivably sensitive, to the large random variation of the cell cycle durations and the gene duplication time.
| Single cell studies over last decades have shown that gene transcription is inherently a stochastic process in a bursting fashion [1–5]. The transcriptional bursting, whereby a gene promoter transits randomly between short periods of mRNA production and long periods of no productions, has been widely studied and invoked to explain how the fluctuation of mRNA molecules arises among single cells of identical genes [6–8]. Early studies on the origin of variability in gene expression found that the noise is not solely due to the randomness in reactions intrinsic to gene expression [9].
Recent experiments have suggested that cell division cycle is an important source of gene expression noise [10–13]. In virtually all cells, from bacteria to mammalian cells, a conserved class of genes is involved in cell cycle stage-specific gene expression. For instance, SWI5 and CLB2 are responsible for mitotic progression, whose transcripts are stable during the interphase, but exhibit a 30-fold increase in degradation in the mitosis phase [12]. In budding yeast, acetylation of histone 3 suppresses transcription activity to buffer changes in DNA dose for expression homeostasis of other genes during DNA replication [13]. During cell division processes, genome duplication involves DNA dosage increase at discrete times in S phase, and introduces considerable variations in gene copies [13–15]. Moreover, the time spent between two successive cell-division events [11], the DNA replication catalyzed by DNA polymerases [16, 17], the variation in transcription kinetics between different cell cycle stages [9, 15, 18], and the partition of molecules between two daughter cells [19], are all observed to be stochastic and may contribute to cell-to-cell variability in transcript counts.
It remains largely unexplored how these random events govern mRNA outputs and their fluctuation among individual cells [1]. In this work, we initiate a mathematical approach by coupling the classical two-state model with cell division cycles to delineate the combined contribution of transcription activities and cell divisions in the variability of transcript counts [4, 6, 20]. In view of gene dosage, a cell cycle is divided into S 1 and S 2 stages. In each stage, the target gene transits randomly between active and inactive states with constant rates. As usual, we use the mean, the noise, and the noise strength to characterize stochastic gene transcription. For a given random variable N, we denote by E[N], E[N2], and Var[N] = E[N − E[N]]2 its mean, the second moment, and variance, respectively. Its noise and the noise strength are defined by
η 2 ( N ) = Var [ N ] ( E [ N ] ) 2 = E [ N 2 ] - ( E [ N ] ) 2 ( E [ N ] ) 2 , and Φ ( N ) = Var [ N ] E [ N ] . (1)
We will formulate the master equations for the model and derive the differential equations of the mean and the second moment. The analytical forms of the mean, the noise, and the noise strength at steady-state will be given.
We measure the fold change of mRNA copy numbers from S 1 to S 2 by r * = m 2 * / m 1 *, where m 1 * and m 2 * are the mean transcription levels at the two stages. Although r* may take any prescribed value in theory, we find that when the transcription kinetic rates are similar in the two stages, the fold change buffers against the DNA dosage variation and stays within (1, 2), as observed in yeast [12] and mammalian cells [15]. Furthermore, if stage transitions are considerably slower than transcription state transitions and mRNA turnover, then r* ≈ 2. The accuracy of our theoretical results is tested by numerical examples that generate nearly the same fold change measured in a mouse embryonic stem cell line [15]. Increasing either of the cell cycle durations up-regulates transcription with less noise, and rapid transitions between cell cycle stages are a major source of highly noisy transcription. Our numerical examples also demonstrate that transcription homeostasis does not bring significant changes in transcription noise. If transcription homeostasis is attained by varying a single kinetic rate in the two cell cycle stages, then the homeostasis nearly minimizes transcription noise. Motivated by increasing evidences that many cellular processes depend mainly on the concentration rather than the absolute number of enzymes [18, 21, 22], we continue to study the noise profile when the transcript concentration homeostasis is maintained. Our analysis reveals an interesting phenomenon that the transcription noise is relatively stable when the concentration homeostasis is maintained, either by reducing the transcription burst frequency or by increasing the burst size in late cell cycle phase, over a wide range of cell cycle stage durations. The reduction degree in the burst frequency is nearly a constant, while the increase in the burst size is conceivably sensitive, when responding to a large random variation of the cell cycle durations and the gene duplication time.
In past two decades, the two-state model has been a prevailing tool to characterize stochastic gene transcription in single cells, from bacteria, yeast, to mammalian cells [4–6, 8, 20, 23]. In the model, as depicted in the diagram
geneOFF ⇄ γ λ geneON → ν mRNA → δ ∅ , (2)
it is postulated that the gene promoter transits randomly between inactive (gene OFF) and active (gene ON) states with constant activation rate λ > 0 and inactivation rate γ > 0. The transcripts are produced only when the gene is active with a synthesis rate ν > 0, and are turned over with a degradation rate δ > 0. Apparently, as the four rates are all assumed to be constants, the transcription described by the model is independent of many important cellular processes such as cell growth and cell division.
Actively dividing eukaryote cells go through several stages known collectively as the cell division cycle, including Gap 1 phase (G1) for cell growth, the synthesis phase (S) for DNA replication, Gap 2 phase (G2) for DNA repairing, and the mitotic phase (M) for cell division; see Fig 1. During S phase, each gene is duplicated into two copies that are transcribed independently in the same cell [15]. During M phase, a cell is divided into two daughter cells and residual mRNA molecules are randomly partitioned. Cell division cycle has global effects on mRNA and protein synthesis, and is also an important source of gene expression noise [10–13]. In recent years, many real-time monitoring methods, such as single molecule fluorescent in situ hybridization (smFISH), have been developed to estimate mRNA copy numbers in different cell cycle stages. In mouse embryonic stem cells, nascent Oct4 and Nanog mRNAs were measured in different phases using smFISH method [15]. It was found that the ratio of the average number of mRNA copies in G2 phase to the average in G1 phase is 1.28 ± 0.09 for Oct4 mRNA, and 1.51 ± 0.15 for Nanog mRNA. In yeast cells, CLB2 mRNAs accumulate apace in late S phase and are degraded almost completely before cytokinesis [12]. From the measurements of [12], we estimated that the median of cytoplasmic CLB2 mRNA copy numbers is ∼10 in G2/M phase, and ∼5 in S phase. It remains an essential and widely open question to quantify how the transition of cell cycle phases, the variation of DNA content and transcription kinetics in different phases, and the random partition of mRNAs in daughter cells affect the dynamics and noise of gene transcription.
In this work, we initiate a quantitative approach to this important question by developing a model that couples gene transcription with cell cycles. During DNA replication in S phase, the two complementary strands in each double helix are separated and serve as templates for the production of their counterparts. After the completion of the whole DNA replication process, which takes hours in some cells [24], each gene copy is doubled with two copies. Normally, the duplication of a single gene takes much shorter time and is completed within seconds to minutes [17, 25]. For instance, the genome of Escherichia coli K12 has ∼ 4.64 million base pairs with ∼ 4375 genes [26], and is replicated at ∼470 ± 180 bp/s [17]. The average duplication time of each gene takes 1.63 ∼ 3.66 seconds. In our model, we treat the short duplication process of our target gene as instantaneous, and accordingly, divide a cell cycle into two stages:
We make the following assumptions to complete the description of the model:
We do not assume constant durations in S 1 and S 2 stages in (i), because the time spent in each cell cycle phase is often not fixed, and the timing for the duplication of the target gene is random. The cell cycle duration in mouse embryonic stem cells measured by flow cytometry varies in 11 ∼ 16 hours [27, 28], that are roughly distributed in G1 (26%), S (52%), and G2/M (22%) estimated by the percentage of cells in these phases [28]. The times spent in cell cycle phases were also measured by time-lapse microscopy and single cell tracking in T and B lymphocytes from reporter mice, and the total division time data were found to be well approximated by the sum of consecutive independent exponential and Gaussian distributions [11]. We assume that the transcription is turned off at the beginning of each stage, as DNA synthesis is catalyzed by DNA polymerase in nucleosomes, and during late S 2 stage, the chromatin shrinks into chromosome [29–31]. In either case, transcription factors and RNA polymerase II are usually prevented from reaching to gene promoters to initiate transcription [32]. Assumption (v) is equivalent to the binomial distribution of mRNA molecules in the two daughter cells, which has been assumed in most theoretical models, and supported by recent experiments. The partition in Escherichia coli measured by the MS2-GFP reporter strongly supports the assumption that each mRNA copy goes to one of the two daughter cells with equal probability [33].
The transcription state of a gene of our interest in a single cell at a time t ≥ 0 can be characterized by the number of active gene copies, the cell cycle stage, and its mRNA copy number. Without loss of generality, we assume that the gene has exactly one copy in S 1 stage, and two copies in S 2 stage, in any single cell of an isogenic cell population. We let I(t) denote the number of active genes in a cell. In S 1 stage, I(t) = 0 if the gene is OFF, and I(t) = 1 if it is ON. In S 2 stage, I(t) = 0 if the two gene copies are OFF, I(t) = 2 if both are ON, and I(t) = 1 in the remaining cases. We let U(t) specify the cell cycle stage, with U(t) = 1 in S 1 stage, and U(t) = 2 in S 2 stage. Let M(t) denote the mRNA copy number for the gene in one cell. Then the transcription state can be fully quantified by the following joint probabilities
P 1 ( i , m , t ) = Prob { I ( t ) = i , M ( t ) = m , U ( t ) = 1 } , i = 0 , 1 ; m = 0 , 1 , 2 , ⋯ , (3)
P 2 ( i , m , t ) = Prob { I ( t ) = i , M ( t ) = m , U ( t ) = 2 } , i = 0 , 1 , 2 ; m = 0 , 1 , 2 , ⋯ . (4)
For clarity and simplicity in the following calculations, we assume that all cells in the isogenic population are synchronized at the beginning of S 1 stage, and count only newly produced mRNA molecules from time zero. Accordingly, we have the initial condition
P 1 ( 0 , 0 , 0 ) = 1 , P 1 ( 0 , m , 0 ) = 0 , m > 0 , P 1 ( 1 , m , 0 ) = P 2 ( 0 , m , 0 ) = P 2 ( 1 , m , 0 ) = P 2 ( 2 , m , 0 ) = 0 , m ≥ 0 . (5)
By using the standard procedure in stochastic process, we calculate the time evolutions of these probabilities based on the basic assumptions (i)-(v) in our model and derive the master equations:
P 1 ′ ( 0 , m , t ) = γ 1 P 1 ( 1 , m , t ) - ( m δ 1 + λ 1 + κ 1 ) P 1 ( 0 , m , t ) + ( m + 1 ) δ 1 P 1 ( 0 , m + 1 , t ) + κ 2 ∑ n = m ∞ ( 1 2 ) n ( n m ) P 2 ( n , t ) , (6)
P 1 ′ ( 1 , m , t ) = λ 1 P 1 ( 0 , m , t ) - ( ν 1 + m δ 1 + γ 1 + κ 1 ) P 1 ( 1 , m , t ) + ν 1 P 1 ( 1 , m - 1 , t ) + ( m + 1 ) δ 1 P 1 ( 1 , m + 1 , t ) , (7)
P 2 ′ ( 0 , m , t ) = κ 1 P 1 ( m , t ) - ( m δ 2 + 2 λ 2 + κ 2 ) P 2 ( 0 , m , t ) + ( m + 1 ) δ 2 P 2 ( 0 , m + 1 , t ) + γ 2 P 2 ( 1 , m , t ) , (8)
P 2 ′ ( 1 , m , t ) = 2 λ 2 P 2 ( 0 , m , t ) + 2 γ 2 P 2 ( 2 , m , t ) + ( m + 1 ) δ 2 P 2 ( 1 , m + 1 , t ) + ν 2 P 2 ( 1 , m - 1 , t ) - ( ν 2 + m δ 2 + λ 2 + γ 2 + κ 2 ) P 2 ( 1 , m , t ) , (9)
P 2 ′ ( 2 , m , t ) = λ 2 P 2 ( 1 , m , t ) - ( 2 ν 2 + m δ 2 + 2 γ 2 + κ 2 ) P 2 ( 2 , m , t ) + 2 ν 2 P 2 ( 2 , m - 1 , t ) + ( m + 1 ) δ 2 P 2 ( 2 , m + 1 , t ) . (10)
The last expression P2(n, t) in (6), defined by
P 2 ( n , t ) = P 2 ( 0 , n , t ) + P 2 ( 1 , n , t ) + P 2 ( 2 , n , t ) ,
gives the probability that the cell resides on S 2 stage with n transcripts, and P1(m, t) in (8), defined by P1(m, t) = P1(0, m, t) + P1(1, m, t), represents the probability that the cell resides on S 1 stage with m copies of mRNA molecules. The technical steps leading to (6)–(10) are given in S1 Text.
The transcription dynamics of a gene in a cell population is best characterized by the mean value m(t) = E[M(t)] of the random process M(t) that counts the number of its mRNA copies. The second moment μ(t) = E[M2(t)] is essential in the calculation of its noise that quantifies the fluctuation of mRNA copy numbers among individual cells. More importantly, as the cell division cycle is integrated into our model, we can extend m(t) and μ(t) to the two cell cycle stages S 1 and S 2. The comparison of these quantities in the two stages can help us understand how the gene duplication contributes to the variation of transcription levels and noises. To start with, we give the formal definitions of these concepts and present the differential equations that provide a framework from which they can be solved analytically.
For this purpose, we need various probabilities by adding the joint probabilities Pj(i, m, t) introduced in (3)–(4) when i, j, or m runs through all possible values. We use a conventional simplification of notations: If any of i, j and m is removed from Pj(i, m, t), then the new probability is defined by summing Pj(i, m, t) over the range of the removed index. For instance,
P 1 ( m , t ) = P 1 ( 0 , m , t ) + P 1 ( 1 , m , t ) , P 2 ( m , t ) = P 2 ( 0 , m , t ) + P 2 ( 1 , m , t ) + P 2 ( 2 , m , t ) (11)
are the respective probabilities that the cell resides on S 1 and S 2 stages with m copies of mRNA molecules, without specifying the promoter state. A further summation of the two probabilities in (11) defines
P ( m , t ) = P 1 ( m , t ) + P 2 ( m , t ) (12)
as the probability that there are m copies of mRNA molecules in the cell. Similarly, we can define P1(i, t) and P2(i, t). To avoid the confusion with these probabilities defined in (11), we change them to P1i(t) and P2i(t) with
P 1 i ( t ) = ∑ m = 0 ∞ P 1 ( i , m , t ) , P 2 i ( t ) = ∑ m = 0 ∞ P 2 ( i , m , t ) . (13)
By adding the probabilities in (13) we have
P 1 ( t ) = P 10 ( t ) + P 11 ( t ) , P 2 ( t ) = P 20 ( t ) + P 21 ( t ) + P 22 ( t ) (14)
as the respective probabilities that the cell resides on S 1 and S 2 stages. By adding the master Eqs (6)–(10) in m, we obtain a closed system of P1i(t) and P2i(t),
{ P 10 ′ ( t ) = κ 2 P 2 ( t ) + γ 1 P 11 ( t ) - ( λ 1 + κ 1 ) P 10 ( t ) , P 11 ′ ( t ) = λ 1 P 10 ( t ) - ( γ 1 + κ 1 ) P 11 ( t ) , P 20 ′ ( t ) = κ 1 P 1 ( t ) + γ 2 P 21 ( t ) - ( 2 λ 2 + κ 2 ) P 20 ( t ) , P 21 ′ ( t ) = 2 λ 2 P 20 ( t ) - ( λ 2 + γ 2 + κ 2 ) P 21 ( t ) + 2 γ 2 P 22 ( t ) , P 22 ′ ( t ) = λ 2 P 21 ( t ) - ( 2 γ 2 + κ 2 ) P 22 ( t ) . (15)
The initial condition for this system can be derived by a summation of the initial data given in (5). This linear system of ordinary differential equations with constant coefficients can be solved analytically, and its solution subject to the corresponding initial condition determines uniquely P1i(t) and P2i(t).
Due to the technical complexity, we break down the process of determining m(t), μ(t), and their extensions in S 1 and S 2 in several steps, and move most involving calculations to S1 Text.
Step 1: The determination of the mean level m(t): With P1(m, t), P2(m, t), and P(m, t) defined in (11) and (12), we have
m ( t ) = E [ M ( t ) ] = ∑ m = 0 ∞ m P ( m , t ) = n 1 ( t ) + n 2 ( t ) , (16)
where
n 1 ( t ) = ∑ k = 0 ∞ k P 1 ( k , t ) , and n 2 ( t ) = ∑ k = 0 ∞ k P 2 ( k , t ) . (17)
As we show in S1 Text, n1(t) and n2(t) satisfy the following system of inhomogeneous linear ordinary differential equations with constant coefficients:
{ n 1 ′ ( t ) = - ( δ 1 + κ 1 ) n 1 ( t ) + κ 2 2 n 2 ( t ) + ν 1 P 11 ( t ) , n 2 ′ ( t ) = κ 1 n 1 ( t ) - ( δ 2 + κ 2 ) n 2 ( t ) + ν 2 [ P 21 ( t ) + 2 P 22 ( t ) ] . (18)
As P11(t), P21(t), and P22(t) can be solved uniquely from (15), we can find n1(t) and n2(t) by solving (18) subject to the initial condition n1(0) = n2(0) = 0, and find m(t) by (16).
Step 2: The determination of the second moment μ(t): Similar to the definition of m(t) in (16), we have
μ ( t ) = E [ M 2 ( t ) ] = ∑ m = 0 ∞ m 2 P ( m , t ) = ω 1 ( t ) + ω 2 ( t ) , (19)
where
ω 1 ( t ) = ∑ k = 0 ∞ k 2 P 1 ( k , t ) , and ω 2 ( t ) = ∑ k = 0 ∞ k 2 P 2 ( k , t ) . (20)
As we show in S1 Text, the time evolutions of ω1(t) and ω2(t) are given by the system
{ ω 1 ′ ( t ) = - ( 2 δ 1 + κ 1 ) ω 1 ( t ) + κ 2 4 ω 2 ( t ) + δ 1 n 1 ( t ) + κ 2 4 n 2 ( t ) + ν 1 [ 2 n 11 ( t ) + P 11 ( t ) ] , ω 2 ′ ( t ) = κ 1 ω 1 ( t ) - ( 2 δ 2 + κ 2 ) ω 2 ( t ) + δ 2 n 2 ( t ) + ν 2 [ P 21 ( t ) + 2 P 22 ( t ) + 2 n 21 ( t ) + 4 n 22 ( t ) ] , (21)
where
n 1 i ( t ) = ∑ m = 0 ∞ m P 1 ( i , m , t ) , i = 0 , 1 , n 2 i ( t ) = ∑ m = 0 ∞ m P 2 ( i , m , t ) , i = 0 , 1 , 2 , (22)
and
n 1 ( t ) = n 10 ( t ) + n 11 ( t ) , n 2 ( t ) = n 20 ( t ) + n 21 ( t ) + n 22 ( t ) .
Apparently, (21) is not a closed system, and finding ω1(t) and ω2(t) requires the following system of n1i(t) and n2i(t):
{ n 10 ′ ( t ) = κ 2 2 n 2 ( t ) - ( δ 1 + λ 1 + κ 1 ) n 10 ( t ) + γ 1 n 11 ( t ) , n 11 ′ ( t ) = λ 1 n 10 ( t ) + ν 1 P 11 ( t ) - ( δ 1 + γ 1 + κ 1 ) n 11 ( t ) , n 20 ′ ( t ) = κ 1 n 1 ( t ) + γ 2 n 21 ( t ) - ( δ 2 + 2 λ 2 + κ 2 ) n 20 ( t ) , n 21 ′ ( t ) = 2 λ 2 n 20 ( t ) + 2 γ 2 n 22 ( t ) + ν 2 P 21 ( t ) - ( δ 2 + λ 2 + γ 2 + κ 2 ) n 21 ( t ) , n 22 ′ ( t ) = λ 2 n 21 ( t ) + 2 ν 2 P 22 ( t ) - ( δ 2 + 2 γ 2 + κ 2 ) n 22 ( t ) , (23)
This system is obtained by multiplying (6)–(10) with m and then taking sums. As P11(t), P21(t), and P22(t) can be solved from (15), it is a closed system of n1i(t) and n2i(t). By substituting its unique solution subject to the zero initial condition into (21), we can determine ω1(t) and ω2(t), and therefore the second moment μ(t).
Step 3: The moment functions on
S
1
and
S
2
stages: To extend the definitions of m(t) and μ(t) to the two cell cycle stages S 1 and S 2, we define the conditional probabilities
p 1 ( i , m , t ) = Prob { I ( t ) = i , M ( t ) = m | U ( t ) = 1 } = P 1 ( i , m , t ) P 1 ( t ) , (24)
p 2 ( i , m , t ) = Prob { I ( t ) = i , M ( t ) = m | U ( t ) = 2 } = P 2 ( i , m , t ) P 2 ( t ) , (25)
for the probabilities P1(t) and P2(t) defined in (14). Then
p 1 ( m , t ) = p 1 ( 0 , m , t ) + p 1 ( 1 , m , t ) , p 2 ( m , t ) = p 2 ( 0 , m , t ) + p 2 ( 1 , m , t ) + p 2 ( 2 , m , t )
are the probabilities that there are m copies of mRNA molecules when the cell resides on S 1 or S 2 stage. The average transcription levels in S 1 and S 2 stages are defined by
m 1 ( t ) = ∑ k = 0 ∞ k p 1 ( k , t ) , m 2 ( t ) = ∑ k = 0 ∞ k p 2 ( k , t ) , (26)
and the second moments are defined by
μ 1 ( t ) = ∑ k = 0 ∞ k 2 p 1 ( k , t ) , μ 2 ( t ) = ∑ k = 0 ∞ k 2 p 2 ( k , t ) . (27)
By comparing (26) with the definition of n1(t) and n2(t) in (17), and (27) with the definition of ω1(t) and ω2(t) in (20), we find the simple relation
m 1 ( t ) = n 1 ( t ) P 1 ( t ) , m 2 ( t ) = n 2 ( t ) P 2 ( t ) , μ 1 ( t ) = ω 1 ( t ) P 1 ( t ) , μ 2 ( t ) = ω 2 ( t ) P 2 ( t ) . (28)
As a cell is either on S 1 or on S 2 stage, we have P1(t) + P2(t) ≡ 1. From the basic assumption (i), the two stages S 1 and S 2 transit each other by constant rates κ1 and κ2. It implies that P1(t) and P2(t) are simply related by
P 1 ′ ( t ) = κ 2 P 2 ( t ) - κ 1 P 1 ( t ) = κ 2 - ( κ 1 + κ 2 ) P 1 ( t ) .
This simple equation can also be derived by adding equations in (15). By the assumption that all cells are synchronized on S 1 initially, we have P1(0) = 1. Hence
P 1 ( t ) = κ 2 κ 1 + κ 2 + κ 1 κ 1 + κ 2 e - ( κ 1 + κ 2 ) t , P 2 ( t ) = κ 1 κ 1 + κ 2 - κ 1 κ 1 + κ 2 e - ( κ 1 + κ 2 ) t . (29)
Our methods for finding n1(t) and n2(t) in Step 1, and ω1(t) and ω2(t) in Step 2, combined with (28) and (29), constitute a complete analytical approach for computing the mean values m1(t) and m2(t), and the second moments μ1(t) and μ2(t), in the two cell cycle stages.
Our discussion in the previous section offers a clear analytical approach for finding the mean value m(t) and the second moment μ(t) of mRNA number M(t), along with their extensions to the two cell cycle stages S 1 and S 2. However, neither of these functions has a simple analytical expression. For simplicity, we will only present their steady-state values in exact forms, and use their temporal forms in numerical simulations. Although the steady-state values are much simpler than the temporal forms, they are still rather complex and capture the delicate involvement of the system parameters as shown by the next two theorems. For a function f(t) that has a finite limit as t → ∞, we let f* denote its limit.
Theorem 1 If the transcription of a gene obeys the model described in Fig 1, then the mean transcription level of the gene in a population of isogenic cells at steady-state is m * = m 1 * · κ 2 κ 1 + κ 2 + m 2 * · κ 1 κ 1 + κ 2 , (30)
a linear combination of the mean levels m 1 * in S 1 stage and m 2 * in S 2 stage, and
m 1 * = 2 ν 1 λ 1 ( δ 2 + κ 2 ) ( λ 2 + γ 2 + κ 2 ) + 2 ν 2 λ 2 κ 1 ( λ 1 + γ 1 + κ 1 ) [ 2 ( δ 1 + κ 1 ) ( δ 2 + κ 2 ) - κ 1 κ 2 ] ( λ 1 + γ 1 + κ 1 ) ( λ 2 + γ 2 + κ 2 ) , (31)
m 2 * = 2 ν 1 λ 1 κ 2 ( λ 2 + γ 2 + κ 2 ) + 4 ν 2 λ 2 ( δ 1 + κ 1 ) ( λ 1 + γ 1 + κ 1 ) [ 2 ( δ 1 + κ 1 ) ( δ 2 + κ 2 ) - κ 1 κ 2 ] ( λ 1 + γ 1 + κ 1 ) ( λ 2 + γ 2 + κ 2 ) . (32)
Theorem 2 If the transcription of a gene obeys the model described in Fig 1, then the second moment of its mRNA copy number M(t) at steady-state is μ * = μ 1 * · κ 2 κ 1 + κ 2 + μ 2 * · κ 1 κ 1 + κ 2 , (33) where μ 1 * and μ 2 * are the second moments in S 1 and S 2 stages given by μ 1 * = m 1 * + 8 ν 1 ( κ 2 + 2 δ 2 ) · m s 1 * + 2 ν 2 κ 1 · m s 2 * 4 ( κ 1 + 2 δ 1 ) ( κ 2 + 2 δ 2 ) - κ 1 κ 2 , (34)
μ 2 * = m 2 * + 8 ν 1 κ 2 · m s 1 * + 8 ν 2 ( κ 1 + 2 δ 1 ) · m s 2 * 4 ( κ 1 + 2 δ 1 ) ( κ 2 + 2 δ 2 ) - κ 1 κ 2 , (35)
with
m s 1 * = ( δ 1 + λ 1 + κ 1 ) m 1 * - κ 1 m 2 * / 2 δ 1 + λ 1 + γ 1 + κ 1 , (36)
m s 2 * = ( δ 2 + κ 2 + 2 λ 2 ) m 2 * - κ 2 m 1 * + 2 ν 2 p 22 * δ 2 + λ 2 + γ 2 + κ 2 , (37) and p 22 * = 2 λ 2 2 / [ ( κ 2 + λ 2 + γ 2 ) ( κ 2 + 2 λ 2 + 2 γ 2 ) ].
The proofs of Theorems 1 and 2 are given in S1 Text. By using definition (1), combined with the analytical expressions (31) and (32) of the stationary mean transcription levels, and (34) and (35) for the second moments, we derive the noise strengths of mRNA copy numbers in S 1 and S 2 as
Φ 1 * = 1 - m 1 * + 1 m 1 * · 8 ν 1 ( κ 2 + 2 δ 2 ) · m s 1 * + 2 ν 2 κ 1 · m s 2 * 4 ( κ 1 + 2 δ 1 ) ( κ 2 + 2 δ 2 ) - κ 1 κ 2 , (38)
Φ 2 * = 1 - m 2 * + 1 m 2 * · 8 ν 1 κ 2 · m s 1 * + 8 ν 2 ( κ 1 + 2 δ 1 ) · m s 2 * 4 ( κ 1 + 2 δ 1 ) ( κ 2 + 2 δ 2 ) - κ 1 κ 2 . (39)
The noises η 1 2 * and η 2 2 * are given by
η 1 2 * = 1 m 1 * - 1 + 1 ( m 1 * ) 2 · 8 ν 1 ( κ 2 + 2 δ 2 ) · m s 1 * + 2 ν 2 κ 1 · m s 2 * 4 ( κ 1 + 2 δ 1 ) ( κ 2 + 2 δ 2 ) - κ 1 κ 2 , (40)
η 2 2 * = 1 m 2 * - 1 + 1 ( m 2 * ) 2 · 8 ν 1 κ 2 · m s 1 * + 8 ν 2 ( κ 1 + 2 δ 1 ) · m s 2 * 4 ( κ 1 + 2 δ 1 ) ( κ 2 + 2 δ 2 ) - κ 1 κ 2 . (41)
Our gene transcription model coupling with cell division cycles offers six quantities to characterize the fluctuations of mRNA numbers in single cells: The noise η2* and the noise strength Φ* in cells without referring to cell cycle stages, along with η 1 2 * and Φ 1 * in S 1 stage, and η 2 2 * and Φ 2 * in S 2 stage. Theorems 1 and 2 provide the basic formulas by which these quantities can be computed from the system parameters.
The relations between these quantities are far more complicated than our intuition may envisage. We use simple example to demonstrate the delicacy of their relations: Assume (43) and fix the kinetic rates as in (51), and take the stage transition rates κ1 = κ2 = 1/1250 hr−1. Then applying Theorem 2, (38) and (39) gives Φ 1 * = 43 . 2008 and Φ 2 * = 43 . 8410. Since Φ 1 * and Φ 2 * are nearly equal, one might expect by intuition that Φ* is about equal to each of Φ 1 * and Φ 2 *, which is in conflict with
Φ * = 89 . 3329 > Φ 1 * + Φ 2 *
obtained by using Theorems 1 and 2. Moreover Φ* = 89.3329 given here is significantly higher than Φ* ≈ 1 reported in various single cell measurements, including the classical studies by Taniguchi et al. [38] and Yu et al. [39]. In these studies, the genes were active most of time, so that m* and η2* exhibited a strict reciprocal relation, implying Φ* ≈ 1. The stochastic switching between gene promoter ON and OFF states, combined with transitions between cell cycle stages, may induce much noisier transcriptions.
Due to the wide range of the six noise measures and their complex relations, we will discuss their profiles in three special cases below:
Gene transcription involves inherently various probabilistic steps that create fluctuations in mRNA and protein counts [1–3]. The random transitions between the active and inactive promoter states have been widely invoked to explain the fluctuation of mRNA numbers among individual cells of identical genes [6, 7]. Recent experimental studies have revealed that the cell division cycle has global effects on transcriptional outputs, and is thought to be an additional source of transcription noise [10–13].
In this work, we integrated cell division cycles into an extended two-state model to delineate the combined contribution of transcription activities and cell divisions in the variability of transcript counts [4, 6]. In the model, a cell division cycle is divided into S 1 stage before the duplication of a target gene and the late stage S 2, on which the durations are independently and exponentially distributed with rates κ1 and κ2. When a cell divides, each mRNA molecule has an even chance of being partitioned to one of the two daughter cells. We defined two joint probabilities to quantify the system state, and derived the master equations of their time evolutions. From the master equations we obtained the differential equations of the mean and the second moment of mRNA numbers in single cells. By solving these equations we presented in Theorem 1 the steady-state mean transcription level m* in cells, together with the means m 1 * and m 2 * on the two stages. The analytical forms of the second moments are presented in Theorem 2, which in turn help us determine six noise measures: the noise η2* and the noise strength Φ* without referring to cell cycle stages, along with η 1 2 * and Φ 1 * in S 1, and η 2 2 * and Φ 2 * in S 2. The fold change of mRNA counts from S 1 to S 2 is quantified by r * = m 2 * / m 1 *. As a cell contains twice as many copies of each gene in S 2 as that in S 1, one may envisage by intuition that r* ≈ 2. However, our Theorem 3 shows that r* can take any prescribed value in theory, although we also found that r* ≈ 2 when the transcription kinetics are unchanged in the two stages, and stage transitions are considerably slower than mRNA turnover and transcription state transitions. The dependance of r* on κ1 and κ2 is examined deeply in Theorem 4, where the necessary and sufficient conditions for r* < 1 or r* > 2 are identified. In particular, it is proved that r* has an upper bound strictly less than 2 when κ1 ≤ κ2.
We tested the accuracy of our analytical results against various experimental data. For a gene in a mouse embryonic stem cell line, our result predicts r* = 1.2791, which offers a good match with r* = 1.28 ± 0.09 measured in [15]. The analysis also indicates that if the transcription kinetics do not change considerably in the two cell cycle stages, then the average mRNA counts increase about 1 to 2 folds from S 1 stage to S 2 stage as observed in mouse embryonic cells [15] and yeast [12]. The mean m* increases while the noise η2* decreases in each of the cell cycle durations. Rapid transitions between cell cycle stages were identified as a major source of highly noisy transcription. Eukaryotic cells have a DNA dosage-compensating mechanism to reduce mRNA production in late cell cycle stage, resulting in gene transcription homeostasis that overall transcription remains constant across S 1 and S 2 stages [13, 18]. Our analysis reveals that transcription homeostasis does not bring significant changes in transcription noise. If transcription homeostasis is attained by varying a single kinetic rate in the two cell cycle stages, then the homeostasis nearly minimizes transcription noise. As many cellular processes depend on the concentration of enzymes rather than their absolute numbers for proper cellular function [18, 21, 22], we also studied the noise profile when the transcript concentration homeostasis is maintained that the mean transcription level scales with the cellular volume in S 1 and S 2. We found that the transcription noise is relatively stable when the transcript concentration homeostasis is maintained, either by reducing the transcription burst frequency or by increasing the burst size in late cell cycle phase, in the face of a large cell cycle stage duration variation. The reduction degree in the burst frequency is relative robust, while the increase in the burst size is conceivably sensitive, to the large random variation of the cell cycle durations and the gene duplication time.
This work provides one of the first theoretical explorations on how the coupling of stochastic promoter state transitions and cell cycle progressions regulates transcription abundance and noise. It presents a core model for further inclusion of more complex transcription kinetics and cell cycle progressions. The kinetic rates may display large variations in different cell cycle phases or within the same phase, or oscillate periodically in the cell cycle progression [44]. With the expansion of the model, motivated and tested by more upcoming experimental data, the approach initiated here is expected to be developed further to help understand the role played by the cell cycle dependent gene expression in cell functions and cell fate decision [45, 46].
|
10.1371/journal.pgen.1007275 | Identification and functional analysis of glycemic trait loci in the China Health and Nutrition Survey | To identify genetic contributions to type 2 diabetes (T2D) and related glycemic traits (fasting glucose, fasting insulin, and HbA1c), we conducted genome-wide association analyses (GWAS) in up to 7,178 Chinese subjects from nine provinces in the China Health and Nutrition Survey (CHNS). We examined patterns of population structure within CHNS and found that allele frequencies differed across provinces, consistent with genetic drift and population substructure. We further validated 32 previously described T2D- and glycemic trait-loci, including G6PC2 and SIX3-SIX2 associated with fasting glucose. At G6PC2, we replicated a known fasting glucose-associated variant (rs34177044) and identified a second signal (rs2232326), a low-frequency (4%), probably damaging missense variant (S324P). A variant within the lead fasting glucose-associated signal at SIX3-SIX2 co-localized with pancreatic islet expression quantitative trait loci (eQTL) for SIX3, SIX2, and three noncoding transcripts. To identify variants functionally responsible for the fasting glucose association at SIX3-SIX2, we tested five candidate variants for allelic differences in regulatory function. The rs12712928-C allele, associated with higher fasting glucose and lower transcript expression level, showed lower transcriptional activity in reporter assays and increased binding to GABP compared to the rs12712928-G, suggesting that rs12712928-C contributes to elevated fasting glucose levels by disrupting an islet enhancer, resulting in reduced gene expression. Taken together, these analyses identified multiple loci associated with glycemic traits across China, and suggest a regulatory mechanism at the SIX3-SIX2 fasting glucose GWAS locus.
| Type 2 diabetes risk and levels of glucose, insulin, and HbA1c are heritable traits correlated with risk of other diseases and mortality. To identify genetic variants associated with these traits, we studied up to 7,178 men and women from nine provinces in China. We found established variants that could affect fasting glucose located in or near genes named G6PC2 and SIX3. One of the variants at G6PC2 changes the protein sequence and is predicted to affect how the protein functions. Variants located near SIX3 that are associated with levels of glucose are also associated with levels of expression of the genes SIX3 and SIX2 in pancreatic islets. These variants are located in a genomic region predicted to enhance gene expression. We used laboratory assays to show that alleles at one variant, rs12712928, demonstrate significant differences in transcriptional activity, suggesting that this variant influences levels of the SIX3 and SIX2 genes in islets, ultimately increasing glucose levels.
| Type 2 diabetes (T2D) is a chronic disease affecting over 422 million people worldwide[1] with over 30% of cases occurring in East Asian populations [2]. Large-scale genome-wide association studies (GWAS) have identified >100 loci associated with T2D and >80 loci associated with fasting glucose, fasting insulin, and glycated hemoglobin (HbA1c), many of which have also been implicated in T2D susceptibility [3–6]. While the largest GWAS of glycemic traits and T2D to date have been performed in populations of predominantly European ancestry [3, 6–9], other studies have identified glycemic trait and T2D associations in East Asian individuals [5, 10, 11]. As glycemic trait profiles, allele frequencies, and environmental contributions differ between populations, continued investigation of genetic factors can discover additional loci influencing inter-individual variation in fasting glucose, fasting insulin, and HbA1c levels and T2D.
A new resource for genetic analyses, the China Health and Nutrition Survey (CHNS) is an ongoing, household-based, longitudinal survey aimed at examining economic, sociological, demographic, and health questions in a diverse Chinese population [12]. Using a multistage random-cluster design and stratified probability sampling to select counties and cities, data were collected from 228 communities across nine provinces (Guangxi, Guizhou, Heilongjiang, Henan, Hubei, Hunan, Jiangsu, Liaoning, and Shandong) that constituted 44% of China’s population as of the 2009 census. In addition to nearly 30 years of longitudinal survey data collected during 9 survey rounds from 1989–2011, quantitative biomarker measurements and DNA are available on 8,403 subjects in the CHNS.
Individual GWAS loci can harbor multiple association signals. More than one association signal has been reported at G6PC2 and PCSK1 for fasting glucose and at KCNQ1, ANKRD55, CDKN2A/B, DGKB, HNF4A, and CCND2 for T2D [5]. Imputation reference panels generated from large sample sizes can facilitate identification of additional signals. For non-European populations, the 1000 Genomes Phase 3 reference panel is currently the most comprehensive, containing information for more than 88 million variants in >2,500 individuals from 26 diverse populations [13]. Identification of additional association signals at trait-associated loci could explain additional heritability and provide further insights into the biology between the locus and the trait or disease.
GWAS have been an efficient method for studying genetic factors influencing biological mechanisms underlying glycemic traits and T2D, but for many of the identified loci, the underlying gene(s), direction of effect, and disease mechanism are largely unknown [14]. For variants located in non-coding regions of the genome, bioinformatic datasets can be used to annotate and predict regulatory variants, target genes, and direction of effect [15–18], and these variants can be tested for allelic differences in regulatory activity with in vitro laboratory assays [19, 20]. For example, among previous functional studies of variants associated with fasting glucose at the G6PC2-ABCB11 locus, two variants in the promoter were shown to affect G6PC2 expression levels by altering FOXA2 binding, and two variants located in the third intron of G6PC2 were shown to affect G6PC2 splicing [21–23]. However, the majority of the glycemic trait-associated variants have not been examined.
To further clarify the genetic contributions to normal variation in glycemic traits in a multi-provincial Chinese population, we performed a GWAS of fasting glucose, insulin, and HbA1c levels and T2D in subjects from the CHNS, using genetic data imputed to 1000 Genomes Phase 3 in up to 7,178 subjects [12]. We examined the population substructure within the CHNS and evaluated candidate functional regulatory variants at one locus using annotation and in vitro laboratory assays.
To evaluate population substructure among 8,403 CHNS subjects with genotype data available, we constructed principal components (PCs) using a subset of variants (MAF > 0.05; pairwise LD r2 <0.02 in a sliding window of 50 variants). Compared to HapMap 3 populations, the CHNS participants clustered closely with the Han Chinese in Beijing (CHB), the Han Chinese in Denver (CHD), and the Japanese in Tokyo (JPT) populations, with greater diversity than any of these populations (S1A Fig). A comparison to only the East Asian samples showed more clearly that the distribution of the CHNS extends beyond that of the CHB, CHD, and JPT samples (S1B Fig). Within the CHNS, the subjects showed two axes of diversity (Fig 1, S2 Fig). PC1, which explained 4.2% of the variance, appeared to cluster by province, while PC2, explaining 0.6% of the variance, showed diversity among subjects within the Guangxi and Guizhou provinces in southern China. PC2 could be partially explained by differences in self-reported ethnicity, particularly among subjects from the Guizhou province, as PC2 appeared to characterize the Miao and Buyi ethnic groups (S3 Fig). To account for population substructure in subsequent association analyses with glycemic traits, we included PC1 as a covariate and performed analyses using an efficient mixed model approach that accounts for sample structure between individuals [24].
We performed genome-wide association analyses of fasting glucose and fasting insulin levels in up to 8,045,193 genotyped and imputed variants (MAF >0.01) from 5,786 non-diabetic individuals in the CHNS who provided fasting blood samples (S1 Table, S4 Fig). We also performed a genome-wide association analysis of HbA1c in 7,178 nondiabetic individuals who provided fasting or non-fasting samples. In addition, 5,731 unrelated subjects were used to assess variant association with T2D status, including 748 cases and 4,983 controls. For each trait, we also searched for additional signals by conditioning on the lead variants (reciprocal conditional analyses). Overall, a majority of CHNS subjects were female (54%) with a normal BMI (mean = 23.2 kg/m2), and subjects with T2D were older (cases: 59.7 years; controls: 51.2 years) with a higher BMI, higher fasting glucose levels, and higher fasting insulin levels (S1 Table).
Analysis of fasting glucose confirmed eight loci previously identified in East Asian and European samples (G6PC2, SIX3-SIX2, PROX1, ABCB11, GCK, KANK1, GLIS3, and TCF7L2; S2 Table), two of which achieved genome-wide significance (rs34177044, near G6PC2, P = 6.9 x 10−12, Fig 2; rs895636, near SIX3-SIX2, P = 2.3 x 10−8, Table 1, Fig 3A, S5 Fig) [11, 25–27]. At these two loci, we used stepwise conditional analyses to identify additional association signals at a locus-wide threshold of P <1 x 10−5. Conditional analysis including rs34177044 at the G6PC2 locus revealed a second signal (rs2232326, MAF = 0.04, Punconditioned = 1.8 x 10−9, Pconditioned = 2.0 x 10−6, Fig 2). When conditioning only on rs2232326, rs34177044 was attenuated but remained significantly associated with fasting glucose (Pconditioned = 7.0 x 10−9); the attenuation suggests the two signals are distinct yet not fully independent. Haplotype analyses (S3 Table) and regression models containing an interaction term with both variants (P = 0.69) do not suggest a haplotype effect between the two signals, providing further evidence that the two signals are separate. While conditional analyses could be influenced by the moderate imputation quality of rs34177044 (r2 = 0.70) in CHNS, genotypes from the 1000 Genomes project show that the minor allele of rs2232326 is only inherited with the major allele of rs34177044 (East Asian LD r2 = 0.04, D’ = 1.0). No additional association signals were identified at the SIX3-SIX2 locus (S5 Fig).
The lead variant in the second signal at G6PC2 (rs2232326) is a missense variant (S324P). Amino acid 324 is located in a helix spanning the cell membrane [28], and the substitution of a proline for a serine in the middle of a helix may add kinks to the protein [29]. In addition, both SIFT and PolyPhen [30] predict this variant to be “probably damaging”, suggesting that it may affect function of the G6PC2 protein. Based on data from 1000 Genomes Phase 3, rs2232326 is rare in all ancestry populations (MAF: African, 0.2%; Admixed American, 0.3%; European, 0.3%; South Asian, 0.3%) except in East Asians (MAF 5%), and it has few (Admixed American, rs34102076; East Asian, rs139014876), to no (African, European, and South Asian) proxy variants (LD r2>0.80). This variant contributed to a significant G6PC2 gene-based association with glucose in Europeans [31] and other protein-coding variants within G6PC2 have been individually associated with fasting glucose levels (e.g. rs492594, rs138726309, rs2232323) [21]. In the CHNS, rs492594 was nominally associated with fasting glucose levels (P = 0.002); other previously described coding variants were either monomorphic or did not pass imputation quality control thresholds in CHNS (S4 Table).
We examined whether the strength of fasting glucose associations at SIX3-SIX2 and G6PC2 varied by province (Table 2). At rs895636 (SIX3-SIX2), the minor allele frequencies (MAF) differed by as much as 0.12 between provinces. Most of the provinces in which individuals have a relatively lower minor allele frequency (0.35–0.38) showed a stronger association between the variant and fasting glucose levels than similarly sized samples of individuals with higher MAF of 0.43–0.47. The opposite pattern was observed at rs2232326 (G6PC2), for which the province in which the largest MAF (0.09) showed the strongest association with fasting glucose levels. The allele associated with higher levels of fasting glucose trended from less frequent in the northern provinces (MAF = 0.02) to more frequent in the southern provinces (MAF = 0.09). Although allele frequencies between provinces were not statistically different, observed allele frequency differences are consistent with genetic drift and the observed population substructure (Fig 1) [32], demonstrating that study samples from across China have differing power to detect specific associations.
Analysis of fasting insulin validated (P <0.05) two loci previously reported in European and Hispanic samples (PPARG and LOC284930; S5 Table) [33, 34] and did not reveal any genome-wide significant loci (Table 1). The most significant association was identified near CNTN6 (rs13078376, P = 3.22 x 10−7; Supplementary Materials, S6 Fig). Previous studies have demonstrated genome-wide significant associations between variants approximately 1 Mb upstream of the CNTN6 gene and both fasting insulin (rs9841287) [33] and HbA1c levels (rs892295) [35], although the rs13078376 is not a proxy for either of the two previously reported variants (East Asian LD r2<0.01). Data from the CHNS strengthen the evidence for these nominally significant loci near CNTN6.
Analysis of HbA1c validated (P <0.05) nine loci previously identified in East Asians and Europeans (FN3KRP, MYO9B, PIEZO1, ANK1, GCK, SPTA1, HBS1L, MTNR1B, and ABCB11; S6 Table) [10, 36, 37], and did not reveal any genome-wide significant loci. The most significant variant was located within an intron of FN3KRP (rs9895455, P = 3.5 x 10−7; Supplementary Materials, S7 Fig). rs9895455 is in high LD (East Asian, r2 = 0.99) with a variant previously reported to be associated with HbA1c in East Asians (rs1046875) [10]. Three additional variants in high LD with rs9895455 have previously demonstrated moderate associations in Europeans with both HbA1c (P = 4.1x 10−7) and modified Stumvoll insulin sensitivity index (P = 0.02) [38]. While power to detect associations is limited, data from the CHNS provide further support for these established loci.
Association analyses for T2D validated (P <0.05) sixteen loci previously identified in East Asians and/or Europeans (POU5F1/TCF19, SLC30A8, CUBN, MIR17HG, TMEM18, GLP2R, GIPR, MC4R, BCL2L11, PAX4, IGF2BP2, PRC1, KCNQ1, CDKN2A/B, TLE4, and PAM/PPIP5K2; S7 Table) [4, 26, 39–41], and did not reveal any genome-wide significant loci. Of the validated loci, the most significant are at SLC30A8 (rs3802177, P = 3.0 x 10−4) and POU5F1/TCF19 (rs2073721, P = 2.3 x 10−4). The three most significant variants not described previously were located near RTN4RL1 (rs62069176, P = 2.7 x 10−7), SOCS6, (rs2581685, P = 2.6 x 10−6), and ARID1B (rs6557473, P = 3.3 x 10−6) (S8 Table, S8 Fig). Of these, prior suggestive associations (P<0.05) have been detected previously between rs6557473 at ARID1B and both fasting insulin and T2D in Europeans [9, 42, 43]. The CHNS data provide suggestive evidence for these loci.
To aid in the identification of candidate genes at the strongest association signals, we examined whether any of the variants associated with glycemic traits are also associated with expression of nearby transcripts in pancreatic islets, blood, subcutaneous adipose, or tissues from GTEx (S9 Table) [44–46]. These expression quantitative trait locus (eQTL) datasets were generated predominantly from European ancestry donors. GWAS and eQTL signals more clearly coincide when the GWAS variant and the variant most strongly associated with expression level of the corresponding transcript exhibit high pairwise LD (r2>0.80). To allow for differences in LD patterns across ancestries in the GWAS and eQTL datasets, we considered GWAS and eQTL signals to be possibly coincident at a less stringent threshold for pairwise LD values (r2>0.60, East Asian 1000G Phase 3). The association signal for fasting glucose in East Asians at the SIX3-SIX2 locus contained fifteen variants meeting this criterion (lead GWAS variant rs895636 and fourteen variants with East Asian LD r2≥0.60), and the association signal for islet SIX3 expression in Europeans contained 14 variants (lead variant and 13 variants with European LD r2≥0.60) (Fig 3B, S9 Table). One variant, rs12712928, exhibited high LD (r2>0.80) with both the lead GWAS and eQTL variants (Fig 3C). rs12712928-C showed strong association with higher fasting glucose (P = 3.4 x 10−8), similar to the lead fasting glucose GWAS variant (rs895636, P = 2.3 x 10−8), and strong association with lower SIX3 expression level in pancreatic islets (P = 4.7 x 10−8), similar to the lead SIX3 eQTL variant (rs12712929, P = 1.7 x 10−8). In addition to SIX3, rs12712928-C was strongly associated with lower expression level of SIX2 (P = 1.4 x 10−4), SIX3-AS1 (P = 4.8 x 10−6), and two other long non-coding transcripts (S9 Table) [47]. Assuming the fasting glucose GWAS and islet eQTL signals are shared across ancestries, then the strongest candidate variant that may be responsible for both associations is rs12712928.
To establish a set of candidate functional variants at the SIX3-SIX2 locus, we used regulatory chromatin marks (open chromatin and histone states) to predict which variants may affect the transcription of nearby genes. Of 19 candidate variants at the SIX3-SIX2 locus (including the lead GWAS variant rs895636, the lead islet eQTL variant rs12712929, variants in EAS LD r2>0.60 with rs895636, and variants in EUR LD r2>0.6 with rs12712929), only five variants (rs10192373, rs10168523, rs12712928, rs12712929, and rs748947) overlap pancreatic islet active enhancer and pancreatic islet open chromatin (DNase or FAIRE) marks, as well as predicted transcription factor binding motifs (S9 Fig). All five of these variants have EAS LD r2 0.66–0.87 with lead GWAS variant rs895636. These data suggest that these five variants are the strongest candidates to affect transcription of the gene(s) at this locus.
To evaluate the allelic differences in enhancer activity of the five candidate functional variants, we conducted transcriptional reporter assays in MIN6 mouse insulinoma cells. We tested 4–6 independent constructs corresponding to each allele or haplotype for a 312-bp DNA region located 18 kb downstream of SIX3 and 37 kb from the 3’ end of SIX2 spanning rs10192373, rs10168523, rs12712928, rs12712929 (tested as a haplotype) and for a 365-bp region located 20 kb from the 3’ end of SIX3 spanning rs748947 (S9 Fig). While the rs748947 construct showed no enhancer or allele-specific activity (S10 Fig), the haplotype construct had haplotype-differences in enhancer activity in both orientations (Fig 4A). This 4-variant construct containing the fasting glucose-increasing alleles (rs10192373-A, rs10168523-G, rs12712928-C, and rs12712929-T) showed significantly decreased enhancer activity of ≥ 1.4-fold in magnitude (forward, P = 0.0008; reverse, P = 0.0001) compared to the haplotype containing the non-risk alleles. To determine whether rs12712928 could account for the allele-specific effects, we used site-directed mutagenesis to create two additional haplotype constructs. Haplotype constructs containing rs12712928-C exhibited a 1.5-fold decrease in enhancer activity compared to the haplotype constructs containing rs12712928-G (Fig 4B). Taken together, these data show that rs12712928 exhibits allelic differences in transcriptional enhancer activity and suggest it functions within a cis-regulatory element at the SIX3-SIX2 fasting glucose-associated locus.
We next asked whether alleles of rs12712928 or the other three variants differentially affect DNA binding to nuclear proteins. A DNA-protein complex specific to the rs1272928-C allele was observed using electrophoretic mobility shift assays (EMSA) with MIN6 nuclear lysate (S10 Table, S11–S12 Figs). Competition with excess unlabeled C-allele probe more efficiently competed away allele-specific bands than excess unlabeled G-allele probe, providing further support for allele-specificity of the protein-DNA complexes (Fig 4C). Based on these results, we hypothesized that rs12712928-C is located in a binding site for a transcriptional regulatory complex that may be disrupted by the rs12712928-G allele. The sequence containing rs12712928-C is predicted to include a consensus core-binding motif for several transcription factors, and a ChIP-seq peak for CTCF also overlaps this region [48–50]. To identify transcription factor(s) binding to rs12712928, we used a DNA-affinity capture assay. A protein band showing allele-specific binding to the C allele was identified as the alpha subunit of GABP using MALDI TOF/TOF mass spectrometry. In EMSA supershift assays using antibodies to GABP-α, we observed a supershift of the allele-specific band (Fig 4C), suggesting that GAPB may act as a repressor to reduce enhancer activity at this locus (Fig 4D).
We used a similar approach to identify potentially functional variants at the G6PC2 locus. The first signal is comprised of two intergenic variants (GWAS index variant and one variant in LD r2 >0.80). GWAS variant, rs34177044, is ~3.2 kb upstream from the transcription start site of G6PC2 and does not overlap any predicted open chromatin marks. rs1402837 (LD r2 = 0.97) is located 646 bp upstream of the G6PC2 transcription start site and 187 bp and 208 bp upstream, respectively, of other promoter variants previously shown to exhibit allelic differences on transcriptional activity, rs573225 and rs2232316 (S13 Fig) [51]. rs1402837 also overlaps open chromatin marks, suggesting rs1402837 may play a regulatory role in fasting glucose levels in the context of other G6PC2 promoter variants.
In this study of genetic associations with T2D and related glycemic traits in Chinese individuals from 9 provinces in the CHNS, we observed associations with fasting glucose at SIX3-SIX2 and G6PC2, including a coding variant representing an additional signal at G6PC2. We also showed that the SIX3-SIX2 fasting glucose locus colocalizes with eQTL associations for SIX3, SIX2, SIX3-AS1, RP11-89K21.1, and AC012354.6 in pancreatic islets, and we showed evidence that rs12712828 functions as a regulatory variant at the SIX3-SIX2 fasting glucose locus. Genetic associations in CHNS also supported (P<0.05) previously reported associations at 6, 2, 9, and 16 loci with fasting glucose, fasting insulin, HbA1c, and T2D, respectively. The moderate sample size of CHNS prohibited us from identifying additional associations.
Hundreds of genes contribute to the heritability of complex traits [52]. As more GWAS and genome-wide meta-analyses are conducted across genetically diverse populations, identification of additional association signals and loci will help to explain the levels of heritability. The CHNS adds to the growing number of population-based cohorts available for the study of metabolic traits. With its multi-provincial study design, the CHNS includes subjects of differing ethnicities, from both urban and rural areas across China. Additionally, linkage of the genotype data with biomarkers and decades of longitudinal phenotype data (e.g. nutrition, health outcomes, environment) will allow environmental and societal contributions to trait or disease outcomes to be evaluated.
Alterations in regulatory elements or the coding sequence of G6PC2 can impact levels of fasting plasma glucose. G6PC2 encodes an enzyme belonging to the glucose-6-phosphatase catalytic subunit family responsible for the terminal step in gluconeogenic and glyconeogenic pathways that lead to the release of glucose into the bloodstream [53]. Several previous studies have identified >1 fasting glucose association signals at this locus in populations of European and African ancestries, two of which include nonsynonymous coding variants [25, 26, 31, 33, 54–57]. We identified two distinct signals at G6PC2 associated with fasting plasma glucose levels. Variants within the primary CHNS association signal have been associated with fasting glucose in East Asian populations previously [11]. The lead variant in the second signal at G6PC2 (rs2232326) is a missense variant (S324P). We were unable to assess evidence of association with other coding variants in G6PC2 as the variants were either monomorphic in CHNS or did not pass quality control thresholds.
To date, the association between variants near SIX3 and glycemic traits remains specific to East Asian populations. rs895636 was reported as the lead variant associated with fasting plasma glucose in a GWAS of >17,000 Korean and Japanese subjects (P = 9.9x10-13) and in a separate GWAS meta-analysis of up to 46,085 East Asians (P = 2.5x10-13) [27]. However, in Europeans, nominal to no association has been observed between rs895636 and fasting glucose (P = 0.002, n>96,000) [33], HbA1c (P = 0.05, n>46,000) [36], fasting insulin (P = 0.73, n>96,000) [33], and T2D (P = 0.41, n>120,000) [3]. Allele frequency is a possible explanation for ancestry differences. In East Asians, the MAF of rs895636 is 0.42 while the MAF in Europeans is only 0.16. Larger sample sizes of European-ancestry individuals may be needed to identify the association between variants at the SIX3-SIX2 locus and glycemic traits. Other genetic and environmental factors may also be playing a role in the fasting glucose association at SIX3-SIX2 in East Asian populations that are not present in other populations.
We provide compelling evidence that rs12712928 is a regulatory variant at the SIX3-SIX2 fasting glucose locus. The rs895636-T allele is associated with increased fasting glucose levels and decreased SIX3, SIX2, and other transcript expression in pancreatic islets [47]. rs12712928 is in high LD (East Asian and European r2>0.87) with both rs895636 and the lead pancreatic islet eQTL variant, rs12712929, and was the strongest candidate for an effect of regulatory function based on its location in a putative islet enhancer element. Compared to rs12712928-G, the allele rs12712928-C demonstrated decreased transcriptional activity, as well as allele-specific binding to the alpha subunit of GABP, which suggests that at least GABP, and possibly other transcription factors, bind to the C allele and repress expression of SIX3 and SIX2. rs12712928 may be responsible for the GWAS signal, or given that some GWAS signals are affected by multiple functional variants [58, 59], other variants at this locus may also contribute to variation in fasting glucose.
SIX3 is a strong candidate for a target gene at the SIX3-SIX2 fasting glucose locus. Highly expressed in pancreatic islets [44], SIX3 encodes sine oculis homeobox-like protein 3, a transcription factor that localizes to the nucleus of adult beta cells to regulate insulin production and secretion. Decreased expression of SIX3 results in the misregulation (i.e. decreased levels) of insulin [60], which promotes the uptake of glucose into fat, liver, and skeletal muscle cells, thus lowering blood glucose levels [61]. Consistent with the effects of SIX3 in mice, the risk allele rs12712928-C is associated with both decreased expression of SIX3 and increased levels of fasting glucose. In CHNS, rs12712928-C was also moderately associated with decreased fasting insulin levels (P = 7.6x10-5) and an increased risk for T2D (P = 0.03). However, in the islet eQTL data, rs12712928 was also associated with expression level of SIX2, SIX3-AS1, RP11-89K21.1, and AC012354.6. SIX2 is also believed to play a role in the regulation of islet beta cell functions such as insulin output [60]; however, less is known about its biologic function compared to SIX3. Additionally, the roles of SIX3-AS1, RP11-89K21.1, and AC012354.6 are not well characterized. One or more of these transcripts could be a target gene underlying the association signal and contributing to the biological effect on fasting glucose.
In conclusion, this study confirmed many previously identified loci associated with T2D and related glycemic traits and validated a recently described G6PC2 missense variant associated with fasting glucose. We report a functional variant at the SIX3-SIX2 locus, rs12712928, and provide evidence of a potential mechanism by which this variant affects expression of at least SIX3, leading to decreased levels of fasting glucose. Our use of a denser reference panel of >8 million variants in a diverse Chinese population allowed us to conduct higher resolution genetic analyses than reported previously. Further functional analyses of the variants identified in this study is the next step to confirm which variants and genes are affected. Replication of the moderately-significant associations would be useful to better understand the genetic architecture of glycemic traits.
The China Health and Nutrition Survey (CHNS) is a nationwide, longitudinal survey aimed at examining economic, sociological, demographic, and health questions in a Chinese population. Details of subject selection and study design have been described elsewhere [12]. Briefly, a stratified probability sample with a multistage, random cluster design was used to select counties and cities within 9 diverse provinces (Guangxi, Guizhou, Heilongjiang, Henan, Hubei, Hunan, Jiangsu, Liaoning, and Shandong), stratified by income and urbanicity using State Statistical Office definitions. A total of 4,560 households from 228 communities were then randomly selected from within each stratum. Health data was collected during nine rounds of surveys from 1989–2011 (1989, 1991, 1993, 1997, 2000, 2004, 2006, 2009, and 2011). The 2009 survey was the first to collect fasting blood samples. The CHNS was approved by the Institutional Review Boards at the University of North Carolina at Chapel Hill (#07–1963, #05–2369), the Chinese National Human Genome Center at Shanghai (#2017–01), and the Institute of Nutrition and Food Safety at the China Centers for Disease Control (#201524–1). All participants provided written informed consent.
The present analysis was limited to subjects who participated in the 2009 survey round and for whom blood biomarker traits were available (n = 9,551). For the glucose and insulin analyses, subjects were only included if their blood sample was obtained after an overnight fast (n = 6,779). Subjects were excluded from a particular analysis if their biomarker trait value exceeded 4 standard deviations beyond the group mean or have type 1 diabetes. Fasting blood samples were not required for the HbA1c or T2D analysis. Additionally, one member of each first-degree relative pair was randomly removed for analyses of T2D, as current software to analyze associations with binary traits do not control for the high number of related individuals with CHNS.
Following an overnight fast, a blood sample (12 mL) was collected by venipuncture. Glucose and insulin were measured in the central laboratory of the China-Japan Friendship Hospital. Detailed descriptions of the laboratory procedures for measuring glucose (GOD-PAP method; Randox Laboratories Ltd, UK), and insulin (radioimmunology in a gamma counter, XH-6020 analyzer; North Institute of Bio-Tech, China), levels have been described previously [62]. HbA1c was measured in the central laboratory of the China-Japan Friendship Hospital [Guizhou and Hunan, HLC method, HLC-723 G7 (machine), Tosoh, Japan (reagent). Theory: Boronic Acid Afinity HPLC] or in the field [Guangxi and Henan: HPLC method, Primus Ultra 2 (PDQA1c), Primus, USA; Heilongjiang, Hubei, Jiangsu, and Liaoning: HLP method, Bio-Rad (D10), Bio-Rad, USA; Shandong: HLC method, HLC-723 G7, Tosoh, Japan. Theory: Ion exchange HPLC]. Different methods and machines were calibrated with the same quality control products made in Bio-Rad, USA. Participants were classified as having T2D if they were at least 18 years old and met at least one of the following criteria: 1) HbA1c ≥6.5%, 2) fasting blood glucose ≥126 mg/dL, 3) received the diagnosis from a physician after the age of 20 (self-report), or 4) reported taking diabetes medication (self-report) (S1B Table).
DNA samples were extracted and genotyped at the Chinese National Human Genome Center, Shanghai, China. Genotyping was performed with the Illumina HumanCoreExome chip using the standard protocol recommended by the manufacturer. Genotyping was attempted on the 10,131 unique samples, 316 duplicates, and 1 set of triplicates (total n = 10,449). 1,513 samples were unable to be genotyped due to inadequate DNA concentrations (<10 ng/uL or OD 260/280 outside the 1.5–2.0 range), and an additional 69 samples were excluded for poor quality. Using KING, we identified 7 pairs of samples that were unintentionally duplicated; one sample from each pair was excluded. We used PLINK v.1.9 to compare genotype heterozygosity on the X chromosome to self-reported gender and excluded 129 mismatched samples and six samples with apparent XXY or XXXY genotypes. The CHNS data contained 4 sets of identical twins; 1 subject was randomly excluded from each twin pair. Finally, based on the principal component analysis described below, we excluded two samples that were outliers from HapMap samples of Han Chinese in Beijing, China (CHB), Chinese in Metropolitan Denver, Colorado (CHD), and Japanese in Tokyo, Japan (JPT). After exclusions, 8,403 samples were successfully genotyped and passed all genotyping quality control.
We applied variant quality control checks in PLINK v.1.9 on the 8,403 remaining CHNS samples that were successfully genotyped. Of the initial 538,448 variants, we discarded 4,306 variants due to call rate <95% and/or deviation from Hardy-Weinberg equilibrium (P <10−6). Of the remaining 534,143 variants, 193,236 (36.2%) were monomorphic and 340,906 (63.8%) were polymorphic.
Because of the household-based study design of CHNS, we expected the CHNS to include many first- and second-degree relatives. Using KING [63], we calculated kinship coefficients for all pairwise relationships and identified 3,681 first-degree relative pairs and 1,567 second-degree relative pairs.
We performed genotype imputation of the autosomal chromosomes of 8,403 samples with the 1000 Genomes Project Phase 3 v5 reference panel [13, 64] using the Michigan Imputation Server [65]. We used Eagle2 [66] for pre-phasing, followed by imputation with Minimac3 software. We also imputed the X chromosome using with the 1000 Genomes Project Phase 3 v5 reference panel. We imputed male (n = 3,927) and female (n = 4,476) samples separately, using Mach for pre-phasing and Minimac2 for imputation. Imputation yielded data for 47,095,001 variants, and the 534,143 directly genotyped variants were also assigned imputed genotypes. We removed variants with an imputation r2 <0.30 (35,615,501 variants) or a MAF <0.01 (37,891,969 variants) as additional quality control procedures. In total, we tested 8,045,193 variants for association with fasting glucose, insulin, and HbA1c levels and T2D.
We constructed principal components (PCs) to capture population substructure among the CHNS subjects. We identified a set of 55,601 independent variants with MAF > 0.05 and pairwise LD r2 <0.02 in a sliding window of 50 variants and used the variants to construct PCs in 8,403 CHNS subjects (Fig 1; S1–S3 Figs). The set of 55,601 variants was trimmed to match a list of 47,032 variants that were also available in HapMap Phase III samples. Individuals from CHNS and HapMap III were plotted based on the first two eigenvectors produced by the PC analysis (S1 Fig). We tested for the association between each of the first 10 PCs and each of the phenotypic traits to identify PCs associated at P<0.05; the first PC was included as a covariate in the regression models.
Fasting glucose, fasting insulin, HbA1c, and T2D were adjusted for age, age2, BMI, gender, and PC1. Residuals were then inverse normal-transformed to satisfy model assumptions of normality. Efficient mixed model associations (EMMAX) accounting for population structure and relatedness were performed using EPACTs v.3.1.0 [24]. Because EMMAX was designed for analysis of linear traits, GWA analyses for T2D were performed using the Firth bias-corrected logistic regression likelihood ratio test implemented in EPACTs [67]. For all analyses, genotype was modeled as an additive effect, with the genotype dosage values used as the primary predictor of interest. Due to the correlation of the glycemic traits, we used a genome-wide significance threshold of P <5 x 10−8 to define a single result as genome-wide significant, as used in previous association studies of this scale and high trait correlation [68]. A conservative experiment-wide Bonferroni-corrected P-value for four could be considered as P<1.25x10-8. We created regional association plots using LocusZoom [69] with LD estimates generated from the CHNS subjects. All variant positions correspond to build hg19.
At loci that exhibited evidence of genome-wide significant association (P <5 x 10−8), we identified additional association signals using conditional analysis. We added the most strongly associated variant into the regression model as a covariate and tested all remaining regional variants (+/- 1 Mb from the initial lead GWA variant at each locus) for association. Since we were focusing on a much narrower region of variants during the conditional analyses, we set a less stringent locus-significance threshold of P <1 x 10−5 based on ~5,000 variants in a 2 Mb region. We performed sequential conditional analyses until the strongest variant no longer met the P-value threshold.
We used summary data available in the Type 2 Diabetes Knowledge Portal [70] to explore associations between the newly identified loci and other metabolic traits and outcomes. Association summary statistics available (last assessed June 23, 2017) included coronary artery disease from CARDIoGRAM [71]; kidney-related traits from CKDGen [72]; T2D from DIAGRAM, GoT2D, BioMe AMP, CAMP, and SIGMA [3, 43, 73–75]; BMI and waist-hip-ratio from GIANT [76, 77]; and glycemic traits from MAGIC [26, 36, 78, 79]. Additionally, we used data available from the ICP-GWAS (systolic and diastolic blood pressure) [80] and the AGEN adiponectin GWAS [81]. To identify variants in high LD (r2>0.80) with the lead variants, we used LDlink with all East Asian sample populations from the 1000 Genomes Project as the reference [82].
We used ENCODE [15], ChromHMM [83], and Human Epigenome Atlas [84] data available through the UCSC Genome Browser to determine which of the candidate variants in each association signal overlapped open-chromatin peaks, ChromHMM [83] chromatin states, and chromatin-immunoprecipitation sequencing (ChIP-seq) peaks of histone modifications H3K4me1, H3K4me3, and H3K27ac, and transcription factors in pancreatic islets and the pancreas.
We searched the following publicly available eQTL databases to identify cis-eQTLs at the observed loci: GTEx v7 [85], the University of Chicago eQTL browser [86], the Islet eQTL Explorer (http://theparkerlab.org/tools/isleteqtl/) [47], and the Blood eQTL Browser [45]. We also searched for cis-eQTLs in subcutaneous adipose tissue from the METSIM study [87]. All eQTL data sources used a false discovery rate (FDR) <5% for identifying cis-eQTLs, with the exception of the METSIM study, which used an FDR <1%.
MIN6 mouse insulinoma cells[88] were cultured in DMEM (Sigma) supplemented with 10% FBS, 1 mM sodium pyruvate, and 0.1 mM beta-mercaptoethanol. The cell cultures were maintained at 37°C with 5% CO2.
To measure variant allelic differences in enhancer activity at the SIX3-SIX2 locus, we designed oligonucleotide primers (S10 Table) with KpnI and XhoI restriction sites, and amplified the 312-bp DNA region (GRCh37/hg19 –chr2: 45,191,902–45,192,213) around: rs10192373, rs10168523, rs12712928, rs12712929 (tested as a haplotype). Separately, we amplified a 365-bp region (GRCh37/hg19 –chr2:45,192,357–45,192,721) around rs748947. The 312-bp haplotype construct was altered to create a missing haplotype for rs12712928 using the QuickChange site directed mutagenesis kit (Stratagene). As previously described [19], we ligated amplified DNA from individuals homozygous for each allele into the multiple cloning site of the luciferase reporter vector pGL4.23 (Promega) in both orientations with respect to the genome. Isolated clones were sequenced for genotype and fidelity. 2x105 MIN6 cells were seeded per well, and grown to 90% confluence in 24-well plates. We co-transfected five independent luciferase constructs and Renilla control reporter vector (phRL-TK, Promega) using Lipofectamine 2000 (Life Technologies) and incubated for another 48 hours. 48 hours post-transfection, the cells were lysed with Passive Lysis Buffer (Promega). Luciferase activity was measured using the Dual-luciferase Reporter Assay System (Promega) per manufacturer instructions and as previously described [19].
Nuclear cell protein was extracted from MIN6 cells using the NE-PER nuclear extraction kit (Thermo Scientific). 17 bp oligonucleotide probes were designed centered on each variant: rs10192373, rs10168523, rs12712928, and rs12712929 (S10 Table). The annealed double-stranded oligonucleotide biotin labeled and unlabeled probes for both alleles were generated as previously described [19]. To conduct EMSAs, we used the LightShift Chemiluminescent EMSA Kit (ThermoFisher Scientific) and followed the manufacturer’s recommendations. Briefly, a 20 μl binding reaction consisting of 6 μg nuclear extract, 1X binding buffer, 50 ng/μL poly (dI-dC), and 200 fmol of labeled probe was incubated at room temperature for 25 minutes. For competition reactions, 25-fold excess of unlabeled probe for either allele were incubated for 15 min prior to the addition of 200 fmol labeled probe and incubated for an additional 25 minutes. For supershift assays, 6 μg of polyclonal GABP-α antibody (sc28312X; Santa Cruz Biotechnology) was added to the binding reactions and incubated for 25 minutes prior to the addition of 200 fmol labeled probe. The reaction was further incubated for an additional 25 minutes. Protein-probe complexes were resolved on non-denaturing PAGE on 6% DNA retardation gels (Thermo Scientific), transferred to Biodyne B nylon membranes (PALL Life Sciences), cross-linked on a UV-light cross linker (Stratagene), and detected by chemiluminescence. EMSAs were carried out on second independent day and yielded comparable results.
To identify factors in the protein complex binding rs12712928, we conducted a DNA affinity capture assay as previously described [19]. Briefly, the 450 μL binding reactions consisted of 300 μg of pre-cleared, dialyzed MIN6 nuclear extract, 1X binding buffer, 50 ng/μL poly (dI-dC), and 40 pmol of biotin-labeled probe for either rs12712928 allele (same as EMSA probes) or a scrambled control. Binding reactions were incubated at room temperature for 30 min on a rotator, and then 100 μL of streptavidin-magnet Dynabeads were added to the reaction and incubated for an additional 20 minutes. Beads were washed and bound DNA-proteins were eluted in 1X reducing sample buffer. Proteins were separated on NuPAGE denaturing gel and allelic differences in protein bands was visualized with Coomassie G-250 staining. The UNC Michael Hooker Proteomics Center used a Sciex 5800 MALDI-TOF/TOF mass spectrometer to identify the proteins in the excised protein bands.
|
10.1371/journal.pcbi.0030057 | Posttranscriptional Expression Regulation: What Determines Translation Rates? | Recent analyses indicate that differences in protein concentrations are only 20%–40% attributable to variable mRNA levels, underlining the importance of posttranscriptional regulation. Generally, protein concentrations depend on the translation rate (which is proportional to the translational activity, TA) and the degradation rate. By integrating 12 publicly available large-scale datasets and additional database information of the yeast Saccharomyces cerevisiae, we systematically analyzed five factors contributing to TA: mRNA concentration, ribosome density, ribosome occupancy, the codon adaptation index, and a newly developed “tRNA adaptation index.” Our analysis of the functional relationship between the TA and measured protein concentrations suggests that the TA follows Michaelis–Menten kinetics. The calculated TA, together with measured protein concentrations, allowed us to estimate degradation rates for 4,125 proteins under standard conditions. A significant correlation to recently published degradation rates supports our approach. Moreover, based on a newly developed scoring system, we identified and analyzed genes subjected to the posttranscriptional regulation mechanism, translation on demand. Next we applied these findings to publicly available data of protein and mRNA concentrations under four stress conditions. The integration of these measurements allowed us to compare the condition-specific responses at the posttranscriptional level. Our analysis of all 62 proteins that have been measured under all four conditions revealed proteins with very specific posttranscriptional stress response, in contrast to more generic responders, which were nonspecifically regulated under several conditions. The concept of specific and generic responders is known for transcriptional regulation. Here we show that it also holds true at the posttranscriptional level.
| Large-scale mRNA concentration measurements are a hallmark of our post-genomic era. Usually they are taken as a surrogate for the corresponding protein concentrations. For most genes, proteins are the actual cellular players, but up to now it has been much more difficult to measure protein concentrations than mRNA concentrations. However, due to numerous posttranscriptional regulation mechanisms, mRNA levels only partly correlate with protein concentrations. Based on thoroughly composed reference datasets for protein and mRNA concentrations in yeast under standard growth conditions, we report the best corresponding correlation so far. We took into account additional factors, beyond mRNA concentrations, that influence protein levels in order to improve protein level predictions. Extending our previous approach, where ribosome occupancy and ribosome density were considered, we now also consider ORF-specific translation elongation rates. Different measures for elongation velocity were examined, and the codon adaptation index was found to be most appropriate. Moreover, saturation kinetics were introduced to better describe the translation process. The general findings were also applied to four stress conditions. Three new concepts, translation on demand, just-in-time translation, and general and specific posttranscriptional stress responders, are discussed.
| Although mRNA concentrations are widely used as a surrogate for protein abundances, studies comparing mRNA and protein expression on a global scale indicate that mRNA levels only partly correlate with the corresponding protein concentrations [1–12]. It has been estimated that protein concentrations are determined by the corresponding mRNA concentrations by only 20%–40% [11,13]. Thus, for a better interpretation of results obtained from mRNA measurements, a deeper understanding of translational regulation is urgently required [14–16].
To study the fundamental role of posttranscriptional regulation, we focused on S. cerevisiae as one of the most thoroughly investigated model organisms, where mRNA concentrations and even protein concentrations are available for most genes. More specifically, we were interested in (i) identifying the most important factors regulating translation rates and (ii) specific translational regulation under different conditions. The translation rate is proportional to the translational activity (TA) [9,10,17], which we previously calculated as the product of mRNA abundance, ribosome occupancy, and ribosome density [10]. Ribosome occupancy (ribocc) is the fraction of mRNA molecules with at least one ribosome, and the ribosome density (ribden) is the number of ribosomes on active mRNAs divided by the transcript length [18]. Hence, ribden takes into account that longer transcripts take longer to be translated, and require a larger number of bound ribosomes to achieve the same synthesis rate (number of new proteins per time).
Here, we additionally account for ORF–specific translation elongation velocity. It depends on the amino acid composition of the corresponding protein and the availability of the needed tRNAs. We discuss different measures for the elongation velocity based on tRNA concentrations. Interestingly, we found that the codon adaptation index (CAI) was the best measure for the speed of translation elongation, because it improved the correlation between TA and protein concentrations more than any of the other measures tested. The CAI was initially introduced as a measure for selection of optimal codons in ORFs based on highly abundant mRNAs [19]. It is defined as the geometric mean of the relative synonymous codon usage (RSCU) values for all codons of a given ORF, normalized by the maximum possible mean RSCU value (the RCSU value for a codon is the observed codon frequency divided by the expected frequency for equal codon usage). The CAI has also been used to predict protein concentrations [1,2].
Whereas up to now linear kinetics have been assumed for the TA [8,10,20], here we demonstrate that accounting for nonlinear saturation improves the protein concentration prediction: the TA–protein correlation is improved assuming Michaelis–Menten kinetics for the three factors influencing the translation initiation. Using our newly calculated TAs and our newly composed reference dataset for protein concentrations, we were able to deduce degradation rates for 4,125 proteins. Comparison of our predicted values with measured protein half-lives [21] shows that including the CAI and accounting for saturation significantly improved our predictions.
Previous studies on posttranscriptional regulation that included protein concentration measurements in yeast focused either on standard conditions [1,5,8,10] or dealt with just one stress condition [3,4,6,9]. Other studies only measured ribden changes without considering the respective protein concentration changes [22]. Here, we present the first comprehensive analysis of different stress conditions by combining existing experimental data. For this purpose we used all four published large-scale datasets that tested the relative change of both mRNA and protein abundances upon exposure to different stress conditions [3,4,6,9]; two other studies were published after completion of this work and could therefore not be analyzed in detail here [15,16].
Our analyses of the stress data support the finding that considering saturation kinetics improves the quantification of posttranscriptional regulation. We also confirmed the previously introduced concept of translation on demand, concerning proteins that are quickly needed in response to a (stress) stimulus [10]. In such situations the usual order of events, with transcription and subsequent translation, may be too slow for an appropriate physiological reaction. Instead, the cell might keep a constant level of reservoir mRNA, which is blocked for translation. Processing bodies (P-bodies) may play a role in mRNA storage. It has been shown that processing bodies accumulate mRNAs for subsequent degradation [23], but they may also store mRNA for later translation [24]. Translation initiation might also be blocked by 5′ binding proteins or alternative 5′ leaders [25]. If the corresponding protein is rapidly needed (for instance, after cell exposure to perilous conditions), the cell can then immediately start with translation. The detailed analysis of 62 proteins that were measured under all four conditions confirmed the existence of translation on demand and identified additional candidates. Another means for fast stress response is the continuous synthesis and destruction of proteins under normal conditions. Upon stress the protein turnover can be stopped to quickly elevate protein concentration [26]. Since this mechanism does not change the translation rate, it is not considered to be translation on demand.
Based on the available experimental data, it is possible to separate proteins with distinct posttranscriptional regulation under one specific condition from others that are regulated in a more generic way. This extends the previous notion of generic and specific stress response from the level of transcriptional to posttranscriptional expression regulation [27].
Our first goal was to establish a reliable set of experimental data to investigate the correlation between mRNA abundance and observed protein concentrations. We integrated various published datasets for S. cerevisiae [1,2,7,28] and obtained protein and mRNA concentrations under standard conditions for 4,152 ORFs, representing the largest corresponding dataset so far. Correlations between protein concentrations and other properties were computed using the Spearman rank correlation coefficient rs. The rs was preferred over the Pearson correlation coefficient, because the former makes no assumptions about the underlying distributions of the variables. It has previously been shown to perform better for the analysis of this kind of data [2]. The global rs between mRNA and protein abundance was 0.63, which is the best large-scale correlation reported for yeast so far [8,10,16]. This high correlation also underlines the quality of the integrated dataset. After completing the computational part of this study, another large-scale protein concentration dataset was published [15]. This new dataset could not be included in this study. However, we computed the correlation of our integrated dataset to those independent measurements. Our integrated protein concentrations were more strongly correlated to those measurements (rs = 0.65) than any of the individual input datasets alone (rs ranging from 0.39 to 0.62). We thus conclude that our protein dataset is comparably robust and that our conclusions are unlikely to change significantly if the new data were included. However, the remaining deviations between those measurements reenforce the finding of Newmann and co-authors that biological noise in protein concentrations can be considerable [15].
We quantified the relative importance of various factors for the determination of TAs (Figure 1). Considering additional factors always improved the correlation compared with using mRNA levels alone (rs increases from 0.63 to 0.70; Figure 1). In addition to mRNA concentration, ribocc, and ribden, we also took into account sequence-based aspects of elongation efficiency. In Escherichia coli, translation of major codons occurs 3-fold to 6-fold faster [29,30] and 10-fold more accurately [31] than translation of minor codons, which also reduces the cost of GTP-dependent proofreading [32]. We thus used the CAI as an indirect measure of elongation velocity, because it is easily accessible for all yeast ORFs. Figure 1 shows that accounting for the CAI further improved the correlation to protein concentrations (rs = 0.68). Most of the factors contributing to TA are correlated with each other, e.g., CAI correlates with mRNA levels. To demonstrate that each factor independently contributes to the TA, we computed partial correlation coefficients for all relevant combinations of factors (Table 1). This analysis shows that (i) every factor independently carries significant information about the translation rates and (ii) the CAI and the ribosome-related factors (ribocc × ribden) contributed about equally to the overall TA. Further, we determined the significance of regression improvements by randomly subsampling the datasets. Figure 1 shows the standard error of the regression coefficients obtained by randomly subsampling the datasets (two-thirds of the proteins were randomly sampled). To determine if the regression improvements were caused by only few proteins (outliers), we checked if the improvements also hold for the subsamples (see Methods). We tested 1,000 random subsamples and always observed relative improvements when considering more factors for TA prediction.
The CAI is exclusively based on the frequency of codons in highly expressed genes. Thus, it only indirectly accounts for tRNA availability. A direct quantification of tRNA concentrations for the respective codons might give an improved descriptor for an ORF's tuning for fast elongation. To directly account for tRNA availability, we introduce the tRNA adaptation index (tRNA–AI) ([33]; for calculation, see Methods). However, problems for the calculation of tRNA–AIs are the lack of measured tRNA concentrations and the ambiguous assignment of the 42 yeast tRNAs to the 64 codons. We tested two different tRNA–AIs (see Methods). The correlations to the CAI are rs = 0.91 for tRNA–AI_p and rs = 0.55 for the tRNA–AI_c. Next, we replaced the CAI in TA2 (Equation 2) by either of the tRNA–AIs to test if the tRNA–AIs are equally or more predictive than the CAI. Figure 1 shows that the strict assignment (tRNA–AI_p) yielded a protein–TA correlation slightly better than the correlation with TA1 (rs = 0.66 versus 0.65), whereas using tRNA–AI_c yielded no improvement compared with TA1 (rs = 0.65). Hence, accounting for tRNA–AI_p improved the predicted translational activity, while tRNA–AI_c was not predictive for TA. Thus, (i) the tRNA concentration is indeed an important factor for translation, and (ii) the speed of elongation seems to be mainly determined by the availability of tRNAs with perfectly matching anticodons. This observation was supported by the findings of Rocha [34], who showed that the perfect match model is more likely to mimic in vivo conditions than the frequency model. Yet, the protein–TA correlation based on the CAI was still better than the one based on the tRNA–AI_p. This suggests that the CAI contains information beyond the tRNA gene copy number. For instance, the codon–anticodon interaction strength might also affect the efficiency of elongation [34,35]. Therefore, all following estimations of TA are based on CAI, and the tRNA–AIs were not used any further.
So far we assumed a linear relationship between the TA and the factors defining it. However, because energy, as well as the numbers of tRNAs, amino acids, and ribosomes available in a cell, are limited, actual TA may saturate for high mRNA concentrations. To test this hypothesis, we computed TAs using different possible kinetic relationships, which were all based on Michaelis–Menten kinetics (Figure 2). Ribosome density was the only single factor that showed a slight saturation effect. However, by combining mRNA concentration, ribocc, and ribden in one saturation term (named TA3, see Equation 3), we significantly enhanced the protein–TA correlation (see Methods and Figure 2). This combined term quantifies the number of initiation events for the given ORF.
Several lines of evidence indicate that the small improvement obtained by accounting for saturation is of true biological relevance: first, we wanted to know if the saturation may be biased by a specific experimental technique. To test for such bias, we divided the dataset into a training set and a test set based on different measurement techniques (Figure S1 in Protocol S1). This analysis showed that, even if we train the Michaelis constant, Km, on protein concentrations obtained with one experimental method, it also improved the correlation for protein concentrations from other experimental methods. Thus, the type of kinetics suggested by Equation 3 is independent of the experimental techniques employed. Second, we randomly split the data into training (two-thirds) and test (one-third) sets. We determined Km based on the training data and applied it to the test data. We repeated this test 50 times, and the resulting correlations of TA3 in the test data were consistently high (Table 2). The Km values determined for TA3 were remarkably stable (coefficient of variation CV = 0.18, Table 2). As opposed to that, the Km values for the alternative models were more sensitive, which additionally supports that our model better fits the observed data. Finally, we analyzed the respective correlations for different functional groups of proteins (Figure S2 in Protocol S1). We observed at least a slight improvement when using TA3 for 17 out of 18 functional groups. The final TAs predicted for 6,063 ORFs, as well as respective protein half-life descriptors (PHD, see Methods) for 4,125 ORFs, are provided in Table S1. The PHD is only a relative descriptor of protein half-life [10], but we expect that the predicted PHDs correlate with measured protein half-lives. We compared our PHDs with recently published protein half-lives for yeast [21]. PHDs determined with TA1 or TA2 were not significantly correlated with the measured protein half-lives (|rs| < 0.1, p > 0.01), whereas PHDs based on TA3 exhibited a weak but significant correlation (rs = 0.24, p < 10−40). The predicted PHDs are based on five values, most of which are noisy (except for CAI); the measured protein half-lives are also subject to biological noise. Nevertheless, the above significant correlation provides confidence that protein stability prediction via PHD is possible—at least within certain limits. We expect that future PHDs will improve further as more and more precise data become available.
To systematically identify candidates for translation on demand, we developed an integrated score. Translation on demand is indicated under normal conditions by (i) a small number of ribosomes translating a given ORF, and (ii) by a small number of proteins produced per transcript. Thus, ribocc × ribden serves as a first indicator, and the ratio of proteins per mRNA (PRR) is used as the second predictor. A low value for both, ribocc × ribden and PRR, indicates strong evidence for translation on demand. Next, we computed an integrated score to rank genes with respect to their potential for translation on demand on a single scale. This score is defined as the weighted sum of ribocc × ribden and PRR (both normalized by the median values). A low value of this score is indicative for translation on demand. We also tested other scores, but they performed significantly worse than this scoring scheme. For instance, the (weighted) product of ribocc × ribden and PRR may yield low scores even if one of the two descriptors is high. This is particularly problematic for proteins with short half-lives, because they might exhibit low PRRs without actually being translated on demand.
Due to the necessity for relatively fast responses, components of signal transduction chains are probable candidates for regulation at the translational level (in addition to covalent modifications governing their activities). To test this hypothesis, we assessed the overlap of the top 100 and the top 500 genes according to our score with gene categories from the Munich Information Center for Protein Sequences (MIPS) (http://mips.gsf.de) and the Saccharomyces Genome Database (SGD) (http://www.yeastgenome.org). As shown in Tables S3 and S4, these top-scoring candidates were indeed significantly enriched for genes with functions related to signal transduction and, interestingly, to transcriptional regulation. This suggests that transcription factors are among the first proteins that are rapidly synthesized upon detection of stress signals. We also analyzed the correlation of protein and mRNA concentrations within groups of genes with similar functions (functional modules [10]). Figure S2 in Protocol S1 shows that modules related to signal transduction (i.e., those likely to contain many proteins that are subject to translation on demand) exhibit particularly weak protein–mRNA correlations. However, after accounting for posttranscriptional regulation, those correlations improved. Hence, the protein–mRNA correlation in regulatory modules gets distorted by posttranscriptional processes.
Next we asked whether our results, obtained under standard conditions, also hold under stress conditions. We considered all four available datasets with measured mRNA and protein concentrations under standard and stress conditions (see Methods). Unfortunately, most studies did not provide ribosome densities and occupancies, so we only considered changes in protein and mRNA concentrations under stress.
Both mRNA and protein concentration changes were measured for 1,216 ORFs in response to at least one of the four stimuli. When a given ORF is translated on demand under one of the tested conditions, one expects to see a greater change in its protein concentration than in its mRNA level. Therefore, the ratio of protein to mRNA concentration changes was computed for all available genes and conditions. These ratios served as additional evidence for translation on demand, and may help identify targets for future research. Accordingly, we took a closer look at high-ranking genes (rank < 500) exhibiting strongly elevated protein concentration changes under at least one of the four tested stress conditions. Concentrations of 20 proteins were elevated at least two times more than their corresponding mRNA (Table S5). It would be most interesting to analyze the posttranscriptional regulation of these genes more in detail in the light of these results.
Figure 3 shows the protein change (PC) and change of translation rate (TC) for different stimuli and functional groups. Viewing the data this way emphasizes the high specificity of transcriptional and posttranscriptional responses under different conditions. Whereas a lack of amino acids (Figure 3C) induces dramatic posttranscriptional changes, the mating pheromone induces stronger transcriptional changes (Figure 3D). The figure also highlights the functional difference in protein concentration regulation. For instance, proteins involved in signal transduction (group ST) are clearly strongly upregulated at the posttranscriptional level in the minimal medium. This once more underlines the importance of fast posttranscriptional changes of protein concentrations (e.g., translation on demand) especially for signaling proteins. Proteins involved in protein synthesis, on the other hand (group PS), were downregulated, presumably because the overall protein synthesis was reduced in minimal media. When switching the energy source (Figure 3A), changes in protein synthesis were less drastic, while energy-related proteins (group EG) changed most at both the transcriptional and posttranscriptional levels.
The experimental datasets contain 62 ORFs with mRNA and protein ratios measured under all four stress conditions. By combining these data, we were able to analyze the differential regulation of these proteins under several conditions. One noteworthy candidate for translation on demand is YPL028w (also known as ERG10, LPB3, or TSM0115), which codes for an acetyl-CoA C-acetyltransferase. This enzyme is involved in the first step of mevalonate biosynthesis [36]. Accordingly, it was substantially upregulated in the galactose experiment (6.7-fold change) without significantly changing its mRNA concentration (0.7-fold change). Another protein catalyzing the first step of a pathway (serine and glycine biosynthesis) is Ser3p [37]. Although the mRNA concentration was also upregulated (4.2-fold change) upon exposure to minimal medium [4], the increase in protein concentration was much higher (18.2-fold change). Whereas it is well-established that genes are transcribed in the order of the appearance of their products in metabolic pathways (just-in-time transcription [38,39]), YPL028w and Ser3p represent the first hints for a similar regulation at the translational level (just-in-time translation). Thus, enzymes catalyzing important first steps in metabolic pathways might also be regulated at the translational and post-translational level (e.g., allosteric control or covalent modifications) to ensure a fast cellular response.
Table 3 shows the PC/TC ratio for selected proteins. The complete table, with all 62 proteins measured under all four conditions, is shown in Table S6. On the top of the table are all proteins whose maximum PC/TC is at least 2.5 times higher than for any of the other conditions—these are specific posttranscriptional responders. At the bottom of Table 3 are general responders, i.e., proteins without a distinct posttranscriptional upregulation under one specific condition. Hence, there are two classes of proteins: the first class contains proteins that are strongly posttranscriptionally upregulated under just one of the four tested conditions. These proteins are likely to be stress specific. The second class contains proteins that are either not distinctly regulated at the posttranscriptional level (i.e., the protein and mRNA changes were similar) or that are upregulated under several stress conditions. We found that proteins involved in amino acid metabolism or protein fate exhibit condition-specific posttranscriptional regulation, in contrast to proteins with more generic cellular functions (those at the bottom of Table 3, e.g., proteins involved in translation). Importantly, this pattern only emerged after accounting for the saturation of translational activity (Equation 3, Table S6).
It is increasingly recognized that mRNA abundances are only a weak surrogate for the corresponding protein concentrations [13], and it has been proposed that posttranscriptional control of gene expression is at least as important as the better-studied transcriptional regulation [14]. Our work contributes to a better understanding of posttranscriptional regulation by taking into account as much information as possible in addition to mRNA concentrations. In our previous work [10], we did not consider any data describing the translation elongation velocity. Here we introduced the CAI as an additional factor and we demonstrated that it is currently the best corresponding measure. A systematic analysis of all TA factors reveals that the CAI contributes additional independent information for understanding translation rates, and it is at least as important as ribocc and ribden together. Moreover, for the first time we tackled the problem of the functional relationship between the TA and the contributing factors. The proposed Michaelis–Menten kinetics implies that the concentrations of highly abundant mRNAs have to change much more drastically to achieve a significant change of protein concentrations. The transcripts of many signalling proteins, like transcription factors, are often expressed at comparably low concentrations, facilitating sensitive and significant changes in response to stress. These findings are also in line with the previous observation that protein concentrations tend to be less noisy if transcript levels are high [20]. We found that only the initiation-related factors (mRNA concentration, ribocc, and ribden) were subject to saturation, whereas the CAI contributed linearly to TA. In other words, once translation started, it progressed as quickly as permitted by the sequence [18]. Importantly, our results imply a protein-specific saturation, as opposed to a global reduction in translation (e.g., due to energy exhaustion). First, each protein requires a distinct set of amino acids, and hence also a distinct set of tRNAs, for its synthesis. Hence, these resources could be exhausted if certain mRNAs were excessively translated. Second, translation is often conducted in a site-specific manner, i.e., transcripts are transported to specific cytoplasmic sites where the protein products are needed [40,41]. Excessive translation can therefore exhaust resources in those cellular regions, whereas translation may remain unaffected at other sites. In this context it is also important to remember that log-growth conditions are not the normal environmental conditions to which yeast cells have been adapted. Natural conditions are much more characterized by nutrient limitations. Hence, it is likely that several proteins are synthesised beyond their optimal limits in fast-growing cells under ideal lab conditions.
It is well-established that the input data used for this study are noisy [5,8,10]. General conclusions would be affected by a systematic bias caused by the noise in the data. Newman et al. [15] found a correlation between protein abundance and biological variability. However, such bias does not affect average concentrations for populations of cells. In fact, Newman et al. report a good correlation of their protein concentrations with previous population-based measurements. It should be noted that our main results are robust to noise in the data: an additional ORF-specific factor accounting for the speed of translation always improves the TA–protein correlation, regardless of whether we use the tRNA–AI or the CAI (Figure 1). Also, there is an improvement of TA–protein correlations by using saturation kinetics for a range of about two orders of magnitude for the Km value (Figure 2).
Based on our newly calculated TAs, we propose degradation rates for 4,125 proteins. Comparison with the recently published study of the first large-scale measurement of protein turnover [21] reveals that our calculation outperformed previous approaches [10]. Deviations between our predicted values and the measurements were partly due to noise in the data, but they might also pinpoint potential additional posttranscriptional control steps, which should trigger more detailed investigations of these ORFs.
The consideration of all available large-scale data on stress response in S. cerevisiae enabled us to confirm the previously introduced concept of translation on demand [10]. Additionally, based on the analysis of all 62 genes with measured protein concentrations under all four conditions, we demonstrated the first evidence to our knowledge for generic and specific posttranscriptional stress responders. Several proteins that were posttranscriptionally upregulated under only one of the tested conditions might, of course, also respond under other, yet-untested conditions. However, the distinct patterns that already emerged based on the available data indicate that cells use similar regulatory schemes of generic and specific responses to tackle threats at the transcriptional and posttranscriptional level. Many of the translation-on-demand candidates did not show any significant upregulation under any of the tested conditions. The majority of them were not even measured under all four conditions. Also, these conditions only represented a small subset of the possible threats that yeast has adapted to. Clearly, the investigation of the posttranscriptional stress response is lagging behind the corresponding analysis at the transcriptional level. By combining all available information, it might be possible to nail down those conditions under which the translation-on-demand candidates respond at the posttranscriptional level and to experimentally verify the predictions. The in silico analyses presented here will help to streamline those experimental efforts.
A complete list of all data used is presented in Table S1. Only genes occurring in MIPS and/or SGD were considered. The mRNA concentrations for standard conditions were taken from our previous work [10], which were derived from a pool of 36 independent mRNA abundance measurements from different research groups. Protein concentrations for standard conditions of four measurements [1,2,7,28] were normalized by nonlinear regression, and the median was taken as the reference value for each ORF (Table S1). The following equations were used to map the reported measurement value (meas) onto a common scale:
Ghaemmaghami et al. [7]: selected value = original data
Prot Futcher et al. [2] [103 copies/cell]: selected value = 9,710.7 ·
meas0.4293
Prot Gygi et al. [1] [103 copies/cell]: selected value = 10,977 ·
meas0.2662
Prot Liu et al. [28] [relative abundance]: selected value = 2,108.2 ·
meas0.516
The four stress datasets were taken from [3] (shift from glucose to galactose), [6] (ethanol), [4] (minimal medium), and [9] (exposure to pheromone). Genes were grouped according to the functional protein classification in MIPS, http://mips.gsf.de/genre/proj/yeast/, where one gene can be assigned to several groups.
The TA is a measure for the ORF-specific translation rate. The true translation rate is the product kp × TA [10], where we assume an ORF-independent rate constant, kp. We tested different variants of estimating the TA from measured data. The first (TA1) has been suggested previously [9,10]:
where mRNA is the mRNA concentration of that gene [9,18]. Next, we included the CAI:
Finally, we assumed Michaelis–Menten kinetics:
The Km for the Michaelis–Menten kinetics (Km = 0.06) was determined by maximizing the correlation between TA3 and protein abundance (Figure 2).
The protein half-life descriptor, PHD, was calculated according to [10], but using TA3 as an improved measure of translational activity:
where prot is the reference protein concentration of the corresponding gene (see the section Data used).
When comparing the predicted PHDs with measured protein degradation rates from [21], we used two different datasets: the first set contained all proteins with the half-lives as reported by [21], and the second set contained only proteins with half-lives shorter than 300 min. The second dataset was used because short half-lives are more reliable [21]. Correlations with both datasets were very similar, especially the significant correlations (rs = 0.2427 and 0.2425, respectively).
Correlations were quantified using the Spearman rank correlation coefficient, which is in agreement with previous studies [5,10]. All reported correlations were based on at least ten data points. Although some of the correlations in Figure 1 can be computed for more proteins than others, all correlations reported were performed on the same set of proteins to avoid biases due to different sample sizes.
All correlations reported in Figure 1 were significantly different from zero (p < 10−16). Variability of the correlation coefficients was tested by randomly subsampling two-thirds of the proteins. The rs was computed for each subsample, and its standard error was computed (Figure 1). To test the significance of correlation improvements, the correlations (rs|mRNA, rs|TA1, rs|TA2, rs|TA3) were compared for each subsample individually. In all 1,000 subsamples, rs|TA1 > rs|mRNA, rs|TA2 > rs|TA1, and rs|TA3 > rs|TA2. This test demonstrated that the respective improvements were not dependent on some specific outliers. We also tested for the significance of individual factors by computing partial correlation coefficients (see main text and Table 1).
The definition of the tRNA–AI is similar to the CAI [19], whereby the RSCU value was replaced by the gene copy number (GCN) of the corresponding tRNA. Under normal growth conditions, the GCN can be used as a measure of tRNA concentrations [42]. The relative adaptation value wk of a codon k is the GCN of that tRNA compared with the maximal GCN for that amino acid:
Two assignments of tRNAs to codons were tested. According to [42], each codon gets assigned only one tRNA (the perfectly matching tRNA according to Watson–Crick base pairing, w_p). The wobble rule introduced by Crick [43] assumes that some codons can be recognized by several tRNAs. In the corresponding second model, the GCNs were added up (w_c) [33].
In accordance with the definition of the CAI, we defined the tRNA–AI of a gene with L amino acids as the geometric mean of wk:
|
10.1371/journal.ppat.1006846 | A new cell culture model to genetically dissect the complete human papillomavirus life cycle | Herein, we describe a novel infection model that achieves highly efficient infection of primary keratinocytes with human papillomavirus type 16 (HPV16). This cell culture model does not depend on immortalization and is amenable to extensive genetic analyses. In monolayer cell culture, the early but not late promoter was active and yielded a spliced viral transcript pattern similar to HPV16-immortalized keratinocytes. However, relative levels of the E8^E2 transcript increased over time post infection suggesting the expression of this viral repressor is regulated independently of other early proteins and that it may be important for the shift from the establishment to the maintenance phase of the viral life cycle. Both the early and the late promoter were strongly activated when infected cells were subjected to differentiation by growth in methylcellulose. When grown as organotypic raft cultures, HPV16-infected cells expressed late E1^E4 and L1 proteins and replication foci were detected, suggesting that they supported the completion of the viral life cycle. As a proof of principle that the infection system may be used for genetic dissection of viral factors, we analyzed E1, E6 and E7 translation termination linker mutant virus for establishment of infection and genome maintenance. E1 but not E6 and E7 was essential to establish infection. Furthermore, E6 but not E7 was required for episomal genome maintenance. Primary keratinocytes infected with wild type HPV16 immortalized, whereas keratinocytes infected with E6 and E7 knockout virus began to senesce 25 to 35 days post infection. The novel infection model provides a powerful genetic tool to study the role of viral proteins throughout the viral life cycle but especially for immediate early events and enables us to compare low- and high-risk HPV types in the context of infection.
| Current cell culture models for the study of the human papillomavirus (HPV) life cycle depend on immortalized keratinocytes harboring episomal HPV genomes. However, the requirement for immortalization restricts the study to only a few HPV types and does not allow investigating immediate early events of the viral life cycle. Despite many efforts, efficient infection of primary keratinocytes has not been achieved until now. Using pre-binding of virus to extracellular matrix deposited by keratinocytes, we now achieve very efficient infection of primary keratinocytes. The infection model allows studying the complete viral lifecycle. It could be extended to HPV types that do not immortalize keratinocytes and allows an extensive genetic screen of the contributions of viral factors throughout the viral lifecycle. It should aid the investigations of processes leading to HPV-induced immortalization.
| High-risk HPV types such as HPV16 are the infectious agents most commonly associated with human cancers such as but not restricted to cervical and oropharyngeal squamous cell carcinoma. Approximately 5% of all human cancers can be linked to HPV infection. HPV encodes two major viral oncoproteins, E6 and E7, which drive immortalization and transformation of HPV infected cells. Their roles in cancer development can be mostly attributed to the inactivation of the p53 [1–3] and pRb family of tumor suppressors [4], respectively. The viral oncogenes have been extensively studied over the past three decades mainly using transfection models and recombinant retroviruses to express them in established and primary keratinocytes.
However, immortalization and transformation are not the default outcome of an HPV infection. Instead, oncogene expression is tightly regulated in a natural infection. Our understanding of this regulation is very limited. The lack of knowledge is partly due to the fact that the HPV life cycle is strictly dependent on the terminal differentiation process of keratinocytes making the studies technically difficult. Our current view is that HPV gains access to stem and post stem cells of the basal layer through (micro)lesions by preferentially binding to the basement membrane (BM) [5]. After reaching the nucleus, it is assumed that viral genome is initially amplified. This is based on the observation that up to several hundred copies of viral genome can be found in infected basal keratinocytes [6]. After establishment of infection, the viral genome copy number is maintained in the basal compartment by maintenance replication. Viral transcription occurs at a low rate and it is assumed that the infection spreads by cell division. When HPV-harboring keratinocytes enter the terminal differentiation program, viral transcription is activated [7]. Uninfected keratinocytes exit the cell cycle at this time and commit to terminal differentiation. However E7 protein, which negates the function of the pRb family members, allows HPV-harboring cells to maintain cell cycle competence. As a consequence, E1 and E2 protein in concert with the host cell replication machinery amplify the viral genome [8]; a process that requires activation of the DNA damage response [9] and the function of the E4 and E5 viral proteins through poorly understood mechanisms [10, 11]. Inactivation of p53 by E6 protein prevents cell cycle arrest due to unscheduled DNA replication. The viral life cycle is completed following structural (late) gene expression and assembly of progeny virions in highly differentiated cells of the uppermost layers of the stratified epithelium [12].
Most of our current knowledge is based on studying HPV-harboring keratinocytes either derived from lesions or established after transfection of the viral genome. However, establishment of these cell lines requires outgrowth of immortalized keratinocytes, which in turn depends on viral oncogene expression. According to current models, immortalization is associated with increased expression of E6 and E7 [13]. Therefore, HPV-harboring cells likely display deregulated viral oncogene expression and may not be suitable for the investigation of viral early promoter regulation after infectious entry. Thus, essentially no information is available regarding the early events that regulate viral oncogene expression in an HPV-infected basal cell; despite our detailed understanding of processes leading to tumor progression. Similarly, many assumptions about establishment of infection and shift to maintenance such as genome amplification during the establishment phase lack robust experimental support. This lack of knowledge can be attributed to the fact that no cell culture model has been available to study the immediate early events of the HPV life cycle, despite more than 20 years of effort by many researchers in the field. While significant recent advances have allowed generation of virions using packaging cell lines or organotypic raft cultures [14–18], we have been unable to infect primary keratinocytes efficiently for the study of the complete viral life cycle. Even though two reports published in 2009 described efficient infection of primary keratinocytes with HPV18, the system does not seem to be robust as no follow up studies were reported [19, 20].
We have now succeeded in developing an infection model that mimics immediate early events of the HPV life cycle. The infection model is amenable to extensive genetic screens, could possibly be expanded to essentially all HPV types and allows the completion of the viral life cycle. This represents a significant technological advance that will enable the HPV and cancer research community to fill in huge gaps in our understanding of the regulation of oncogene expression and its deregulation in the early stages of tumor development. Our model will also be extremely helpful in gaining a better understanding of the HPV life cycle. It should allow a direct comparison of high- and low-risk HPV types for the first time.
Direct binding of HPV16 to primary keratinocytes yields very inefficient infection rates for unknown reasons. However, it was reported that HPV16 preferentially binds in vivo and in vitro to the basement membrane and the extracellular matrix (ECM) secreted by keratinocytes, respectively [21–23]. The interactions with ECM-resident receptors such as LN332 and heparan sulfates were shown to be sufficient to induce conformational changes in viral capsid proteins that are important for infectious entry. Mutational analyses of receptor binding sites also suggested a unique contribution of LN332 to conformational shifts in capsid proteins [24–26]. Based on these findings, we hypothesized that pre-binding of virions to ECM depositions would mimic in vivo infection and improve infection of primary keratinocytes. To test this, HaCaT cells were grown in culture dishes for 48 h and subsequently removed by treatment with EDTA. Next, HPV16 viral particles generated using the 293TT packaging cell line were added to the ECM depositions left behind on the culture dish, incubated for 2 h and followed by seeding of primary keratinocytes. With this protocol, we were able to deliver EdU-labeled pseudogenome to the nuclei of close to 50% of primary human foreskin keratinocytes (HFK) at 40 hours post infection (hpi) using ECM-to-cell transfer (Fig 1A and 1B). We observed that HPV16 E1^E4 transcripts were 10-fold higher following ECM-to-cell transfer of HPV16 virions as compared to direct binding to HFK at 72 hpi (Fig 1C). E7 and E1^E4 transcript levels were further increased up to 50-fold when HFK were left on the ECM for 7 instead of 2 days with transcripts arising mostly from the early promoter (Fig 1D). Detaching infected cells from virus-loaded ECM at day 2 and reseeding on ECM-coated dishes did not yield higher transcript levels (S1A Fig). This finding suggests that increased transcript levels over time can be attributed to the continual delivery of viral genome rather than increased promoter activity. We were also able to efficiently infect primary human tonsilar epithelial (HTE) cells using this ECM-to-cell transfer to deliver viral genome (Fig 1E). HeLa cell secretions, which lack LN332, do not support efficient HPV16 infection (S1B Fig). This is in line with previous observations, which suggested that ECM-resident LN332 plays an important role in efficient ECM-to-cell transfer [22]. HPV16 virions harboring a translation termination linker (TTL) mutation in E1 failed to establish infection since viral transcripts were hardly detectable (Fig 1F) suggesting that E1 is essential for establishment of HPV16 infection and providing indirect support for the amplification of incoming viral genome.
HPV16 early transcripts are transcribed from the early promoter p97 and are differentially spliced resulting in different quantities of viral open reading frames (ORF) (for a review see [27]). The late promoter p670 is activated when infected keratinocytes enter the terminal differentiation program. As expected when primarily the p97 early promoter is active, the most abundant transcripts contained the E6, E7 and E4 ORFs, whereas the early E1, E5 and E2 transcripts were present at significantly lower levels (Fig 2A). The late L1 and L2 ORFs were essentially undetectable at two days post infection (dpi) of HFK and just barely reached our limit of detection at 7 dpi suggesting that the late promoter is under tight control in infected HFK. Similar results were obtained with HPV16-infected HTE (S2 Fig). When we compared viral transcript levels between HPV16-infected and -immortalized HFK, we found that most early transcripts were present at 2- to 4-fold lower levels in HPV16-infected HFK (Fig 2B), with the exception of E1 encoding transcripts for which we found similar levels. The transcripts containing the late L1 and L2 ORFs as well as the E5 ORF were found at up to 20-fold lower levels in HPV16-infected compared to -immortalized HFK. The data imply a very tight control of the late promoter after HPV infection.
We profiled RNA derived from HPV16-infected HFK at 2, 4 and 7 dpi using next generation sequencing (NGS) and compared the outcome to RNA isolated from HPV16-immortalized HFK. The overall profile of the viral transcripts isolated from HPV16-infected and -immortalized HFK is very similar despite differences in read depths, providing further support for the validity of the infection model (Fig 2C and 2D). Two major splicing events use the 226 and the 409 (E6*I) and the 880 and 3358 (E1^E4) splice acceptor and donor sites, respectively. Approximately 40 to 45% of all early transcripts are spliced at the 226/409, 40 to 43% use the 880/3358 splice donor and acceptor pair. Additional previously described junctions 226/526 (E6*II; 2.5–3.1%), 226/3358 (3–5.8%), 880/2709 (E2; 3.1–3.8%), and 880/3391 (2.4–3.1%), were also found at lower frequency (Fig 2D and 2E; S1 Table). In addition to the splice acceptor site at 3358, an alternative site at 3361 is being used at low frequency. The splice variant with E8^E2 coding potential (1302/3358) is the only one, whose relative levels increase significantly over time post infection compared to other early transcripts (S1 Table) suggesting that it may be important for a switch to maintenance replication and offering support for previous reports suggesting a repressive role for E8^E2 [28–31]. Less than 5% of the early transcripts have coding potential for full-length E6. The NGS results also confirm the low abundance of E1 and E2 encoding RNAs (Fig 2A). Some minor splice variants previously reported in HPV16-immortalized cells and confirmed by our analysis were not present in HPV16-infected cells (S1 Table).
To test whether the incoming viral HPV16 genome is responsive to differentiation, we subjected HFK infected for 5 days with HPV16 virions to growth in semi-solid methylcellulose (MC) media, which is well established to induce differentiation of keratinocytes and to activate the viral late promoter [32]. Differentiation was confirmed by increased expression of differentiation markers loricrin and keratin 10 by RT-qPCR (Fig 3A) and by Western blot (Fig 3B), respectively. Activation of the late promoter was observed by RT-qPCR and confirmed by NGS giving rise to late L1- and L2-encoding transcripts (Fig 3C–3E). In addition, the early promoter was activated as evidenced by a 7-fold increase of early transcripts (Fig 3C). This was seen when HFK were grown in the presence and absence of the ROCK inhibitor. We would like to point out that the E1^E4 transcript measured in Fig 3C can arise from both the early and late promoter. In contrast, growth of HPV16-immortalized HFK in MC activated the late but only weakly the early promoter (Fig 3F). Southern blot analysis of viral genome also suggested increased viral genome levels after growth of HPV16-infected HFK in MC (Fig 3G). These data indicate that the viral genome delivered by HPV16 particles establishes infection and responds to differentiation. Furthermore, our data suggest that not only the late but also the early promoter responds to differentiation, thus providing the first experimental evidence of what has been previously implied, albeit indirectly, by RNA in situ hybridization of naturally infected lesions [33, 34].
We next subjected HFK infected for 5–7 days with HPV16 to organotypic raft cultures, which have previously been shown to support completion of the viral life cycle [18]. Uninfected and HPV16-immortalized HFK served as negative and positive controls, respectively. As shown in Fig 4A, both early and late transcripts were detectable in rafts and the expression profile of viral RNA isolated from rafts derived from infected and immortalized HFK were similar, albeit total viral RNA levels tended to be lower in rafts from HPV16-infected cells. We also observed that HPV16 genome was retained in the raft cultures, thereby suggesting replication of viral genome has occurred (Fig 4B). Indeed, HPV16-specific fluorescent in situ hybridization (FISH) identified cells with replication foci in rafts derived from both HPV16-immortalized and -infected HFK (Fig 4C). Immunofluorescent staining for E1^E4 and L1 protein were positive in many cells of the upper layers of the raft tissues (Fig 4D and 4E). Furthermore, markers of cell proliferation such as MCM7 and PCNA were present throughout the parabasal and spinous layers of the stratified epithelia and p53 signal was greatly diminished in HPV16-infected but not mock-infected cells (Fig 5). These results confirm our previous observation that most cells had been infected. Taken together, amplification of the viral genome and the presence of L1 protein suggest that the ECM-to-cell transfer infection model allows recapitulation of the complete viral life cycle.
As proof of principle that the infection model is amenable to genetic analyses, we generated HPV16 mutant viruses harboring translation termination linkers in the E6 and E7 open reading frames. Both mutant viruses established infection as evidenced by the presence of early transcripts (Fig 6A). We subjected extracts derived from HFK infected with respective wild type (wt) and mutant virus at 7 dpi to western blot analysis and a commercially available test for detection of E7 and E6 protein, respectively. E6 and E7 proteins were detected in HFK infected with wt HPV16 but were absent after infection with the respective mutant virus (Fig 6B and 6C). We conclude that expression of E6 is not impaired by E7 knockout and vice versa.
We also subjected HFK infected with mutant and wt HPV16 to long-term culturing to monitor cell survival, viral transcript, and genome levels. We observed almost complete loss of viral transcripts within 27–33 dpi with E6-TTL mutant virus (Fig 6D). This was accompanied by a loss of viral genome (Fig 6E). In contrast, HFKs infected with the E7-TTL mutant retained high levels of viral transcripts (Fig 6D). To test whether viral genomes were maintained as episomes, we developed an assay determining the resistance of HPV16 genome to exonuclease 5. Intact double-stranded circular DNA is not a substrate for this enzyme. DNA was isolated from HFK infected with wt, E6-, and E7-TTL mutant virus at 29–33 dpi, treated with exonuclease 5 and subjected to qPCR. 18S ribosomal DNA was completely digested in all samples indeed confirming that the nuclease treatment was sufficient for removal of linear DNA (Fig 6F). In contrast, mitochondrial DNA was mostly resistant as expected for a circular DNA molecule. We found that HPV DNA isolated from cells infected with wt and E7-TTL mutant virus was mostly resistant confirming that they are not substrates for exonucleases and thus likely present as circular DNA. In contrast, the low levels of viral genome still present at late times post infection with E6-TTL mutant virus was sensitive to exonuclease indicating that the remaining viral genome was either integrated or compromised otherwise. Upon long term culturing, HFK infected with E6- and E7-TTL mutant virus as well as mock-infected HFK started to senesce approximately 25 to 35 dpi, when they reached the end of their life span. The exact timing varied between different HFK lots used. Wt HPV16-infected HFK, however, continued to grow and express high levels of viral transcripts (Fig 6D). We have cultured these cells for additional 50 days without any sign of senescence, suggesting that they are immortalized. Taken together, these results suggest that neither E6 nor E7 are essential for establishing infection. However, E6 protein is essential for episomal genome maintenance, whereas loss of E7 protein does neither impair genome maintenance nor the viral transcription program in the maintenance stage of infection. However, E7 is absolutely necessary for immortalization of primary HFK under our conditions.
Herein, we describe a novel cell culture system that allows the study of the complete HPV16 life cycle following infectious delivery. Rather than binding virus directly to the cell surface, which has been documented to restrict uptake by primary keratinocytes for unknown reasons [35, 36], we used an ECM-to-cell transfer for infection of primary cells. This approach resulted in efficient uptake of viral genome by the majority of cells. Throughout the development of this infection model we used primary cells grown in the presence or absence of the Rho kinase inhibitor Y-27632 and found no significant difference in infection efficiency. Y-27632 has previously been shown to promote immortalization of primary keratinocytes [37, 38]. Taken together, this suggests that immortalization and/or the use of Y-27632 is not essential for increased infection rates. The model mimics natural infection in that (i) it utilizes pre-binding of virions to the basement membrane equivalent; (ii) only the early but not the late promoter is active in undifferentiated HFK; (iii) early and late promoter are responsive to differentiation triggered by growth in methylcellulose or organotypic raft cultures; (iv) viral genome remains episomal and is amplified upon differentiation; and (v) capsid proteins are expressed in the upper layers of organotypic rafts. At this time, we can only speculate why ECM-to-cell transfer is superior for infecting primary keratinocytes over direct binding to the cell surface. Since our data taken together with previously published observations suggest that the presence of the ECM component LN332 is important for efficient infection [22, 23], we assume that the interaction of the HPV16 capsid with LN332 induces unique conformational changes possibly allowing for direct transfer to the cellular uptake receptor. Indirect evidence for unique contributions of LN332 to HPV16 infection has been presented before using heparan sulfate binding-deficient mutants, which were shown to be non-infectious when bound to the cell surface but fully infectious when pre-bound to ECM in the absence of heparan sulfate moieties [26]. Furthermore, HFK are polarized and uptake via the basolateral surface may be more efficient than uptake by the apical surface. Similar observations have been made for other epitheliotropic viruses [39].
Most models of the HPV life cycle assume that incoming viral genome is amplified, which is followed by subsequent genome maintenance and low transcriptional activity of viral promoters after infections have been established in the basal cell compartment. They also depict early promoter activation when infected keratinocytes enter terminal differentiation, in addition to the well-studied late promoter activation. It is unclear whether genome amplification requires an initial boost of transcription and whether the shift to genome maintenance is accompanied by early promoter repression. Also, no robust experimental data exist in support of viral genome amplification following infectious delivery. The current cell culture models using immortalized cells do not allow studying the temporal regulation of viral promoters during the immediate early stages of the viral life cycle. In addition, the early promoter is only weakly upregulated upon differentiation. We now find that the p97 early promoter strongly responds to differentiation, which, in turn, suggests that the early promoter is repressed in the basal cells. We also found that the splice variant encoding for E8^E2 is the only early transcript whose relative levels increase over time post infection of monolayer cells. E8^E2 is a potent inhibitor of viral replication and transcription and has been shown to restrict viral genome copy numbers in HPV-harboring immortalized cells [28–30]. E8^E2 is transcribed from a recently identified promoter located in the E1 ORF [40]. The E8 promoter has not been studied in great detail, notably, knowledge about its temporal regulation post infectious delivery of viral genome is completely lacking. The infection model will provide a potent platform to study the temporal regulation of the E8 promoter following infectious delivery of viral genome. It is tempting to speculate that its regulation may allow the E8^E2 repressor to orchestrate the shift from establishment of infection, which has been suggested to involve a boost of viral transcription and genome amplification [30], to maintenance transcription and replication.
Despite extensive studies regarding the functions of early viral proteins in immortalization, transformation and transcriptional regulation, we still know very little about their roles during the viral life cycle; owing mainly to our inability to establish cell lines carrying mutations in many viral genes. We generate HPV16 virions in the HEK 293TT cell line, which does not require HPV factors other than the capsid proteins expressed from a heterologous expression vector. Therefore, the system is amenable to extensive mutational manipulation. As a proof of principle that the infection model will allow investigation of the contributions of individual viral proteins to the complete viral life cycle, we tested E1-, E6-, and E7-TTL mutant viruses for their ability to establish infection and retain episomal genome. As expected, the E1-TTL mutant was unable to efficiently establish infection. Viral transcripts are present, however, at levels 1% below that of wt HPV16 at 6 dpi. In turn, this indirectly suggests that viral genome is amplified following infectious entry. However, it is also conceivable that replication is essential for efficient transcription and further experimentation is required to clarify this point. In contrast, E6- and E7-TTL mutant virus established infection, suggesting they are not essential for immediate early events of the viral life cycle. However, viral transcript levels were consistently lower after infection with E6- compared to E7-TTL mutant and wt virus. Analysis of infected cells at subsequent passages suggests that E6-TTL failed to retain episomal viral genome and viral transcripts were not detectable. Published data using mutants of HPV16 and HPV31 are somewhat conflicting. For HPV31, it was shown that both E6 and E7 were required to establish stably transfected cell lines containing episomal viral genome. In contrast, HPV16 genome harboring E7 mutations were episomally maintained in immortalized NIKS keratinocytes [41, 42]. It is interesting to note that previously described E7-mediated changes to the host cell transcriptome, many of which involve S phase genes (for recent reviews see [43, 44]), do not seem to be essential for genome maintenance, as the cells infected with E7-TTL mutant virus retain episomal genomes until they senesce. However, we have not yet compared the host transcripts from cells infected with wild type and E7-TTL mutant virus to formally show which alterations to the transcriptome are seen in wild type-infected cells and which of these are due to E7 expression.
The infection model will provide a unique platform to identify host cell factors transcriptionally regulated by the viral oncoproteins after infectious delivery of viral genome without the requirement for immortalization. Analyses of transcripts isolated from individual layers of the stratified epithelia obtained after growth of infected and immortalized HFK as organotypic raft cultures may provide important clues regarding the involvement of altered pathways in the viral life cycle. In future studies, it should be possible to link alterations of the transcriptome to specific functions of the oncoproteins by using mutant viruses. While many of the biological functions and interacting partners of E6 and E7 are identical between low- and high-risk HPV types, it is still not clear, which activities of the high-risk HPV types are ultimately responsible for immortalization. The infection model should be extendable to the study of low-risk HPV types such as HPV6 and 11, which cannot be studied with the current cell culture systems due to their inability to immortalize keratinocytes. A comparative analysis combined with a genetic approach should identify activities absolutely essential for completion of the viral life cycle of both virus groups and may in turn identify functions mediating immortalization. The low-risk HPV types are known not only to cause genital warts but also recurrent respiratory papillomatosis, a debilitating disease requiring repeated surgical procedures, for which no treatment other than surgery is currently available [45]. The extension of the herein described infection model to low-risk HPV types will provide the first platform to investigate and test potential drug candidates for treatment. The infection model may also allow the investigation of skin cancer-linked HPV types from the β-genus and their cooperation with UV irradiation, including the proposed hit and run mechanism of carcinogenesis [46]. Overall, the establishment of this infection model will provide a new experimental tool for the study of the HPV life cycle and will help further our understanding of the biological processes leading to immortalization. Furthermore, it will be helpful for the emerging field of studying the synergy of different pathogens in the development of tumors such as oropharyngeal squamous cell carcinoma [47, 48].
Human embryonic kidney 293TT and HeLa cells were obtained from John Schiller and Daniel DiMaio, respectively [15, 49]. They were cultured in DMEM supplemented with 10% FBS, non-essential amino acids, antibiotics, and L-Glutamax. Spontaneously immortalized human keratinocytes HaCaT cells were purchased from the American Type Culture Collection (ATCC) and grown in low glucose DMEM containing 5% FBS and antibiotics. Human foreskin keratinocytes (HFKs) were derived from neonatal human foreskin epithelia and maintained in E medium containing mouse epidermal growth factor (EGF) and mitomycin-treated mouse 3T3 J2 fibroblasts as previously described [50]. Pooled primary epithelial keratinocytes were also purchased from the ATCC (PCS-200-010) and used in some experiments. In early experiments where indicated, we maintained and infected primary keratinocytes in the presence of the Rho kinase inhibitor (ROCK) Y-27632, which was reported to increase their lifespan [37, 38]. However, the ROCK inhibitor was excluded prior and during experiments involving long term culturing of infected primary cells. Stable cell lines containing HPV16 episomes were created by co-transfection of pEGFP-N1-HPV16 containing the HPV16 genome (W12 strain) with an expression vector for Cre recombinase and a Neomycin resistance plasmid. Cells were transfected using polyethyleneimine (PEI; Polysciences), selected with G418, and expanded as previously described [51]. Episomal maintenance of the viral genome was confirmed using Southern blotting [51]. Differentiation was induced by suspending cells in 1.5% methylcellulose (MC) for 24 hours followed by washes in phosphate buffered saline [32]. Human primary tonsil cells were isolated from tonsils and maintained in E medium with mitomycin-treated mouse 3T3 J2 fibroblasts. Before harvesting RNA or DNA, fibroblast feeders were removed by short trypsin treatment, followed by two washes in PBS.
Foreskin and tonsillar keratinocytes were collected from discarded tissue following routine circumcisions and tonsillectomy from anonymous donors attending University Health, Shreveport. Because the samples were de-identified, would otherwise have been discarded, and were not collected specifically for our studies, the LSUHSC-S IRB ruled that they fell under the NIH’s definition of “exempt” from human subjects research, including informed consent (Institutional IRB approval number: STUDY00000187).
The pSheLL16 L1/L2 packaging plasmid and pfwB plasmid, expressing enhanced green fluorescent protein (GFP) were a kind gift from John Schiller, Bethesda, MA. The plasmid pEGFP-N1 containing the entire floxed HPV16 genome (pEGFP-N1-HPV16) and pBCre plasmid have been described previously [52]. Quasivirions were generated using 293TT cells following the improved protocol of Buck and Thompson [16] with minor modifications. Briefly, 293TT cells were first cotransfected with the pSheLL16 L1/L2 and pEGFP-N1-HPV16 plasmids and 24 hours later transfected with the pBCre plasmid. An additional two days later, cells were harvested and viral particles were purified as described previously [16, 53]. Because activity of the Cre recombinase generates two circular plasmids of packable size (pEGFPN1 and HPV16 genome), isolated viral particles comprise a mixture of pseudovirions (pEGFPN1 plasmid) and quasivirions (HPV16 genome). Pseudovirions harboring GFP were also generated in 293TT cells as described by Buck et al. [16, 53]. For pseudogenome detection by fluorescence microscopy, pseudogenomes were labeled with EdU (5-ethynyl-2’-deoxyuridine) by supplementing the growth medium with 100 μM EdU at 6 hours post transfection as described [54] during generation of pseudovirions. The viral genome equivalence (vge) was determined by real-time quantitative PCR (RT-qPCR) of encapsidated DNA isolated using the NucleoSpin Blood QuickPure (Macherey-Nagel; 740569.250).
To introduce a TTL into the E1 ORF, the pEGFP-N1-HPV16 plasmid was digested with ApaI for 1 h at 25 C. The subsequent ~4500 bp fragment was excised from the gel, purified using DNA gel clean-up kit (Macherey-Nagel, 740609.50), and subcloned into pBlueScript KS II. Next, we used site-directed mutagenesis to substitute a single nucleotide at position 892 within the E1 ORF, which introduced an in-frame TAA ‘stop’ codon just downstream of the E1 start codon (as previously described by [55]). Once confirming the substitution by DNA sequencing, we re-digested the plasmids with ApaI, excised and gel-purified the mutated fragment and vector, and re-ligated it back into the original pEGFP-N1-HPV16 plasmid. The correct insert was again confirmed by sequencing. Primers used: Forward 5'- CCA TGG CTG ATC CTG CAG GTA CCA ATG GGT AAG AGG GTA CGG GAT GTA ATG G -3', Reverse 5'- CCA TTA CAT CCC GTA CCC TCT TAC CCA TTG GTA CCT GCA GGA TCA GCC ATG G -3'. The E7-TTL mutant has been described previously [51]. Site directed mutagenesis to create TTL mutations in the E6 open reading frame of pEGFP-N1-HPV16 was performed using the QuickChange II Site Directed Mutagenesis kit (Agilent) using primers 5’-GCAATGTTTCAGGACCCATAGTAGTGACCCAGAAAGTTAC-3’ and 5’-GTAACTTTCTGGGTCACTACTATGGGTCCTGAAACATTGC-3’ and confirmed by sequencing.
HaCaT cells were seeded in 60 mm cell culture dishes and grown for 24-48h until they reached confluency to allow secretion of ECM. Cells were incubated in Dulbecco’s PBS supplemented with 0.5 mM EDTA for up to 2 h in order to remove the cells. To prevent outgrowth of residual HaCaT cells, the dish surface was treated with 8 μg/ml mitomycin for 4 h. Optiprep-purified viral particles (>5 x 107 vge/dish) diluted in 2 ml E medium were added to the ECM for at least 2h at 37°C. At this time, 5 x 105 low passage primary keratinocytes were added. Two hours later approximately 1 x 105 mitomycin-treated fibroblast feeder cells were added in addition. When different sized culture dishes (ranging from 12 well plates to 100 mm dishes) were used, cell and vge numbers were scaled proportionally to the surface area. Infection was continued for up to 7 days or until cells reached confluency.
In order to induce differentiation of HFK cells, cells were suspended in methylcellulose at 5 to 7 days post infection with HPV16 quasivirions as described [56]. Samples were collected 24 or 48 hours later. Increased levels of differentiation markers were confirmed by Western Blot and RT qPCR.
Organotypic raft cultures generated from HFK cells infected for 5 to 7 days with HPV16 quasivirions were grown as described [50, 57]. Briefly one million keratinocytes were seeded onto the surface of the collagen gel containing fibroblasts feeders. Following attachment, the gel with keratinocytes layer was lifted and placed onto a stainless steel grid in a culture dish. Culture medium was added to the dish so that the keratinocyte/collagen plug was exposed to the air from above and to the medium from below. The medium was changed every other day maintaining the air-liquid interface. Rafts were grown for 14 days and samples were collected for RNA/DNA analysis and immunofluorescent staining and FISH. Rafts generated from uninfected HFK seeded on ECM were used as control.
HFK cells were infected with EdU-labeled pseudovirions using ECM-to-cell transfer on glass slides. EdU staining was performed according to the manufacturer’s directions. In brief, at the indicated times post infection, cells were washed with PBS and fixed with 4% paraformaldehyde for 15 min at room temperature, washed, permeabilized with 0.5% Triton X-100 in PBS for 10 min, washed, and blocked with 5% goat serum in PBS for 30 min followed by a 30 min incubation with Click-iT reaction cocktail containing AlexaFluor 555 for EdU-labeled pesudogenome detection. After extensive washing, cells were incubated for 30 min with anti-PML (BETHYL; A301-167A), and anti-laminin A/C (Sigma; SAB4200236) primary antibodies at room temperature, washed again extensively, and subsequently incubated with AlexaFluor488- and AlexaFluor647-tagged secondary antibodies (Molecular Probes; A11029, A21245) for 30 min. After extensive washing with PBS, cells were mounted in ‘Gold Antifade’ containing DAPI (Life Technologies; P3693).
HPV16 genomic DNA probes for FISH were prepared by gel purification of the entire HPV16 genome from pUC-HPV16 digested with BamHI and generated using BioNick labeling system according to the manufacturer’s protocol (Invitrogen, 18247–015). When mentioned, raft sections were stained for the presence of viral proteins prior to in situ hybridization. Paraffin wax embedded sections were dewaxed in series of xylene and alcohol washes, followed by antigen retrieval using microwave heating at 100°C in citrate buffer with 0.05% Tween for 20 minutes. Slides were permeablized with 0.5%Triton x100 for 45 minutes and blocked with 5% goat serum for 1h. Primary antibodies: anti-L1-7, anti-E1^E4 (a kind gift from J. Doorbar), anti-PCNA (Santa Cruz Biotechnology; sc-7907), anti-p53 (Santa Cruz Biotechnology; sc-126) or anti-MCM7 (Abcam; ab52489) were added for overnight incubation at 4°C. After extensive washing with PBS, sections were incubated for 1h with AlexaFluor-tagged secondary antibodies (Molecular Probes; A21236; A21245; A11030; A11035) for 1 hour. After extensive PBS washing, sections were fixed and slides were treated with 100 ug/ml RNase A in 2× SSC for 1 hour at 37°C and for 5 min with micrococcal nuclease (NEB; M0247S). Enzymatic activity was blocked by adding 20mM EGTA for 5 min. Subsequently, the slides were washed three times with 2× SSC, then dehydrated for 2 min each in 70% EtOH, 80% EtOH and 100% EtOH at room temperature. Slides were then denatured in 70% formamide-2× SSC at 76°C for 3 minutes, followed by dehydration for 2 min each in 70% EtOH (−20°C), 80% EtOH and 5 min in 100% EtOH at room temperature. The probe was denatured at 74°C for 10 minutes prior to hybridization overnight at 37°C. After overnight incubation, the slides were washed multiple times, and tyramide-enhanced fluorescence was carried out according to the manufacturer's instructions (Molecular Probes, T20932). After extensive final washing with PBS, cells were mounted in ‘Gold Antifade’ containing DAPI (Life Technologies; P3693). All IF images were captured by using a Leica CTR6000 fluorescence microscope or by confocal microscopy with a 63x objective using a Leica TCS SP5 Spectral Confocal Microscope and processed with Adobe Photoshop software.
Total RNA from HFK cells was extracted using the RNeasy Plus Mini RNA Isolation Kit (Qiagen; 74236). RNA samples from raft cultures were extracted using RNA Stat-60 (amsbio LLC) according to manufacturer’s protocol. Isolated RNA samples were treated with DNase I (NEB; M0303L) prior to reverse transcription. 1 or 0.5 μg total RNA was used to reverse-transcribe into cDNA using ImProm-II Reverse Transcriptase kit (Promega). Equal amounts of cDNA were quantified by RT-qPCR using the IQ SYBR Green Supermix (BIO-RAD) and a CFX96 Real-Time System (BIO-RAD). PCR reactions were carried out in triplicate, and transcript levels were normalized to cyclophilin A. Mock reverse-transcribed samples were included as negative control. A list of oligonucleotide sequences used is provided in S2 Table. The BIO-RAD CFX Manager 3.1 software was used to analyze the data.
Total RNA was harvested as described above. RNA quality was assessed on an Agilent Tapestation Bioanalyzer. All samples showed an RNA Integrity Number (RIN) greater than 7. An mRNA sequencing library was prepared with the NEBNextUltra directional library kit and the TruSeq stranded mRNA kit (Illumina). Paired end sequencing (2x75cycles) was performed on an Illumina NextSeq 500 obtaining over 25 million reads per sample. Reads were aligned to the HPV16 (NC_001526.3) genome using STAR_2.4.2a and counted using RSEM 1.2.31.
HFK cells were infected with HPV16 quasivirions using ECM to cell transfer. Uninfected cells served as a control. Genomic DNAs (gDNAs) were isolated from the cells cultured in monolayer for 4 days or cultured in monolayer for 4 days followed by 48h in methylcellulose. Cell pellets were resuspended in lysis buffer (400 mM NaCl, 10 mM Tris-HCl [pH 7.4], and 10 mM EDTA); then, RNase A (50 μg/ml), proteinase K (50 μg/ml) and 0.2% SDS were added, and the lysates were incubated overnight at 37°C. DNA was extracted with phenol-chloroform and precipitated with ethanol. Approximately 5 μg of gDNA was digested with BglII (which does not cut the HPV16 genome) and resolved on a 0.8% agarose gel. Genomic DNA fragments were transferred from the gel to DuPont GeneScreenPlus nylon membrane (NEN Research Products, Boston, Mass.) as described by the manufacturer using alkaline transfer. Prehybridization of the membrane was performed for 1h at 42°C using a solution containing 50% formamide, 4× SSC, 5× Denhardt's solution, 1% SDS, 10% dextran sulfate, and denatured salmon sperm DNA (0.1 mg/ml). The HPV16 probe was prepared by gel purification of the entire HPV16 genome from pUC HPV16 digested with BamHI and labeling with the Ready-To-Go DNA labeling kit (Amersham Pharmacia). Labeled probe was then purified with ProbeQuant G-50 Micro columns (Amersham Pharmacia), denatured, and added to fresh hybridization solution, which was incubated with membrane at 42°C overnight. Membrane was washed twice with 2× SSC-0.1% SDS for 15 min at room temperature, twice with 0.5× SSC-0.1% SDS for 15 min at room temperature, twice with 0.1× SSC-0.1% SDS for 15 min at room temperature, and once with 0.1× SSC-1% SDS for 30 min at 50°C. Hybridizing species were visualized by autoradiography.
Whole-cell extracts were obtained from cell pellets lysed in 1x Laemmli Sample Buffer (BIO-RAD) supplemented with 2-mercaptoethanol. Proteins were resolved on SDS-PAGE and transferred to nitrocellulose membranes (BIO-RAD). Membranes were blocked 1 hour in 5% Blotting-Grade Blocker (BIO-RAD) in 1× TBST) and incubated at 4°C overnight with anti-cytokeratin 10 (Santa Cruz Biotechnology; sc-52318), anti-E7 (Santa Cruz Biotechnology, sc-6981) or anti-b-actin (Santa Cruz Biotechnology; sc-47778) primary antibodies. After incubation, membranes were washed 3 × 15 minutes in 1× TBST wash buffer. Membranes were then incubated with horseradish peroxidase–tagged goat anti-mouse and goat anti-rabbit secondary antibodies (1:2500, Jackson ImmunoResearch) at room temperature for 1 hour, washed 3 × 15 minutes in 1× TBST. Signals were detected by enhanced chemiluminescence (Thermo Scientific). Equal protein loading was confirmed by probing with β-actin monoclonal antibody.
The presence of E6 protein in infected cells was detected using a kit from ArborVita according to the manufacturer’s protocol. Briefly, the cell lysate was incubated with alkaline phosphatase conjugated high-affinity E6 HPV16/18 monoclonal antibodies. Next, a nitrocellulose test strip with two capture lines consisting of immobilized mAbs to HPV16/18 E6 was placed into the lysate/mAb-AP mix. The solution was allowed to migrate through the strip by capillary action. E6-mAb-AP present in the sample is forming a ternary complex with the immobilized antibodies on the strip. The complex was visualized as a purple line in the respective location on the strip by the addition of an enzyme substrate solution provided in the kit.
Genomic DNA was isolated using the QIAamp DNA Blood Mini Kit (Qiagen) according to the manufacturer’s instructions and stored at 4°C. DNA from UMSCC47 and HPV16-infected 293TT cells served as an HPV16 integration control and episomal HPV16 control, respectively. 100 ng of DNA was either treated with exonuclease V (RecBCD, NEB) or left untreated for 1 hour at 37°C followed by heat inactivation at 95°C for 10 minutes. 10 ng of digested/undigested DNA was then quantified by real time PCR using a 7500 FAST Applied Biosystems thermocycler with SYBR Green PCR Master Mix (Applied Biosystems) and 300nM of each primer in a 15 μl reaction. Nuclease free water was used in place of the template for a negative control. The following cycling conditions were used: 50°C for 2 minutes, 95°C for 10 minutes, 40 cycles at 95°C for 15 seconds, and a dissociation stage of 95°C for 15 seconds, 60°C for 1 minute, 95°C for 15 seconds, and 60°C for 15 seconds. Separate PCR reactions were performed to amplify HPV16 E6 (1) (F: 5’-GAGAACTGCAATGTTTCAGGACC-3’ R: 5’-TGTATAGTTGTTTGCAGCTCTGTGC-3’), human mitochondrial DNA (2) (F: 5’-CAGGAGTAGGAGAGAGGGAGGTAAG-3’ R: 5’-TACCCATCATAATCGGAGGCTTTGG -3’), and human 18S_ribosomal DNA (3) (F: 5’-GCAATTATTCCCCATGAACG-3’ R: 5’-GGGACTTAATCAACGCAAGC-3’). Human mitochondrial DNA and 18S ribosomal DNA served as episomal and multi-copy linear DNA internal controls, respectively. Primer efficiencies were based on a standard curve generated using a 5-fold dilution series of undigested UMSCC47 DNA and used to calculate the relative amount of DNA per sample. The percent of DNA resistant to exonuclease digestion was calculated relative to undigested DNA.
|
10.1371/journal.pmed.1002473 | Shortages of benzathine penicillin for prevention of mother-to-child transmission of syphilis: An evaluation from multi-country surveys and stakeholder interviews | Benzathine penicillin G (BPG) is the only recommended treatment to prevent mother-to-child transmission of syphilis. Due to recent reports of country-level shortages of BPG, an evaluation was undertaken to quantify countries that have experienced shortages in the past 2 years and to describe factors contributing to these shortages.
Country-level data about BPG shortages were collected using 3 survey approaches. First, a survey designed by the WHO Department of Reproductive Health and Research was distributed to 41 countries and territories in the Americas and 41 more in Africa. Second, WHO conducted an email survey of 28 US Centers for Disease Control and Prevention country directors. An additional 13 countries were in contact with WHO for related congenital syphilis prevention activities and also reported on BPG shortages. Third, the Clinton Health Access Initiative (CHAI) collected data from 14 countries (where it has active operations) to understand the extent of stock-outs, in-country purchasing, usage behavior, and breadth of available purchasing options to identify stock-outs worldwide. CHAI also conducted in-person interviews in the same 14 countries to understand the extent of stock-outs, in-country purchasing and usage behavior, and available purchasing options. CHAI also completed a desk review of 10 additional high-income countries, which were also included. BPG shortages were attributable to shortfalls in supply, demand, and procurement in the countries assessed. This assessment should not be considered globally representative as countries not surveyed may also have experienced BPG shortages. Country contacts may not have been aware of BPG shortages when surveyed or may have underreported medication substitutions due to desirability bias. Funding for the purchase of BPG by countries was not evaluated. In all, 114 countries and territories were approached to provide information on BPG shortages occurring during 2014–2016. Of unique countries and territories, 95 (83%) responded or had information evaluable from public records. Of these 95 countries and territories, 39 (41%) reported a BPG shortage, and 56 (59%) reported no BPG shortage; 10 (12%) countries with and without BPG shortages reported use of antibiotic alternatives to BPG for treatment of maternal syphilis. Market exits, inflexible production cycles, and minimum order quantities affect BPG supply. On the demand side, inaccurate forecasts and sole sourcing lead to under-procurement. Clinicians may also incorrectly prescribe BPG substitutes due to misperceptions of quality or of the likelihood of adverse outcomes.
Targets for improvement include drug forecasting and procurement, and addressing provider reluctance to use BPG. Opportunities to improve global supply, demand, and use of BPG should be prioritized alongside congenital syphilis elimination efforts.
| A single dose of low-cost benzathine penicillin G (BPG) ends syphilis infectivity in adults with no documented risk of antibiotic resistance. In spite of this, syphilis continues to infect millions globally.
Pregnant women with syphilis are particularly vulnerable, as fetal transmission of the infection can cause stillbirth. The only recommended treatment for syphilis in pregnant women is BPG.
Congenital syphilis remains a significant contributor to early infant mortality, particularly in low- and middle-income countries.
There are several reasons for this, but one of the most important is a global shortage of BPG.
In 2015, WHO began to receive anecdotal country reports of BPG stock-outs.
WHO decided to assess these shortages, describe global supply and demand drivers, and propose viable policy solutions.
The team completed 3 surveys to assess global BPG shortages occurring during 2014–2016.
The first was distributed to 41 countries and territories in the Americas and 41 African countries.
The second was emailed to 28 US Centers for Disease Control and Prevention country directors.
The third was used in in-person interviews in 14 countries by the Clinton Health Access Initiative.
In all, 95 of 114 unique countries and territories responded to the surveys. Of these, 39 reported a BPG shortage and 56 reported no BPG shortage.
The team discovered 3 major issues. First, countries often obtain their product from a single wholesaler, which often obtains its products from a single final dose formulator, which often obtains its active ingredient from a single manufacturer. This “sole sourcing,” combined with a highly consolidated market, makes alternative supply difficult if there are production, quality, regulatory, or specification changes within a country’s supply chain.
Second, as an off-patent medication, BPG commands a market price of pennies per dose. However, as a sterile injectable, it is also expensive to manufacture. These economics have led manufacturers to either abandon BPG production or implement stringent ordering protocols that compromise supply for low- and middle-income countries.
Third, inaccurate country forecasts, weak procurement systems, and clinical knowledge gaps about syphilis treatment have compromised demand for and procurement of BPG.
Widespread BPG shortages compromise treatment of adult syphilis, prevention of congenital syphilis, and treatment of other BPG-indicated conditions (including rheumatic heart disease).
An uninterrupted supply of quality-assured active pharmaceutical ingredient and final formulated product will simplify BPG procurement for high-burden countries.
Countries must strengthen their supply chain, purchasing, forecasting, and procurement infrastructure to mitigate shortage risk and reduce demand-side stock-outs.
Countries must also strengthen testing for and treatment of maternal syphilis (and prevention of congenital syphilis) with BPG.
| The World Health Organization (WHO) estimates that there are 5.6 million new cases of syphilis annually and 18 million prevalent cases [1]. In 2012, WHO estimated that there were 930,000 pregnant women with syphilis, resulting in 350,000 adverse pregnancy outcomes, with over half of these being stillbirth or neonatal death [2]. Both screening and treatment for syphilis remain suboptimal in low- and middle-income countries (LMICs) [3], despite the diagnosis and prevention of mother-to-child transmission (MTCT) of syphilis being feasible, inexpensive, and cost-effective [4]. Benzathine penicillin G (BPG) is the only recommended treatment for syphilis in pregnant women to prevent MTCT, as other drugs are contraindicated, do not cross the placenta to treat the fetus, or are less effective than BPG [5]. Treatment of syphilis-infected pregnant women with 2.4 million international units (IU) of intramuscular of BPG given at least 28 days prior to delivery can result in an 82% reduced risk of stillbirth and 80% reduction in neonatal mortality [6]. In recognition of this, WHO published The Global Elimination of Congenital Syphilis: Rationale and Strategy for Action [7] and set country targets for the elimination of MTCT of HIV and syphilis, which will bring BPG shortages into sharp focus.
Syphilis’s causative organism (Treponema pallidum) has not developed resistance to first-line treatment with BPG, which is also indicated for secondary prophylaxis of rheumatic heart disease, primary treatment of group A streptococcal pharyngitis, and yaws [5,8,9]. As a result, BPG is considered an essential medicine by WHO [10]; it is typically available as a powder for reconstitution or as a suspension product (marketed in the US, Canada, Australia, New Zealand, and Brazil) [11,12].
As BPG is off patent, it sells for pennies a dose, but as a sterile injectable medication, it is expensive to manufacture. Several active pharmaceutical ingredient (API) manufacturers that make BPG’s active ingredient, and final dose formulators (FDFs) that formulate, package, and label the final product, have stopped producing BPG because of these economics, which has dramatically increased the stock-out risk.
WHO is aware of this vulnerability [13] and advocated a systematic approach to manage these shortages [14]. In May 2016, BPG was recognized by the 69th World Health Assembly as an essential medicine at high risk for stock-out [15]. Despite procurement agency acknowledgment that chronic stock-outs have occurred [14,15], the supply and demand factors that contribute to these shortages have not been previously described.
In 2016, WHO partnered with the Bill & Melinda Gates Foundation and the Clinton Health Access Initiative (CHAI) to (1) identify the extent of BPG stock-outs, (2) determine key contributors to shortages, and (3) propose strategies to mitigate future shortages.
Data for this analysis were collected from 3 data collection efforts. First, a survey was designed by WHO’s Department of Reproductive Health and Research, adapted and translated from English into Spanish, and distributed in both languages by the Pan American Health Organization (PAHO)/WHO Regional Office for the Americas to the ministries of health of 41 countries and territories (henceforth “countries”) in Latin America and the Caribbean (S1 and S2 Appendices). The same survey was translated into French and used by the WHO Regional Office for Africa (AFRO) in 41 southern, eastern, and western African countries (S3–S5 Appendices). It is important to note that the regional offices were free to adapt the survey based on their needs. AFRO in particular used the survey to also get more information on sexually transmitted infection (STI) management in general, making the answers a bit different from those in the PAHO survey, although the core survey questions were preserved. We also provided the survey electronically, but several respondents to the AFRO survey, and to the US Centers for Disease and Prevention (CDC) survey described below, responded verbally.
Second, WHO conducted an informal email survey of 28 countries (selected because they were PEPFAR [US President’s Emergency Plan for AIDS Relief] countries for which a CDC director was currently assigned) to identify additional countries experiencing BPG shortages or stock-outs (S6 Appendix) (CDC director feedback was normally via email). An additional 13 countries were in contact with WHO for related congenital syphilis prevention activities and also reported on BPG shortages (S7 Appendix).
Third, CHAI developed data collection instruments for in-person interviews in 14 countries (where CHAI has active operations) with BPG distributors, purchasers, ministry of health (MoH) staff, and national leads (S8–S11 Appendices), to understand the extent of stock-outs, in-country purchasing and usage behavior, and available purchasing options.
For countries participating in multiple surveys, a result of BPG stock-out in one survey and no stock-out in another was counted as a stock-out. Countries not involved in these surveys but for whom WHO was in contact regarding issues related to maternal syphilis treatment and congenital syphilis prevention also contributed data on BPG shortages.
In all, 114 countries and territories were approached to provide information on BPG shortages occurring during 2014–2016. Of unique countries and territories, 95 (83%) responded or had information evaluable from public records. A detailed list of sources and indicators used for data collection in these 3 data collection efforts is provided in Table 1.
CHAI supplemented the 14 in-depth interviews with a search of grey literature. A systematic review was not completed; however, a snowball technique was employed. CHAI conducted a Google search using a combination of the following terms: “benzathine penicillin” with “brand,” “manufacturer,” or “supplier.” A list of benzathine penicillin names and manufacturers was extracted from the first 5 pages of Google results, which provided a starting list of potential manufacturers and brand names of benzathine penicillin. Duplicates were removed, and a Google search of each brand name with “benzathine penicillin” was conducted.
Relevant information (or lack of a finding) from manufacturer and drug regulatory agency websites was extracted into an Excel database. In addition, specific regulatory databases known to the researchers were searched for “benzathine penicillin.” To validate the hypothesis that stock-outs are a global phenomenon with impact beyond LMICs, CHAI also completed a desk review to scan databases of stock-outs in 10 high-income countries, which were also included in the analysis.
A list of the indicators and sources used in the BPG country surveys is provided in Table 2.
Supply-side stakeholders were interviewed by CHAI to understand constraints and challenges that may contribute to stock-outs, including capacity, lead times, order patterns, and pricing. CHAI conducted site visits with the 4 API manufacturers, which all operate integrated FDFs currently active in the market.
CHAI requested interviews from 12 FDFs (S12 Appendix) that were purposively selected based on general criteria including available contact information, global reach, and stock-out patterns identified through the survey activities described above. Five FDFs responded to the request for interview, and CHAI was able to conduct semi-structured interviews with all 5. Lastly, CHAI interviewed 4 global wholesalers that sell product to governments.
The CHAI data collection instrument allowed CHAI to gather country-specific demand-side information on disease prevalence, current clinical practice, product registrations, pricing, historical procurement volumes, and stock-out reports. Interviews were conducted with purchasing agents, national program leads for syphilis, clinicians, and MoH officials, depending on the country. On-the-ground staff in the 14 CHAI countries tailored the data collection methods to the specific country context to collect the requested data. Procurement, pricing, and stock-out data reports were requested and reviewed from the correct procurement or ministry officials. Desk reviews of country-specific treatment guidelines were also conducted.
These inquiries were planned and executed in a stepwise fashion, in the following order: country surveys, supplier interviews, purchasing agent interviews, and clinician and MoH interviews. Each inquiry was informed by results from the previous inquiry. CHAI investigators recorded notes of their discussions by hand.
The purpose of this assessment was to (1) review BPG availability to prevent and control disease manifestations that represent an immediate risk of untreated or inadequately treated maternal or congenital syphilis, (2) document BPG availability as a public health problem, and (3) improve public health programming. No human research was intended or conducted at the time of collection of these data, and thus this project did not undergo ethical review.
A total of 114 countries and territories were approached to provide information on BPG penicillin shortages or stock-outs occurring during 2014–2016; 24 were part of 2 or 3 of the above-mentioned surveys. After accounting for countries in multiple surveys, 95 (83%) unique countries and territories responded or had evaluable information from public records. Of these 95 countries, 39 (41%) reported a BPG shortage, and 56 (59%) reported no BPG shortage. Ten of these countries reported use of alternative treatments for maternal syphilis including ceftriaxone, amoxicillin, and erythromycin. Three of these countries reported exclusive use of antibiotic alternatives to BPG for treatment of syphilis, and 7 countries reported use of alternatives in addition to reporting BPG shortages (Fig 1; Table 1).
Of the 41 countries and territories in Latin America and the Caribbean surveyed by PAHO, 29 responded (71%). Five reported shortages of BPG (Brazil, Jamaica, Panama, Suriname, and Trinidad and Tobago). Costa Rica reported problems purchasing the 1.2 million UI dose of BPG; Chile reported problems with purchase of the 2.4 million UI dose of BPG. Chile and Nicaragua reported backlogged orders.
Of the 41 countries that received a questionnaire from WHO AFRO to assess availability of BPG, 35 responded (85%), and 12 experienced stock-outs (Benin, Congo, Guinea, Madagascar, Sierra Leone, Eritrea, South Sudan, Comoros, Mozambique, Namibia, Rwanda, and Zimbabwe).
Of the 28 CDC country directors surveyed, 20 (71%) responded, and 6 reported BPG stock-outs (South Africa, Sierra Leone, Namibia, Malawi, Ghana, and Ethiopia).
Of the 14 countries CHAI surveyed to assess availability of BPG, all responded (100%), and 10 reported stock-outs (Cambodia, Cameroon, Ethiopia, Liberia, Malawi, Nigeria, South Africa, Uganda, Zambia, and Zimbabwe).
Thirteen additional countries were in contact, either in person or by phone, with WHO regarding issues related to syphilis. These countries were questioned regarding BPG shortage. Four reported BPG shortage (Philippines, Greece, Croatia, and Switzerland), 1 reported exclusive use of alternative antibiotics for treatment of syphilis (Japan), and 8 reported no shortages.
The study identified a supply shortage of Pfizer’s BPG product in 6 high-income countries (United States, Canada, Australia, New Zealand, Netherlands, and Germany). Large procurement or wholesale agents, such as the United Nations Children’s Fund (UNICEF), were also identified by the CHAI assessment as experiencing challenges sourcing BPG for government and non-profit buyers.
The CHAI assessment demonstrated that, as an older, off-patent formulation, BPG sells at an average of US$0.11 for a 1.2 million IU dose and US$0.20 for a 2.4 million IU dose in LMICs, its largest markets. (This average is based on a CHAI analysis of prices obtained from country-specific procurement data from CHAI BPG country surveys and interviews with distributors.) Additionally, some countries (including India and Brazil) have set a maximum price for BPG, which reduces the commercial attractiveness of this product to API manufacturers.
Significant infrastructure and investment is required to produce a sterile injectable medication like BPG, which limits new market entrants and contributes to supplier exits when production is shifted to less resource-intensive and more commercially attractive products (Box 1). Six API manufacturers and more than 40 FDFs have left the BPG market since the early 2000s.
Even as higher-margin medications are prioritized in FDF production schedules, 4 of 5 FDFs interviewed noted that the time it takes to switch production lines to manufacture other pharmaceutical products can delay lead times between ordering, production, and delivery for up to 1 year.
While API manufacturers have the production capacity to meet global demand, the strategies that they (and FDFs) employ to optimize their commercial position often also constrain supply. For example, minimum purchase order quantities limit the ability of buyers (for example, smaller countries) to make a smaller BPG order. Without pooling of purchase orders across buyers, individual-country demand for BPG is often not enough to meet the minimum purchase order required by API or FDF suppliers. The complexity of the current supply situation and the knock-on effect of changes in any single component of the supply chain are illustrated in Fig 2.
API manufacturer quality is also a concern. None of the 3 API manufacturers currently has market authorization from a stringent regulatory authority for BPG, and 2 of the 3 have experienced GMP quality issues in the past few years [21] (Box 2).
Another issue is that API manufacturers produce APIs with different technical specifications (e.g., particle size and excipient specifications) that adhere to different regulatory constraints. In turn, surveyed FDFs report difficulty in sourcing an alternate “like-for-like” API substitute should their first API manufacturer option go offline for regulatory, quality, or other reasons.
While supply-side delays can significantly impact the availability of BPG, demand-side issues drove stock-outs in 9 of the 10 countries reporting stock-outs surveyed by CHAI; these demand-side issues included (1) poor forecasting, (2) inflexible purchasing cycles, (3) lack of funding, and (4) limited BPG product registrations.
Forecasting inaccuracy resulted in the under-procurement of BPG in 5 of the countries that reported stock-outs to CHAI. Forecasting inaccuracy stems from a lack of product usage data at the facility level. In the absence of facility-level data, countries often rely on historical consumption, purchasing, or prevalence/incidence of BPG-responsive health conditions to estimate how much BPG should be ordered. These data sources may be outdated, be compromised by sub-national differences in incidence or prevalence, be uncorrected for previous stock-outs, be plagued by inaccurate record keeping, or fail to reflect clinician substitution behavior or account for the numerous other bacterial indications for which BPG may be used.
Any combination of these inaccuracies can lead to under-procurement of BPG, as many countries do not have available supplementary budget to purchase buffer stock (a supply of BPG to be held in reserve in case of future supply or demand fluctuations) or additional product. Even when the necessary funds are available, BPG purchasing can be complicated by a procurement system based on annual or 2-year tenders that may also have a minimum purchase order that must be met before fulfillment.
In these cases, even if countries recognize that their original BPG forecast was flawed, these tender processes cannot accommodate orders made mid-cycle, those that request small product volumes, or both. Consequently, countries are often unable to procure product until the next annual tender or an emergency procurement solution is implemented.
The problem is exacerbated when tenders are granted to a single supplier, as the supplier might not have the production flexibility to respond on short notice (Box 3). The study showed that the higher the number of BPG suppliers registered for use in the country, the lower the prospect of a stock-out. Given the lack of a quality-assured BPG product as assessed by a stringent regulatory authority, individual products from individual FDFs must be individually registered for use in each country. If there is a shock to the supplier market, countries without multiple products registered are at higher risk of stock-outs, as they have limited alternatives and must spend considerable time identifying new suppliers and having their products registered for national use.
Holding buffer stock is one mitigation strategy for supply disruptions. All 6 countries with no buffer stock available or planned had experienced stock-outs, while 3 of the 4 countries that had 3 to 6 months of buffer stock available had not experienced a stock-out in the past 3 years, as this buffer stock allowed them to cover shortages as they arose.
Accordingly, policies to strengthen and improve demand forecasting, in-country purchasing behaviors, and purchasing strategies are required. If these structural issues remain unaddressed and/or unmitigated, demand-side risk of stock-outs may remain even if global supply stabilizes in the next few years.
CHAI’s 14-country survey identified decreasing demand for BPG due to the substitution of newer antibiotic classes (cephalosporins, macrolides) in maternal syphilis treatment regimens. Clinically undesirable substitutions are driven by limited BPG availability, intermittent stock-outs, and a lack of BPG knowledge among pharmacists and providers, along with misperceptions of clinical indications, quality, and the possibility of adverse outcomes.
A summary of all the supply and demand elements of this global shortage is shown in Fig 3, and select responses from providers describing barriers to administration of BPG are provided in Table 3.
CHAI’s survey identified 3 countries where BPG was not the recommended first-line treatment for adult syphilis in clinical guidelines and 2 additional countries where BPG was not the most widely prescribed treatment for maternal syphilis despite clear guideline recommendations. Substitutions for adult, and specifically maternal, syphilis treatment were highlighted during interviews in 4 additional countries.
Although not specifically asked in the questionnaire, 3 countries in CHAI’s survey voluntarily reported that BPG demand was limited because many pregnant women were not being tested for syphilis during antenatal care visits. This was related to the limited availability of diagnostics, particularly low-cost rapid diagnostic tests, as well as a continued reliance on lab-based diagnostics, which often do not return results in a timely fashion.
In this evaluation, 39 countries of 95 (41%) responding to surveys reported shortages or stock-outs of BPG occurring during 2014–2016 (114 countries/territories surveyed). Both high-income countries and LMICs were affected.
Overall, the BPG market suffers from limited transparency across a fragmented landscape of API manufacturers, FDF suppliers, procurement agents, and local buyers. On the supply side, BPG presents an unattractive business case as it is an older, off-patent, sterile injectable drug that is expensive to make but commands a market price of pennies per dose. This has led to numerous market exits and the adoption of stringent margin optimization strategies by suppliers, such as inflexible production cycles and minimum order quantities. On the demand side, inaccurate forecasts and suboptimal purchasing strategies, such as inflexible purchasing cycles and sole sourcing of supply, lead to under-procurement of BPG.
This is the first comprehensive assessment to our knowledge to describe (1) the scope of BPG shortages at the national and procurement level, (2) the supply and demand drivers for this essential antibiotic, (3) relevant systemic factors leading to shortages, and (4) viable policy solutions to address the issue. Building on relevant reports such as the Global Status of BPG Report [22], this analysis clarifies how shortages of this drug equally affect developing and developed countries, and its detailed insights into suboptimal forecasting and the clinical use of BPG are also valuable for understanding how reduced demand impacts the global supply of this drug.
This study has limitations. The assessment was based on a convenience sample of countries, and of sources from participating countries, and should not be considered globally representative. Other countries may be experiencing BPG shortages that were not included in our study. Persons contacted during the assessment may not have been aware of BPG shortages at the time of the survey. Underreporting of substitution behaviors may have occurred due to desirability bias. Funding for the purchase of BPG by countries was not evaluated.
On the supply side, increasing transparency between the market players, and harmonizing product quality and specification standards across major buyers, is needed. A common product specification accepted by all major stakeholders would lower costs and ease administration for suppliers; it would simplify product substitution of registered and approved API manufacturers and FDFs should any single supplier discontinue operations for any reason, while decreasing supply uncertainty and product variability for buyers throughout the supply chain.
Accordingly, in December 2016, WHO applied for BPG to be listed as a “prequalified” medication [22] and invited API manufacturers and FDFs to apply for WHO prequalification. The WHO Prequalification of Medicines Programme helps ensure that medicines supplied by international procurement agencies such as UNICEF, the United National Population Fund, the Global Fund to Fight AIDS, Tuberculosis and Malaria, and Unitaid meet acceptable standards of quality, safety, and efficacy [23]. This designation can also establish manufacturing standards on particle size and dissolvability that could align the major producers of APIs and FDFs.
From a demand perspective, there must be twin policy emphases. The first is to strengthen the infrastructure that supports national-level BPG forecasting and procurement as a health system priority. When faced with a stock-out, many countries do not have available funds to purchase buffer stock or additional product. Even when necessary funds are available, many countries are constrained by inflexible annual tender processes that cannot accommodate small mid-cycle purchases. There is a need to generate awareness among country procurement agencies and provide technical assistance to help countries assess their need for BPG across the complete spectrum of BPG-treatable pathologies [22] (including primary prevention of rheumatic fever; treatment of pyoderma, yaws, bejel, and pinta; prophylaxis in sickle cell patients following splenectomy; and prophylaxis of recurrent cellulitis) to avoid stock-outs due to forecasts based on syphilis prevention alone. There is a need to generate awareness among country procurement agencies to preemptively mitigate stock-outs via increased product registration, buffer stock management, and rationing procedures to protect against supply disruptions and cover shortages as they arise.
The second policy emphasis should focus on supporting clinical testing and appropriate treatment of syphilis as a public health priority. Clinicians may incorrectly select broader spectrum antibiotics that can treat multiple infections but are less effective against syphilis. This is particularly important in pregnant women, where only BPG is known to cross the placenta barrier and prevent congenital syphilis [5,24,25].
Clinicians may make these incorrect treatment choices due to misperceptions of BPG’s clinical indications, quality, or likelihood of adverse outcomes. Penicillin-related anaphylaxis only occurs in approximately 1 patient per 100,000 administrations (range 0 to 3) [26]. Clinicians may equate even anecdotal reports of adverse events with a perception of a poor-quality drug product, which limits uptake of BPG, further decreases demand, and compromises procurement efforts.
If these structural issues remain unaddressed and/or unmitigated, demand-side risk of stock-outs may remain even if global supply stabilizes in the next few years.
Misperceptions of healthcare providers regarding the clinical indications for use of BPG were addressed by Argentina and Brazil in relation to these shortages affecting congenital syphilis prevention [27,28]. Retraining of healthcare workers on the WHO-recommended treatment of syphilis with BPG [5] and the technique for administration of BPG may be necessary to increase use. Reintroduction of BPG as a component of syphilis testing and treatment into national strategies and guidelines, particularly for maternal and neonatal health, among other health conditions for which WHO recommends the use of BPG, may also increase provider use and demand.
In 2014, WHO published the Global Guidance on Criteria and Processes for Validation: Elimination of Mother-to-Child Transmission of HIV and Syphilis [29]. Country progress towards achievement of validation indicators has led to increased demand for syphilis treatment and has uncovered several BPG supply chain gaps leading to medication shortages in low-, middle-, and high-income countries, many with a high burden of syphilis. Opportunities to improve global supply, demand, and use of BPG should be prioritized alongside congenital syphilis elimination efforts.
In summary, this assessment represents the first comprehensive analysis, to our knowledge, of the supply and demand drivers for the global shortage of BPG. The global targets for congenital syphilis elimination will not be met until this global BPG shortage is addressed both from a supply and demand perspective. Viable policy approaches to both strengthen procurement infrastructure and support the appropriate treatment of syphilis at the national level are needed.
|
10.1371/journal.pntd.0006949 | Vertical transmission of naturally occurring Bunyamwera and insect-specific flavivirus infections in mosquitoes from islands and mainland shores of Lakes Victoria and Baringo in Kenya | Many arboviruses transmitted by mosquitoes have been implicated as causative agents of both human and animal illnesses in East Africa. Although epidemics of arboviral emerging infectious diseases have risen in frequency in recent years, the extent to which mosquitoes maintain pathogens in circulation during inter-epidemic periods is still poorly understood. This study aimed to investigate whether arboviruses may be maintained by vertical transmission via immature life stages of different mosquito vector species.
We collected immature mosquitoes (egg, larva, pupa) on the shores and islands of Lake Baringo and Lake Victoria in western Kenya and reared them to adults. Mosquito pools (≤25 specimens/pool) of each species were screened for mosquito-borne viruses by high-resolution melting analysis and sequencing of multiplex PCR products of genus-specific primers (alphaviruses, flaviviruses, phleboviruses and Bunyamwera-group orthobunyaviruses). We further confirmed positive samples by culturing in baby hamster kidney and Aedes mosquito cell lines and re-sequencing.
Culex univittatus (2/31pools) and Anopheles gambiae (1/77 pools) from the Lake Victoria region were positive for Bunyamwera virus, a pathogenic virus that is of public health concern. In addition, Aedes aegypti (3/50), Aedes luteocephalus (3/13), Aedes spp. (2/15), and Culex pipiens (1/140) pools were positive for Aedes flaviviruses at Lake Victoria, whereas at Lake Baringo, three pools of An. gambiae mosquitoes were positive for Anopheles flavivirus. These insect-specific flaviviruses (ISFVs), which are presumably non-pathogenic to vertebrates, were found in known medically important arbovirus and malaria vectors.
Our results suggest that not only ISFVs, but also a pathogenic arbovirus, are naturally maintained within mosquito populations by vertical transmission, even in the absence of vertebrate hosts. Therefore, virus and vector surveillance, even during inter-epidemics, and the study of vector-arbovirus-ISFV interactions, may aid in identifying arbovirus transmission risks, with the potential to inform control strategies that lead to disease prevention.
| The East African region is endemic to diverse mosquito-transmitted arboviruses, though little is known about the role of vertical transmission in maintaining these viruses within mosquito vector populations during inter-epidemic periods. We sampled mosquito larvae from the Lake Baringo and Lake Victoria regions of Kenya and reared them to adults in the laboratory before screening them for mosquito-associated viruses by multiplex RT-PCR-HRM, cell culture, and sequencing. From the Lake Victoria region, we detected the arbovirus, Bunyamwera, which can cause febrile illness in humans, in Culex univittatus and vector competent Anopheles gambiae mosquitoes. We also identified diverse insect-specific flaviviruses in Aedes aegypti, Aedes luteocephalus, Aedes spp. and Culex pipiens mosquitoes. From the Lake Baringo region, we detected Anopheles flavivirus in An. gambiae mosquitoes. These findings demonstrate that naturally occurring vertical transmission potentially maintains viruses in circulation within the sampled vector species populations. Therefore, mosquitoes may potentially transmit a pathogenic arbovirus during their first bite after emergence. Because various insect-specific flaviviruses have recently been found to either inhibit or enhance replication of specific arboviruses in mosquitoes, their vertical transmission, as observed in this study, has implications as to their potential impact on both horizontal and vertical transmission of medically important arboviruses.
| The East African Great Lakes region is a recognized hotspot for a broad diversity of arthropod-borne viruses (arboviruses) [1] that affect humans and animals [2] and are transmitted by several mosquito genera (mostly Culex Linnaeus, Aedes Meigen, Anopheles Meigen, Mansonia Blanchard, and Aedeomyia Theobald species) [3–5]. Some mosquito species are capable of naturally maintaining viruses in circulation through vertical transmission [6–9]–up to 38 generations for San Angelo (SA) virus in Aedes albopictus, though with progressive decline in filial infection rate (FIR) in laboratory population bottlenecks [10].
The Lake Victoria and Lake Baringo regions of Kenya have historically been associated with arboviral diseases [11] and have unique lake and island biogeographies [12] in which arboviruses exist [5]. Outbreaks in the 1960s around the Lake Victoria basin involved Semliki Forest, chikungunya, and o’nyong-nyong viruses that are vectored by Culex, Aedes, and Anopheles mosquito species, respectively [13]. More recent studies have found seropositivity for arboviruses in humans [14–16]. During the recent 2006–2007 Rift Valley fever (RVF) outbreak in Baringo County, 10 mosquito species were implicated as potential vectors, among which Aedes pembaensis Theobald, Culex univittatus Theobald, and Culex bitaeniorhynchus Giles were reported as potential vectors for the first time [11].
Although widespread arboviral activity in human populations has been documented in the Lake Victoria and Lake Baringo basins, the role of vertical transmission among mosquito vectors in the maintenance of arboviruses within ecologies remains poorly understood [17]. To ascertain the competence of mosquitoes to horizontally transmit arboviruses between hosts, many methods have been used to collect and test different mosquito body parts (abdomen, saliva, and legs) for arboviruses [18]. However, vertical transmission of arboviruses from adult female mosquitoes to their offspring can also maintain viruses in circulation for generations within mosquito populations [6–10]. To investigate how vertical transmission in different mosquito species in Homa Bay and Baringo counties of Kenya may be maintaining endemic arboviruses in circulation, we set out to identify arboviral infections in laboratory-reared adults of field-caught larvae and pupae.
In 2012, immature mosquitoes were sampled from islands and mainland shores of Lake Baringo (in Baringo County along the Great Rift Valley) and Lake Victoria (in Homa Bay County) of Kenya (Fig 1) during the rainy season. In Baringo County, samples were collected in July and October 2012 from Kokwa Island, Nosuguro, Salabani, Kampi ya Samaki, Sirata, and Ruko. In Homa Bay County, samples were collected in April, May and November 2012 from Ringiti, Chamaunga, Kibuogi, Rusinga, Takawiri, Mfangano and Ngodhe Islands, and Ungoye, Luanda Nyamasare, Mbita and Ngodhe mainland sites on the Kenyan part of Lake Victoria. Sampling was conducted on unprotected public land concurrently with an adult mosquito genetic diversity survey conducted in the same study areas [19].
We collected eggs, larvae, and pupae with 350-ml standard dippers (Bioquip Products, USA) from their breeding sites and transported them to the Martin Lüscher Emerging Infectious Disease (ML-EID) Laboratory at the Duduville campus of the International Centre of Insect Physiology and Ecology (icipe) in Nairobi, Kenya. In the laboratory, we reared them to adults in their field-collected breeding water at 28°C temperature, 80% relative humidity, and 12-hour day and night cycles [20,21].
Before sampling, we obtained ethical clearance for the study from the Kenya Medical Research Institute (KEMRI) ethics review committee (Approval Ref: Non-SSC Protocol #310) and no protected species were sampled.
All reared adult mosquitoes were identified and sorted using morphological keys [22–25] in petri-dishes on frozen ice packs to keep them cold and to avoid degradation of any viruses in the samples. The ice packs were wrapped with paper towels to absorb moisture and prevent frosting of the petri-dishes. We stored pools of ≤25 reared adult mosquitoes in well-labelled 1.5 ml microcentrifuge tubes according to species, larval collection sites, sex, and dates in tubes in a -80°C freezer.
Ten pieces of 2.0-mm yttria-stabilized zirconia beads (Glen Mills, Clifton, NJ) and 400 μl of cold homogenization media (2% L-glutamine, 15% fetal bovine serum) (Sigma-Aldrich, St. Louis, USA) were added to each tube, which were placed on ice to keep them cold. The mosquito pools were then homogenized for 10 seconds in Mini-BeadBeater-16 (BioSpec, Bartlesville, OK, USA) followed by centrifugation for 10 seconds in a bench top centrifuge (Eppendorf, USA) at 1,500 relative centrifugal force (rcf) and 4°C. Aliquots of 210 μl of each homogenate were used for nucleic acid extraction and the remaining aliquots were stored in -80°C freezer as stock.
Nucleic acid (NA) was extracted from the 210-μl mosquito homogenate aliquots using the MagNA 96 Pure DNA and Viral NA Small Volume Kit (Roche Applied Science, Penzberg, Germany) in a MagNA Pure 96 automatic extractor (Roche Applied Science) and eluted into a final volume of 50 μl according to the manufacturer’s instructions. A reverse transcription-multiplex polymerase chain reaction with high-resolution melting (RT-PCR-HRM) analysis based arbovirus screening protocol recently developed by Villinger et al. [26] was used to rapidly screen many samples and detect the presence of four arbovirus genera, namely, Alphavirus (family Togaviridae), Flavivirus (family Flaviviridae), Bunyamwera-group Orthobunyavirus (family Peribunyaviridae), and Phlebovirus (family Phenuiviridae). Briefly, the High Capacity cDNA Reverse Transcription (RT) kit (Life Technologies, USA) was used to synthesize complimentary DNA (cDNA) of the nucleic acid extracts. cDNA synthesis from 5 μl of extracted nucleic acids was performed in 10-μl reaction volumes with final concentrations of 1x RT Buffer, 4 mM dNTP mix, 2.5 U/μl MultiScribe Reverse Transcriptase, 1 U/μl RNase Inhibitor, 600 μM non-ribosomal random hexanucleotide primers [27]. Reverse transcriptions were performed in a Veriti 96-Well Thermal Cycler (Applied Biosystems, Singapore) at 25°C for 10 minutes, 37°C for 2 hours, 85°C for 5 minutes and held at 4°C.
We used established multiplex RT-PCR thermocycling conditions [26] in a HRM capable Rotor-Gene Q real-time PCR thermocycler (Qiagen, Redwood city, CA, USA) to screen for virus sequences in cDNA templates. Ten microliter reactions consisting of 1 μl cDNA template, 5 μl 2x MyTaq HS Mix (Bioline, UK), 1 μl of 50 μM SYTO-9 saturating intercalating dye (Life Technologies), and multiplex PCR primers at concentrations given in Table 1. The QIAgility robot (Qiagen) for liquid handling was used to set up the reaction mixture. Touchdown PCR cycling conditions as detailed by Villinger et al. [26] included an initial denaturation at 95°C for 5 minutes, followed by 50 cycles of denaturation at 94°C for 20 seconds, annealing at 63.5–47.5°C for 20 seconds, and extension at 72°C for 5–30 seconds, followed by a final extension at 72°C for 3 minutes. Immediately after PCR, the product was held at 40°C for 1 minute before HRM analyses of PCR product double stranded DNA stability by measuring SYTO-9 fluorescence at 0.1°C temperature intervals increasing every 2 seconds from 75°C to 90°C. PCR grade water was used as negative control, and Bunyamwera (Orthobunyavirus), dengue and West Nile (Flavivirus), sindbis and Middelburg (Alphavirus), and Rift Valley fever (Phlebovirus) viruses were used as positive controls. Positive samples were re-run in singleplex reactions (using primers from only one genus; Table 1). Amplicons from singleplex runs were purified with ExoSAP-IT for PCR Product Kit (Affymetrix Inc., USA) and Sanger-sequenced at Macrogen (Korea). Samples that were positive for the Flavivirus genus by HRM analysis were further sequenced from nested PCR products using the 2NS5F (5’-GCNATNTGGTWYATGTGG-3’) and 2NS5Re (5’-TRTCTTCNGTNGTCATCC-3’) primers that amplify longer nucleotide fragments (~930 nt) of Flavivirus NS5 genes [28]. Resulting nucleotide sequences were edited using Geneious R7.1.9 software (created by Biomatters) [29].
To validate that the sequenced targets were truly viral and not viral genome segment inserts in the mosquito genome, a fraction of the original mosquito homogenates that were PCR-positive for potential arboviruses were subjected to cell culture in vertebrate BHK-21 (Kidney of Syrian hamster, Lot: 59300875 from ATCC) and Ae. albopictus clone C6/36 (Whole larva of Asian tiger mosquito, Lot: 60400699 from ATTC) cell lines. Stock mosquito homogenates of 19 samples with sequences that aligned with known viruses on GenBank [32] and RNA virus databases were subjected to cell culture. The homogenates were thawed on ice and clarified by centrifugation at 15,000 rcf and 4°C in a bench top centrifuge (Eppendorf 5417R) for 5 minutes. One hundred microlitres of the clarified supernatant were aseptically inoculated in each of sub-confluent BHK-21 and C6/36 cell lines in a 24-well culture plate. The BHK-21 cells were initially aseptically grown in growth media (GM; pH 7.5) made of 2% Minimum Essential Media (MEM; +Eagle’s salt, +25 Mm HEPES) with 10% FBS, 2% L-glutamine and 1% antimycotic (Sigma-Aldrich). The C6/36 GM contained same proportions of respective constituents as the BHK-21 GM, but with the addition of 1% non-essential amino acids (GIBCO, UK). The inoculated cell lines were incubated for 14 days and observed daily for any change in the morphology of the cell line caused by viral infection, also known as the cytopathic effect (CPE). Virus presence was ascertained as CPE. During the initial 14-day incubation period, any contaminated cell culture was purified using a 0.22 μm syringe filter [33] and re-tested. Further, RNA was extracted from cell culture wells that showed CPE and tested in single-genus arbovirus RT-PCR-HRM reactions and re-confirmed by sequencing, as described above.
Using Basic Local Alignment Search Tool (BLAST) [34], initial searches were performed for comparison of all obtained virus sequences with those in GenBank. This was followed by sequence alignments using the default settings of the MAFFT v7.017 [35] plugin in Geneious software, to identify virus segments. Maximum likelihood phylogenetic relationships of the study’s insect-specific flaviviruses (ISFVs) NS5 sequences with those of related ISFVs were analyzed using PhyML version 3.0 [36], employing the Akaike information criterion [37] for automatic selection of the general time reversible (GTR) sequence evolution model. Tree topologies were estimated using nearest neighbour interchange (NNI) improvements over 1000 bootstrap replicates. Rooting the phylogeny to the yellow fever vaccine strain sequence (GenBank accession NC_002031) as an outgroup, the phylogenetic tree was depicted using FIGTREE version 1.4.2 [38].
A total of 4,453 adult mosquitoes comprised of nine Aedes, six Anopheles, 16 Culex and one Mimomyia species were reared from immatures (Table 2). Among 612 pools of ≤25 mosquito samples per pool, 92 pools were from Baringo County and 520 pools were from Homa Bay County. Among mosquito pools from 32 species sampled in Homa Bay County, Bunyamwera virus (Orthobunyavirus) was the only vertically transmitted arbovirus (pathogenic to vertebrates) detected. It was identified by HRM analysis (Fig 2A), culture, and DNA sequencing (143 nt; 100% identity to GenBank accession KM507344, S1 Fig) from female Anopheles gambiae from Luanda Nyamasare (1/77 pools) and Cx. univittatus from Rusinga (2/31 pools) (Table 2) that were reared from larvae sampled in November 2012. However, no vertically transmitted pathogenic arbovirus was detected in Baringo County samples.
Further, we detected (Fig 2B) and sequenced ISFV NS5 sequences from 12 mosquito pools (Table 2, GenBank accessions: MG372051-MG372060, MK015647-MK015648) among May 2012 collections. Among Baringo samples, we sequenced three ISFVs from female An. gambiae mosquitoes (3/15 pools; one pool from Ruko and two pools from Kampi ya Samaki) collected in October 2012 that were closely related to Anopheles gambiae flaviviruses (An(g)FV) that were previously detected in mosquitoes sampled from Kenya’s North-Eastern Province [26], as well as Western and Coastal Provinces (Fig 3). Among Homa Bay County samples, we found Aedes flavivirus (AeFV) NS5 sequences in Ae. luteocephalus (3/13 pools; two pools from Ungoye and one pool from Mbita) and Aedes sp. (1/15 pools; from Takawiri Island), as well as in Cx. pipiens (1/140 pools; Rusinga Island). We also found cell fusing agent virus (CFAV), the first ISFV originally identified in Ae. aegypti using an Ae. albopictus cell line (C6/36) [39], among Homa Bay County Ae. aegypti (3/50 pools; Mfangano Island) and Aedes spp. (1/15 pools; from Ungoye) samples.
We identified natural infections of Bunyamwera virus and ISFVs in diverse anopheline and culicine mosquito species reared to adults from field-collected larvae, demonstrating that these viruses persist transstadially through development to adult stages from naturally infected immature life stages. Since vertical transmission was first identified of vesicular stomatitis virus by phlebotomine sandflies [40] followed by La Crosse virus in Aedes triseriatus [41, 42], this mode of maintaining arboviruses within ecosystems has been observed in numerous arboviruses of medical importance circulating in East Africa, including West Nile virus by Culex and Aedes mosquitoes [43–47], Ndumu virus [48] by Cx. pipiens, and Zika [49,50], dengue [51–53], chikungunya [54], and RVF [8] viruses by Aedes mosquitoes [55]. However, how widespread or important this mode of transmission is in natural ecologies remains poorly understood. While we attribute the naturally occurring virus infections that were transstadially transmitted from immature life stages in this study to vertical transmission from their parents, we cannot completely rule out the possibility that the immature mosquitoes were infected with these viruses from viral contamination in their aquatic environment during early development. However, this mode of transmission if far less likely as past studies indicate that such infection of immature mosquitoes requires unrealistically high viral doses in their aquatic environment [56].
We documented the vertical transmission of the Orthobunyavirus, Bunyamwera virus, from naturally occurring infections in two mosquito species–An. gambiae and Cx. univittatus–the former of which has previously been found to competently transmit Bunyamwera virus during blood-feeding on suckling mice [57]. This is of public health importance and needs to be monitored closely, as Bunyamwera is an important cause of acute febrile illness in humans (Bunyamwera fever) [58] that is able to reassort with closely related arboviruses to form new viruses, such as Ngari virus, which can cause haemorrhagic fever in humans [59]. With the well-established role of vertical transmission in Ae. triseriatus mosquitoes of the closely related Orthobunyavirus, La Crosse virus [60], the potential of Bunyamwera virus to remain in circulation by vertical transmission within mosquito populations in East Africa, highlights the importance of control strategies focused on vectors and the replication of arboviruses within the vector.
Recent laboratory vector competence studies have found that Bunyamwera virus can be competently transmitted by An. gambiae and Ae. aegypti mosquitoes [57], and can naturally infect Aedeomyia africana, Anopheles coustani, and Mansonia africana mosquitoes [5]. However, Culex quinquefasciatus was found to be refractory to Bunyamwera virus infection experimentally [57]. Our findings demonstrate that Bunyamwera infection persists from larval stages to adults in Cx. univittatus mosquitoes as well as in Bunyamwera competent An. gambiae. This expands the mosquito species, and indeed genera, that may play key roles in maintaining Bunyamwera virus in circulation. Though the vectorial competence of Cx. univittatus to transmit Bunyamwera virus has not been established, the species is thought to prefer birds as a source of bloodmeals [61] and has recently been found to also feed on dogs, donkeys, sheep, and toads [5], as well as humans [62]. Therefore, Cx. univittatus may have a greater potential for transmitting arboviruses between birds and other vertebrates to humans, in contrast to the more anthropophilic An. gambiae.
The vertically transmitted ISFVs, AeFV and CFAV, were only detected in samples from the Lake Victoria region, not only in Aedes mosquitoes, but also in Cx. pipiens (AeFV), though we cannot fully rule out accidental Aedes mosquito contamination in the Cx. pipiens sample. While vertical transmission of ISFVs has been reported experimentally [63–66], which may be as high as 90% [67], this study corroborates its occurrence in natural ecologies [64,65,68,69]. Although ISFVs do not infect mammals and generally have been found to cluster within distinct phylogenetic clades associated with distinct mosquito genera [70–72], Aedes flavivirus, which is phylogenetically distinct from related Culex flaviviruses, has previously also been found in Cx. pipiens mosquitoes sampled in Italy [73]. Our findings therefore support not only the vertical transmission of ISFVs in mosquitoes, but also the potential of occasional horizontal transmission between mosquito species and genera. Therefore, ISFVs in mosquito populations represent a promising model for the study of the evolution of host specificity of flavivirus infectivity [72].
Some ISFVs (Palm Creek flavivirus and Culex flavivirus) have been found to inhibit replication of West Nile and Murray Valley encephalitis viruses in the Ae. albopictus C6/36 cell line and in Cx. pipiens mosquitoes [65,66]. In contrast, CFAV, also identified in this study, has recently been found to increase susceptibility of dengue virus in an Ae. aegypti cell line (Aa20) [74] and to be inhibited by the Wolbachia endosymbiont (wMelPop) used for dengue control in Ae. aegypti mosquitoes [75,76]. Because there is considerable variability in how ISFVs effect arbovirus superinfections, how vertical transmission of ISFVs affects the competence of mosquito populations to transmit arboviruses, either horizontally to vertebrate hosts or vertically to the next generation, remains largely unknown.
We also detected An(g)FVs only in mosquito populations from the Lake Baringo region, despite the more than seven times greater sample size of An. gambiae tested from the malaria endemic Lake Victoria region. While it is curious that this ISFV was only detected in malaria mosquitoes from regions with relatively low malaria transmission rates [77], they have been previously identified in An. gambiae and Anopheles squamosus mosquitoes from malaria endemic North-Eastern Province [26], and Coastal and Western Provinces of Kenya (Fig 3). Other closely related Anopheles flaviviruses (AnFVs) (Fig 3) have since been reported in anopheline mosquitoes from Australia [71], Liberia and Senegal [78], and Turkey [79] (Fig 3). Furthermore, transcriptionally active Flavivirus-derived endogenous viral elements have been identified in Anopheles minimus and Anopheles sinensis genomes via in silico and in vivo analyses [80], which suggests a historical presence of ISFVs in anopheline mosquitoes. Though ISFVs may have important implications in the transmission of medically important arboviruses [70], the study of AnFVs has been limited by their inability to replicate in standard Aedes cell line cultures, or even in cell lines of heterologous Anopheles species [26,71]. Appropriate Anopheles cell line cultures for the in vitro replication of the AnFVs will have to be established to further study their role in co-infection with other arboviruses, and possibly malaria parasites [26].
We recorded more diverse vector mosquito species and viruses in samples from Homa Bay County (Table 2), which concurs with reports from previous studies around the Lake Victoria basin [5,16,19,81]. Although adult Aedes mosquitoes have been sampled in both study areas [5,82], we only sampled Aedes spp. larvae from Lake Victoria. In a previous study, we found that many of the suitable larval habitats for Ae. aegypti sampled in the Lake Victoria region correlated with increased ammonium and phosphate levels, which are key components of commonly used fertilizers [83]. Our larval sampling strategy may have been more favourable for sampling Aedes mosquitoes in the Lake Victoria region where agricultural activity is more intensive in comparison to the Lake Baringo region. Though there was an RVF virus outbreak in Baringo County in 2006/2007 and surveillance studies around the area reported possible mosquito vectors [11,84], none of our mosquito samples from Baringo County tested positive for any pathogenic virus.
Our identification of both a pathogenic arbovirus and three ISFVs in larval mosquitoes from both lake basins suggests complex ecologies involved in their circulation and maintenance. Although Omondi et al. [5] did not detect any virus from blood-fed mosquitoes around the Lake Victoria region where we found vertical transmission of Bunyamwera virus, AeFVs, and CFAV, the study found Bunyamwera virus in blood-fed mosquitoes from the Lake Baringo region, where we found no Bunyamwera infected larvae. Though these discrepant findings may be a result of inadequate sample size required to reliably identify specific arboviruses circulating in a region, the conditions for the maintenance of arboviruses by vertical transmission may depend on environmental factors of the mosquito vector’s reproductive environment. Nonetheless, our findings indicate that in the Lake Victoria region environmental context, An. gambiae, and possibly Cx. univittatus, can act as a reservoir that can both vertically and horizontally transmit Bunyamwera virus, ISFVs, and possibly other arboviruses. This is important towards understanding how arboviruses are maintained and geographically spread in different ecological contexts and can be used to forecast risks and improve prevention and other vector management strategies to mitigate future outbreaks. Continued arbovirus surveillance in diverse mosquito and other arthropod vector species in the region will help to more accurately identify the most important vectors of arboviruses possibly associated with febrile illnesses, while a better understanding of the role of ISFVs in the vertical transmission of arboviruses may open new control strategies.
Insect-specific flavivirus NS5 gene sequences from twelve mosquito pools were deposited into the GenBank nucleotide database (accessions MG372051- MG372060, MK015647- MK015648).
|
10.1371/journal.ppat.1002790 | An Endogenous Foamy-like Viral Element in the Coelacanth Genome | Little is known about the origin and long-term evolutionary mode of retroviruses. Retroviruses can integrate into their hosts' genomes, providing a molecular fossil record for studying their deep history. Here we report the discovery of an endogenous foamy virus-like element, which we designate ‘coelacanth endogenous foamy-like virus’ (CoeEFV), within the genome of the coelacanth (Latimeria chalumnae). Phylogenetic analyses place CoeEFV basal to all known foamy viruses, strongly suggesting an ancient ocean origin of this major retroviral lineage, which had previously been known to infect only land mammals. The discovery of CoeEFV reveals the presence of foamy-like viruses in species outside the Mammalia. We show that foamy-like viruses have likely codiverged with their vertebrate hosts for more than 407 million years and underwent an evolutionary transition from water to land with their vertebrate hosts. These findings suggest an ancient marine origin of retroviruses and have important implications in understanding foamy virus biology.
| The deep history of retroviruses is still obscure. Retroviruses can leave integrated copies within their hosts' genomes, providing a fossil record for studying their long-term evolution. Endogenous forms of foamy viruses, complex retroviruses known to infect only mammalian species, appear to be extremely rare, so far found only in sloths and the aye-aye. Here, we report the discovery of endogenous foamy virus-like insertions within the genome of a so-called ‘living fossil’, the coelacanth (Latimeria chalumnae). We provide evidence suggesting that foamy viruses and their hosts share a coevolutionary history of more than 407 million years, and that foamy viruses accompanied their vertebrate hosts on the evolutionary transition from water to land. These findings indicate that the retroviruses originated in the primeval ocean millions of years ago.
| Foamy viruses are complex retroviruses thought exclusively to infect mammalian species, including cats, cows, horses, and non-human primates [1]. Although human-specific foamy viruses have not been found, humans can be naturally infected by foamy viruses of non-human primate origin [2]–[4]. Comparing the phylogenies of simian foamy viruses (SFVs) and Old World primates suggests they co-speciated with each other for more than 30 million years [5]. Retroviruses can invade their hosts' genomes in the form of endogenous retroviral elements (ERVs), providing ‘molecular fossils’ for studying the deep history of retroviruses and the long-term arms races between retroviruses and their hosts [6], [7]. Although ERVs are common components of vertebrate genomes (for example, ERVs constitute around 8% of the human genome) [8], germline invasion by foamy virus seems to be very rare [9], [10]. To date, endogenous foamy virus-like elements have been discovered only within the genomes of sloths (SloEFV) [9] and the aye-aye (PSFVaye) [10]. The discovery of SloEFV extended the co-evolutionary history between foamy viruses and their mammal hosts at least to the origin of placental mammals [9]. However, the ultimate origin of foamy virus and other retroviruses remains elusive.
The continual increase in eukaryotic genome-scale sequence data is facilitating the discovery of additional ERVs, providing important insights into the origin and long-term evolution of this important lineage of viruses. In this study, we report the discovery and analysis of an endogenous foamy virus-like element in the genome of the coelacanth (Latimeria chalumnae), which we designate ‘coelacanth endogenous foamy-like virus’ (CoeEFV). The discovery CoeEFV offers unique insights into the origin and evolution of foamy viruses and the retroviruses as a whole.
We screened all available animal whole genome shotgun (WGS) sequences using the tBLASTn algorithm using the protein sequences of representative foamy viruses (Table S1) and identified several foamy virus-like insertions (Table S2 and Fig. S1) within the genome of L. chalumnae, one of only two surviving species of an ancient Devonian lineage of lobe-finned fishes that branched off near the root of all tetrapods [11]–[15]. There are numerous in-frame stop codons and frame-shift mutations present in these CoeEFV elements, suggesting that the CoeEFV elements might be functionally defective. Although more than 230 vertebrate genome scale sequences are currently available, endogenous foamy virus elements have been only found in the aye-aye, sloths, and coelacanth, indicating that germline invasion of foamy virus is a rare process [9], [10]. We extracted all contigs containing significant matches and reconstructed a consensus CoeEFV genomic sequence (Fig. S2). The resulting consensus genome shows recognizable and typical foamy virus characteristics (Fig. 1). Its genome has long terminal repeat (LTR) sequences at both 5′ and 3′ ends and encodes the three main open reading frames (ORFs), gag, pol, and env, in positions similar to those of exogenous foamy viruses (Fig. 1). Two additional putative ORFs were found at positions similar to known foamy virus accessory genes but exhibit no significant similarity (Fig. 1). Notably, we found that the Env protein is conserved among foamy viruses and the coelacanth virus-like element (Fig. 2). A Conserved Domain search [16] identified a conserved foamy virus envelope protein domain (pfam03408) spanning most (887 of 1016 residues) of the CoeEFV Env protein, with an E-value of 1.3×10−69 (Fig. 2). The CoeEFV Env protein shares no detectable similarity with other (non-foamy virus) retroviral Env proteins or with retroviral elements within available genomic sequences of other fishes, such as the zebrafish (Danio rerio). Hence, it provides decisive evidence that CoeEFV originated from a foamy-like virus.
To exclude the possibility that these CoeEFV elements result from laboratory contamination, we obtained a tissue sample of L. chalumnae and succeeded in amplifying CoeEFV insertions within the genome of L. chalumnae via PCR with degenerate primers designed for conserved regions of foamy virus pol and env genes.
To establish the position of CoeEFV on the retrovirus phylogeny, conserved regions of the Pol protein sequences of CoeEFV and various representative endogenous and exogenous retroviruses were used to reconstruct a phylogenetic tree with a Bayesian approach. The phylogenetic tree shows that CoeEFV groups with the foamy viruses with strong support (posterior probability = 1.00; Figs. 3 and S3), confirming that CoeEFV is indeed an endogenous form of a close relative of extant foamy viruses. The discovery of CoeEFV establishes that a distinct lineage of exogenous foamy-like viruses existed (and may still exist) in species outside the Mammalia.
Endogenous retroviruses are likely to undergo a gradual accumulation of neutral mutations with host genome replication after endogenization [17]. To date the invasion of CoeEFV into coelacanth genome, we identified two sets of sequences, each of which arose by segmental duplication because each set of sequences shares nearly identical flanking regions (Fig. S4). The two sets contain five and two sequences, respectively. Because the divergence time of the two extant coelacanth species (L. chalumnae and L. menadoensis) is uncertain [11], it is impossible to obtain a reliable neutral evolutionary rate of coelacanth species. Nevertheless, even using the mammalian neutral evolutionary rate [18] as a proxy for the coelacanth rate, the invasion dates were conservatively estimated at 19.3 (95% highest posterior density [HPD]: 15.3–23.6) million years ago for the dataset of five sequences. For the dataset containing two sequences, the divergence between the pair is estimated to be 4.1% and the invasion time is estimated to be approximately 9.3 million years ago. Because the CoeEFV invasion almost certainly occurred earlier than the duplication events within the host genome and because the evolutionary rate of coelacanth species is thought to be lower than other vertebrate species [19], [20], the time of CoeEFV integration might much more than 19 million years. Additional phylogenetic evidence (see below) suggests that its exogenous progenitors likely infected coelacanths for hundreds of millions of years prior to the event that fossilized CoeEFV within its host's genome.
To further evaluate the relationship of foamy viruses, we reconstructed phylogenetic trees based on the conserved region of Pol proteins of foamy viruses and Class III retroviruses, the conserved region of foamy virus Pol and Env protein concatenated alignment, and the conserved region of foamy virus Env protein alignment, respectively. The three phylogenies have the same topology in terms of foamy viruses (Figs. 4, S5, and S6). CoeEFV was positioned basal to the known foamy viruses (Fig. 4), suggesting a remarkably ancient ocean origin of foamy-like viruses: the most parsimonious explanation of this phylogenetic pattern is that foamy viruses infecting land mammals originated ultimately from a prehistoric virus circulating in lobe-finned fishes. The branching order of the three foamy virus phylogenies (Fig. 4, S5, and S6) is completely congruent with the known relationships of their hosts, and each node on the three virus trees is supported by a posterior probability of 1.0 (except the node leading to equine, bovine, and feline foamy viruses on the Env phylogeny, which is supported by a posterior probability of 0.94; Fig. S6). The common ancestor of coelacanths and tetrapods must have existed prior to the earliest known coelacanth fossil, which is 407–409 million years old [21]. The completely congruent virus topology, therefore, strongly indicates that an ancestral foamy-like virus infected this ancient animal. Crucially, the foamy viral branch lengths of the three phylogenies are highly significantly correlated with host divergence times (R2 = 0.7115, p = 1.10×10−5, Fig. 5; R2 = 0.7024, p = 1.41×10−5, Fig. S5; and R2 = 0.7429, p = 4.26×10−6, Fig. S6), a pattern that can reasonably be expected only if the viruses and hosts codiverged. It is worth emphasizing that we used a consensus sequence to represent CoeEFV in these analyses, so its branch length should correspond roughly to that of the exogenous virus that integrated >19 million years ago, rather than within-host mutations since that time.
There are two alternative explanations for these phylogenetic patterns. One is that the exogenous progenitor of CoeEFV is not truly the sister taxon to the mammalian foamy viruses, but a more distant relative. The robust posterior probability (1.00) placing them in the same clade and the absence of evidence for viruses or virus-like elements from other species disrupting this clade argue against this view, as does the significant similarity between the Env proteins of CoeEFV and the foamy viruses (Fig. 2). Moreover, its branch length would be difficult to explain under such a scenario. If the coelacanth foamy-like virus lineage and the mammalian foamy virus lineage did not share a most recent common ancestor in their ancestral host, why is CoeEFV neither more nor less divergent from the mammalian foamy viruses than one might expect if they did?
The other alternative to the hypothesis that these viruses have co-diverged over more than 407 million years is that they somehow moved, in more recent times, from terrestrial hosts to sarcopterygian hosts that inhabited the deep sea, and that the similarity of the coelacanth virus to the mammalian viruses is due to cross-species (in fact cross-class) transmission, rather than shared history. However, as illustrated by the significant correlation between host divergence times and viral distances (Figs. 5, S5, and S6), the long branches leading to CoeEFV and the clade of mammal foamy viruses suggest the virus had already circulated in vertebrates for an extremely long time before the origin of mammal foamy virus. Given that there is strong evidence that placental mammals were already being infected with foamy viruses by about 100 million years ago [9], the distinctness of the coelacanth virus suggests that it would have to have crossed from some other unidentified host, one whose foamy-like virus was already hundreds of millions of years divergent from the mammalian viruses. This seems highly unlikely. Although cross-species transmission of SFVs has been observed [2]–[5], [22], foamy viruses seem to mainly follow a pattern of co-diversification with their hosts [5], [9]. If one accepts that the endogenous foamy viruses within the genomes sloths indicate more than 100 million years of host-virus co-divergence, it seems plausible that CoeEFV extends that timeline by an additional 300 million years.
Moreover, the habitat isolation of the coelacanth and terrestrial vertebrates would have provided limited opportunities for direct transfer of foamy viruses to coelacanths. Taken together, these lines of evidence strongly suggest that foamy viruses and their vertebrate hosts have codiverged for more than 407 million years, and that foamy viruses underwent a remarkable evolutionary transition from water to land simultaneously with the conquest of land by their vertebrate hosts.
Our analyses provide compelling evidence for the existence of retroviruses going back at least to the Early Devonian. This is the oldest estimate, to our knowledge, for any group of viruses, significantly older than the previous estimates for hepadnaviruses (19 million years) [23] and large dsDNA viruses of insects (310 million years) [24]. Although highly cytopathic in tissue culture, foamy viruses do not seem to cause any recognizable disease in their natural hosts [1], [25], [26]. Such long-term virus-host coevolution may help explain the low pathogenicity of foamy viruses. The fact that the Env is well conserved between CoeEFV and foamy viruses is consistent with the fact that these viruses are asymptomatic and mainly co-evolve with their hosts in a relatively conflict-free relationship. It is easy to imagine that previously overlooked examples of such a non-pathogenic virus may yet be found in hosts that fill in some of the gaps in the phylogeny, namely amphibians, reptiles, and birds. It will be of interest to screen these hosts, but also various fish species, for evidence of exogenous and/or endogenous foamy-like viruses.
Dating analyses provide the clearest evidence for when and where retroviruses originated. There is strong evidence that foamy viruses shared a common, exogenous retroviral ancestor more than 400 million years ago (since Env was present in both terrestrial and marine lineages). The discovery of endogenous lentiviruses demonstrates that lentiviruses, a distinct retroviral lineage that includes HIV, are also millions of years old [27]–[30]. Foamy viruses and lentiviruses share a distantly related ancestor (Figs. 3, S3) and the foamy virus clade alone almost certainly accounts for more than 407 million years of retroviral evolution. It follows that the origin of at least some retroviruses is older than 407 million years ago. As with the coelacanth lineage in the foamy virus clade, we found that retroviruses of fishes occupy the most basal positions within both the Class I and Class III retroviral clades (walleye dermal sarcoma virus (WSDV) and snakehead retrovirus (SnRV), respectively, blue asterisks), (Figs. 3, S3). This pattern provides additional evidence of a marine origin and long-term coevolution of these major retroviral lineages. However, to be specific, the phylogenetic reconstruction in Fig. 3 reflects the history of only of the Pol protein, not a comprehensive history of retroviral genomic evolution. Nevertheless, our analyses support a very ancient marine origin of retroviruses.
All available animal whole genome shotgun (WGS) sequences from GenBank were screened for endogenous foamy viruses using the tBLASTn algorithm and the protein sequences of representative exogenous and endogenous foamy viruses (Table S1). Sequences highly similar to foamy virus proteins discovered within the coelacanth WGS were aligned to generate a CoeEFV consensus genome. Conserved domains were identified using CD-Search service [16].
Ethanol preserved Latimeria chalumnae tissue sample was obtained from Ambrose Monell Cryo Collection (AMCC) at the American Museum of Natural History, New York. Genomic DNA was extracted using the DNeasy tissue kit (QIAGEN, MD) following the manufacturer's instructions. Amplification of ∼680 bp gag gene and ∼650 bp env gene fragments was performed with the degenerate primer pairs, FVpol-F (5′-AACAGTGYCTYGACCMAACC-3′) and FVpol-R (5′-TAGTGAGCGCTGCTTTGAGA-3′), FVenv-F (5′-CTGGGGATGACAAYCAGAGT-3′) and FVenv-R (5′-CCACTCRGGAGAGAGGCAAC-3′). PCR was performed in 25 µl of final volume reactions with 0.1 µl Platinum Taq HiFi enzyme (Invitrogen, CA), 1 µl primer mix (10 µM each), 0.5 µl of 10 mM dNTP mixture, 1 µl of 50 mM MgSO4, 2.5 µl of 10× PCR buffer, and 1 µl of template DNA. The PCR reactions were cycled under the following conditions: initial denaturation at 94°C for 2 minutes, 45 cycles of (94°C for 15 seconds, 60°C for 60 seconds, and 72°C for 30 seconds), and final elongation at 72°C for 5 minutes. The PCR products were purified using QIAquick spin columns (QIAGEN, MD). Purified PCR products were cloned into the pGEM-T Easy vector (Promega, WI). Cloned products were sequenced by the University of Arizona Genetics Core with an Applied Biosystems 3730XL DNA Analyzer. The sequences have been deposited in GenBank (Accession Nos. JX006240-JX006251).
All protein sequences were aligned using Clustal Omega [31]. Gblocks 0.91b was used to eliminate ambiguous regions and extract conserved regions from the alignments [32]. To determine the phylogenetic relationship between CoeEFV and other retrovirus, we reconstructed a phylogeny based on the conserved region of Pol proteins of CoeEFV and various representative exogenous and endogenous retroviruses (Table S1; Dataset S1). To further evaluate the relationship and divergence of foamy viruses, the conserved region of the foamy viruses and Class III endogenous retroviruses Pol protein (Dataset S2), the conserved region of foamy virus Pol and Env protein concatenated alignment (Dataset S3), and the conserved region of foamy virus Env protein alignment (Dataset S4) were used to infer phylogenetic trees. We were unable to discern positional homology for the first 143 residues of the Pol protein with reasonable certainty. These regions were excluded from all subsequent analyses. All the phylogenetic analyses were performed with MrBayes 3.1.2 [33] using 1,000,000 generations in four chains, sampling posterior trees every 100 generations. The rtREV amino acid substitution model [34] was used. The first 25% of the posterior trees were discarded. MCMC convergence was indicated by an effective sample size >300 as calculated in the program Tracer v1.5.
For the phylogenetic tree based on the foamy viruses and Class III endogenous retroviruses Pol protein, Class III endogenous retroviruses were used to root the foamy viral phylogeny (Fig. 4). Because there is no obvious outgroup for foamy virus Env protein, we rooted the phylogenetic trees inferred from foamy virus Pol and Env concatenated alignment and Env alignment using midpoint method (Figs. S5 and S6). Because the topologies of the host and virus trees were identical for the foamy viruses (Figs. 4, S5, and S6), we were able to plot host branch length (in millions of years) versus virus branch length (in expected amino acid substitutions per site) for every branch (both internal and external). The vertebrate host divergence times are based on references [21], [35], and [36].
The nucleotide sequences were aligned using MUSCLE [37]. To estimate the age of the CoeEFV invasion, we identified two sets of sequences, which contain five sequences (contig270160, contig184752, contig185880, contig245863, and contig236769) (Dataset S5) and two sequences (contig243355 and contig219087) (Dataset S6). Sharing the same flanking region, each set of sequences arose from segmental duplication. I) For the dataset of five sequences: the best-fitting model of nucleotide substitution was determined using jModelTest [38]. The typical mammal neutral evolutionary rate (2.2×10−9 substitutions per site per year, standard deviation = 0.1×10−9) was used as the rate prior [18]. The HKY substitution model was used. BEAST v1.6.1 (http://beast.bio.ed.ac.uk) was employed for Bayesian MCMC analysis with a strict clock model [39] and Yule model of speciation. MCMC chains were run for 100 million steps twice to achieve adequate mixing for all parameters (effective sample size >200). Tracer v1.5 was used to summarize and analyze the resulting posterior sample. II) For the dataset of two sequences: we calibrated the genetic distance between the pair based on the Kimura two-parameter model, in which transitions and transversions are treated separately.
|
10.1371/journal.ppat.1005675 | Mycobacterium tuberculosis Thioredoxin Reductase Is Essential for Thiol Redox Homeostasis but Plays a Minor Role in Antioxidant Defense | Mycobacterium tuberculosis (Mtb) must cope with exogenous oxidative stress imposed by the host. Unlike other antioxidant enzymes, Mtb’s thioredoxin reductase TrxB2 has been predicted to be essential not only to fight host defenses but also for in vitro growth. However, the specific physiological role of TrxB2 and its importance for Mtb pathogenesis remain undefined. Here we show that genetic inactivation of thioredoxin reductase perturbed several growth-essential processes, including sulfur and DNA metabolism and rapidly killed and lysed Mtb. Death was due to cidal thiol-specific oxidizing stress and prevented by a disulfide reductant. In contrast, thioredoxin reductase deficiency did not significantly increase susceptibility to oxidative and nitrosative stress. In vivo targeting TrxB2 eradicated Mtb during both acute and chronic phases of mouse infection. Deliberately leaky knockdown mutants identified the specificity of TrxB2 inhibitors and showed that partial inactivation of TrxB2 increased Mtb’s susceptibility to rifampicin. These studies reveal TrxB2 as essential thiol-reducing enzyme in Mtb in vitro and during infection, establish the value of targeting TrxB2, and provide tools to accelerate the development of TrxB2 inhibitors.
| Mycobacterium tuberculosis (Mtb) antioxidant systems represent attractive targets for developing novel tuberculosis therapies. We demonstrate that targeting thioredoxin reductase TrxB2 eradicates Mtb during acute and chronic mouse infections. TrxB2 inactivation caused thiol-specific oxidizing stress, perturbed growth-essential processes and resulted in lytic death. Unexpectedly, TrxB2 deficiency did not cause increased susceptibility to oxidative and nitrosative stress. To uncover the mechanistic consequences of depleting TrxB2, or other growth essential proteins, in viable and growing bacteria, we developed a “leaky” knockdown system, with which partial TrxB2 depletion was achieved. Importantly, these leaky mutants revealed that one of two TrxB2 inhibitors kills Mtb via TrxB2 inactivation. They also demonstrated that TrxB2 depletion results in hypersusceptibility to rifampicin suggesting that a TrxB2 inhibitor will synergize with this frontline anti tuberculosis drug.
| Endogenous oxidative stress represents an inevitable challenge for microbes adapted to an aerobic lifestyle [1]. In addition, pathogens like Mycobacterium tuberculosis (Mtb) are confronted with exogenous oxidative stress imposed by the host [2]. The production of antimicrobial oxidants is a critical host defense mechanism against Mtb [3,4]. Patients with germline mutations in phagocyte NADPH oxidase resulting in an impaired macrophage respiratory burst are predisposed to mycobacterial diseases including tuberculosis [5]. Mice lacking inducible nitric oxide synthase succumb to Mtb infection much faster than their wild type littermates [3]. The reactive oxygen and nitrogen species generated by these host enzymes can inactivate microbial iron-dependent enzymes, damage lipids and destroy DNA [1,6].
Not unexpectedly, Mtb is armed with a number of dedicated antioxidant systems to ensure replication and survival within its host. Notable members include catalase, alkyl hydroperoxidase, superoxide dismutase, mycothiol, ergothioneine, thiol peroxidase, thioredoxin reductase and a recently identified membrane-associated oxidoreductase complex [4,7–13]. The thioredoxin system, together with the glutathione system, regulates many important cellular processes, such as antioxidant pathways, DNA and protein repair enzymes, and the activation of redox-sensitive transcription factors [6,14]. Unlike many Gram-negative bacteria, which possess both systems, Mtb lacks the glutathione system [6,10]. Instead, mycothiol has been suggested as substitute for glutathione in Mtb [10]. Mycothiol-deficient Mtb requires addition of catalase for growth in vitro, but is not significantly attenuated in mice [15]. In contrast, there is evidence that thioredoxin reductase (TrxB2) is essential for growth in vitro, implying a unique role for TrxB2 [16–18]. Although purified TrxB2 has been shown to mediate detoxification of H2O2, peroxide, and dinitrobenzene in vitro [12,19,20], its role in oxidative stress defense in physiological conditions and its specific biological functions in Mtb physiology are poorly understood. Bacterial thioredoxin reductases have recently been demonstrated to be druggable targets [18,21], however, it has not been determined whether inactivating TrxB2 in vivo, in acute and chronic infections, attenuates Mtb.
To address these questions, we applied a tunable dual-control genetic switch [22] to generate a conditional TrxB2 mutant and evaluated the impact of TrxB2 depletion. Unexpectedly, depleting TrxB2 not only rapidly killed Mtb, but also led to bacterial lysis. TrxB2 depletion perturbed growth-essential processes, including sulfur and DNA metabolism and death could be prevented by addition of a strong disulfide reductant. In vivo depletion of TrxB2 resulted in clearance of Mtb during both the acute and chronic phases of infection. We generated deliberately leaky knockdown mutants to dissect the contribution of TrxB2 to oxidative stress detoxification and found Mtb with partially depleted TrxB2 highly susceptible to thiol-specific oxidizing stress, but, surprisingly, not to peroxide and reactive nitrogen species. The leaky knockdown mutants were used to evaluate the specificity of two TrxB2 inhibitors and revealed that targeting TrxB2 results in hypersusceptibility to the frontline anti-tuberculosis drug rifampicin.
We first established that TrxB2 is indeed required for growth of Mtb under standard laboratory conditions (S1 Fig). Because a deletion mutant could not be isolated, we generated a TrxB2 dual-control (DUC) strain (S2 Fig). In TrxB2-DUC expression of TrxB2 is controlled by both transcriptional silencing and inducible proteolytic degradation, while TrxC is constitutively expressed from its native promoter [22]. Upon addition of anhydrotetracycline (atc) TrxB2 protein was rapidly depleted and below the limit of detection after 6 hours, which corresponds to less than 5% of TrxB2 amount in wild type (wt) H37Rv (Fig 1A and S3 Fig). TrxB2 depletion not only inhibited Mtb growth in nutrition-rich 7H9 medium, but also led to rapid killing (Fig 1B and 1C). Bacterial viability declined by 2.7 log after 24 hours, and 3.4 log after 4 days of atc treatment, indicating that TrxB2 is required for bacterial growth and survival in replicating conditions. We also assessed the impact of inactivating TrxB2 on non-replicating Mtb, which is known to be tolerant to anti-TB drugs and, in part, responsible for the long duration of anti-TB chemotherapy [23]. TrxB2 depletion was induced with atc after 10 days of incubation in PBS. Remarkably, TrxB2 depletion killed ~90% of the bacilli after 48 h and 99.9% within two weeks of PBS starvation, highlighting that starvation-induced non-replicating Mtb depends on TrxB2 for survival as well (Fig 1D and 1E).
While culturing TrxB2-depleted Mtb in liquid growth medium, we observed that the culture gradually declined in optical density and turned clear (Fig 1F and 1G). This motivated us to ask whether TrxB2 depletion caused lysis of Mtb. Notably, mycobacterial death is not always accompanied by lysis. So far, only a small number of cell-wall targeting compounds have been shown to induce lytic death [24]. To further investigate whether lysis occurred upon TrxB2 depletion, we monitored the release of the cytoplasmic enzymes enolase (Eno), dihydrolipoamide acyltransferase (DlaT) and the proteasome beta subunit (PrcB) into the culture supernatant. Because Eno, DlaT and PrcB are generally not detected in the culture supernatant of intact mycobacterial cells, we consider their release as an indicator of bacterial lysis. Consistent with a previous report that meropenem-clavulanate caused Mtb lysis [24], we found Eno, DlaT and PrcB in the culture filtrate 6 days after exposure to meropenem-clavulanate (Fig 1H). There was no detectable lysis of TrxB2-DUC in the absence of antibiotic or atc, even after 9 days of incubation. In contrast, cytoplasmic proteins were readily detectable in the supernatant of TrxB2-DUC treated with atc for 6 or 9 days, confirming our hypothesis that TrxB2 depletion caused lytic death (Fig 1H). In contrast, depletion of nicotinamide adenine dinucleotide synthetase (NadE) which also rapidly kills Mtb [22], did not result in detectable lysis of NadE-DUC (S4 Fig). Microscopic analysis revealed that lysis of TrxB2-depleted Mtb was preceded by significant cell elongation (Fig 1I and S5 Fig). The majority of TrxB2-depleted bacteria were twice as long as those expressing TrxB2, suggesting that TrxB2 depletion affects processes required for cell division.
To evaluate the importance of TrxB2 for virulence of Mtb, mice were infected with TrxB2-DUC and fed doxycycline (doxy) containing food to inactivate TrxB2 at selected time points. The infection was rapidly cleared in mice given doxy food from the time of infection or during the acute phase of infection on day 10 (Fig 2A and 2B). No pulmonary pathology was observed in these mice (S6 Fig). Even when TrxB2 depletion was initiated during the chronic phase of infection on day 35, colony forming units (CFU) declined rapidly and no bacteria could be isolated from both lungs and spleens on day 160 (Fig 2A and 2B). The decline of CFU was accompanied by progressive healing of lesions in the lungs (Fig 2C and 2D). These results establish that TrxB2 is required for growth and persistence of Mtb in mice and point to the value of targeting TrxB2 to treat TB.
Although purified TrxB2 has been shown to reduce H2O2 and other peroxides, little is known about the detoxification function of TrxB2 in a physiological setting [12,19]. Therefore, we sought to evaluate the impact of partial TrxB2 depletion on the susceptibility of Mtb to oxidative stress. Achieving partial TrxB2 depletion to an extent that does not affect viability but significantly reduces the intracellular TrxB2 protein amount is technically challenging with a DUC strain because of the steep atc dose response curve of this regulatory switch [22]. To circumvent this problem, we generated a panel of TrxB2-TetON mutants that contain point mutations in the operator of the tet promoter resulting in different degrees of constitutive, leaky transcription upon atc removal. Transcription from the mutated tet promoters is similar without TetR, however leaky repression results in a range of promoter activities without atc (Fig 3A). Two of the leaky TrxB2-TetON mutants, TrxB2-tetON-WT and TrxB2-tetON-1C, showed growth defects in the absence of atc (Fig 3B). Their growth defects correlated well with the protein depletion kinetics of TrxB2 (Fig 3C). These mutants thus achieved a phenotypically significant level of TrxB2 depletion yet retained enough TrxB2 to support growth. The moderate impact on growth of TrxB2-tetON-1C permitted the use of standard minimal inhibitory concentration (MIC) assays to measure how inhibition of TrxB2 affects susceptibility of Mtb to different chemical stresses.
Surprisingly, partial inhibition of TrxB2 did not affect Mtb’s susceptibility to growth inhibition or killing by plumbagin, a superoxide generator (Fig 4A and S7 Fig). TrxB2 silencing only caused a 2-fold shift of the MIC of H2O2, and we did not detect significant survival differences between wild type H37Rv and the TrxB2-tetON mutant following H2O2 exposure (Fig 4B and S7 Fig). Additionally, we measured Mtb’s susceptibility to reactive nitrogen species and found that TrxB2-silenced Mtb was only slightly less resistant to acidified nitrite at a high concentration (Fig 4C). In contrast, TrxB2-silenced Mtb was 8–16 fold more susceptible to growth inhibition by diamide, a thiol-specific oxidant (Fig 4D). This hypersusceptibility suggested a specific role for TrxB2 in detoxifying thiol-oxidizing stress.
To determine if thiol-specific oxidizing stress was responsible for the lethality caused by TrxB2 depletion, we tested if supplementation with the strong thiol-reducing agent dithiothreitol (DTT) could prevent death of TrxB2-depleted Mtb. Indeed, DTT rescued viability of TrxB2-DUC in a dose-dependent manner (Fig 4E). In contrast, neither glutathione nor catalase provided any survival benefit (Fig 4F). These results indicate that the primary function of TrxB2 in Mtb is to detoxify thiol-specific oxidative stress and that TrxB2 is the dominant thiol-reducing enzyme in Mtb.
We sought to investigate the pathways affected in TrxB2-depleted Mtb and analyzed the transcriptome changes associated with TrxB2 depletion. We found an early induction of 61 genes after 6 hours of atc treatment (fold change >2, p<0.02), 12 of which belong to sulfur metabolism pathways (Fig 5A). Mtb converts imported inorganic sulfate into adenosine 5’-phosposulfate (APS), which can be used for metabolite sulfation [25,26]. Alternatively, APS can be sequentially reduced for the biosynthesis of essential sulfur-containing metabolites, including cysteine, methionine and mycothiol. The first committed step in this reductive branch, the conversion of sulfate to sulfide by APS reductase (cysH), requires reducing potential supplied by the thioredoxin system [25,26]. We observed extensive up-regulation of sulfate importer genes (cysT, cysW, cysA1 and subI) and genes in the reductive branch, including cysH and the O-acetylserine sulfhydrylase encoding cysK1 and cysK2, which indicates a response to compensate for a defect in sulfur assimilation. Consistent with that, TrxB2 depletion also resulted in increased expression of cysE, encoding a serine acetyl transferase, which is required for de novo cysteine biosynthesis (Fig 5A). The expression of sulfur metabolism genes remained induced at 24 hrs post atc treatment (Fig 5B) and we asked whether death could be prevented or delayed by addition of reduced sulfur metabolites. A cysH deletion mutant was viable and had no growth defect, as long as it was supplemented with either 2 mM cysteine or methionine [27]. However, neither cysteine nor methionine protected TrxB2-depleted Mtb from death, indicating that TrxB2 is required for other essential pathways besides sulfur metabolism (Fig 5C).
Indeed, among the most highly up-regulated genes after 24 h of TrxB2 depletion were those involved in DNA metabolism (Fig 5B). We observed extensive up-regulation of genes involved in three DNA repair pathways, including base excision repair (nei, alkA, ung, ogt and xthA,), nucleotide excision repair (ercc3, uvrA and uvrD2), and homologous recombination (recA, ruvA and ruvC), suggesting that inhibition of TrxB2 was associated with DNA damaging stress. In support of this, we found that partial TrxB2 depletion decreased Mtb’s tolerance to genotoxic stress caused by mitomycin C, a potent DNA crosslinker (Fig 5D).
Of note, several genes involved in cell division were significantly down regulated in TrxB2-depleted Mtb (Fig 5B) consistent with the observed cell elongation (Fig 1I). The induction of antioxidant genes (trxB1, trxC, thiX, ahpC, ahpD and mshA) and whiB3 encoding an intracellular redox sensor and regulator [28] further supports that TrxB2 depletion induces thiol-oxidizing stress.
Because DTT rescued survival of TrxB2-depleted Mtb (Fig 4E and 4F) we investigated its impact on the transcriptional changes caused by TrxB2 depletion. DTT treatment alleviated most of the mRNA changes associated with TrxB2 depletion without affecting atc-mediated transcriptional silencing and proteolytic degradation of TrxB2 (S8 and S9 Figs). It reduced the expression of most antioxidant genes to basal levels, suppressed the induction of sulfur metabolism genes, reduced suppression of cell division genes and decreased the activation of genes involved in DNA repair (S9 Fig).
Together, these data demonstrate that death following TrxB2 depletion was caused by pleiotropic effects on a number of growth-essential pathways, including sulfur and DNA metabolism, and was mediated primarily through exhaustion of thiol-reducing power (S10 Fig).
We utilized the leaky TrxB2 knockdown mutants to evaluate the specificity of two thioredoxin reductase inhibitors, ebselen and auranofin. Ebselen is a substrate of mammalian thioredoxin reductase, a competitive inhibitor of thioredoxin reductase from E. coli, and inhibits growth of Mtb [21]. Mtb’s susceptibility to ebselen was, however, not altered by partial TrxB2 depletion suggesting that ebselen inhibits Mtb growth by affecting other targets (Fig 6A). Auranofin, a gold-containing compound, was recently found to inhibit the enzymatic activity of Mtb’s TrxB2 in vitro and to kill Mtb [18]. Partial depletion of TrxB2 caused a 3.6-fold shift of the MIC of auranofin and sensitized Mtb to killing by 0.65 μg/ml auranofin, a concentration that did not affect viability of wt Mtb (Fig 6B and 6C). However, wt and mutant were killed similarly in the presence of a higher concentration of auranofin (Fig 6C). Our data suggest that TrxB2 is one of the major targets of auranofin, although auranofin likely inhibits multiple enzymes with reactive cysteine residues in Mtb, such as mycothione reductase [18].
To determine whether targeting TrxB2 sensitizes Mtb to other antimicrobial compounds, we screened the leaky TrxB2-TetON-1C mutant against a panel of antibiotics, including most of the first and second line anti-TB drugs. We found TrxB2-depleted Mtb highly susceptible to the cell wall biosynthesis inhibitors vancomycin and moenomycin (Fig 6D and 6F). Moenomycin directly inhibits bacterial peptidoglycan glycosyltransferases, while vancomycin can block both transglycosylation and transpeptidation by binding to the terminal D-Ala-D-Ala residues of the peptide stem [29]. Other inhibitors of peptidoglycan transpeptidation such as ampicillin, did not affect TrxB2-depleted Mtb more than wt Mtb (Fig 6F). Thus inhibiting TrxB2 may impair transglycosylation, which could contribute to the lysis phenotype we observed. Unexpectedly, depleting TrxB2 decreased the MIC of rifampicin by 5.6 fold, suggesting that a compound that inhibits TrxB2 may synergize with this important first line anti-TB drug (Fig 6E).
The paucity of targets that are both biologically validated and susceptible to inhibition by drug-like small molecules, i.e. “druggable”, is a major bottleneck in antimycobacterial drug development. Mtb’s thioredoxin reductase TrxB2 has recently been shown to be druggable, yet its biological evaluation has not advanced beyond the prediction of its essentiality for growth of Mtb on standard agar plates [18]. Auranofin inactivates thioredoxin reductase in vitro but has multiple targets in bacteria, including in Mtb [18,30]. It was thus unknown how the specific inhibition of TrxB2 would affect Mtb in different environments including those encountered during acute and chronic infections. We addressed these questions using genetic strategies and found that inactivating TrxB2 quickly eradicated Mtb during the acute and, importantly, the chronic phase of mouse infection, validating TrxB2 as a valuable target for therapeutic intervention. Deliberately leaky TrxB2 knockdown mutants revealed that a TrxB2 inhibitor may synergize with rifampicin. Treatment combinations of rifampicin and a TrxB2 inhibitor could thus reduce the required drug dosage and limit the frequency of resistant mutants as shown for the synergistic action of carbapenems and rifampicin [31]. We used a leaky TrxB2 mutant to determine the specificity of two TrxB2 inhibitors. The MIC of ebselen was not affected by partial TrxB2 depletion, suggesting that ebselen inhibits Mtb growth primarily through targets other than TrxB2. Ebselen has been shown to bind covalently to a cysteine residue located near the antigen 85 complex (Ag85C) active site and may thereby disrupt the biosynthesis of the mycobacterial cell envelope [32,33]. Auranofin was significantly more active against TrxB2-depleted Mtb than wild type indicating that it exerts its antimycobacterial activity at least partially through inhibiting TrxB2. However, auranofin exhibits a higher affinity for human thioredoxin reductase than for bacterial enzymes [34]. Furthermore, auranofin, an FDA-approved anti-rheumatic drug, has immunosuppressive activities by inhibiting NF-κB signaling and decreasing the production of nitric oxide and pro-inflammatory cytokines, which are critical for anti-TB immune responses [35,36]. It also has anti-tumor activity through inhibition of proteasome-associated deubiquitinases [37–39]. The catalytic mechanisms of mammalian and bacterial thioredoxin reductases are significantly different and the crystal structure of TrxB2 has been solved [6,20]. It should thus be possible to identify inhibitors that are more specific for TrxB2 than auranofin. We expect the leaky TetON mutants we constructed for this study to facilitate the identification of such inhibitors.
In addition to determining TrxB2’s value as a potential target for drug development we wanted to gain insights into the physiological functions of TrxB2, especially its role in detoxifying oxidative stress. TrxB2 expression is induced upon oxidative and nitrosative stress and purified TrxB2 can mediate the reduction of H2O2, peroxide, and dinitrobenzene [12,19]. However, TrxB2-depleted Mtb was hypersensitive specifically to thiol-oxidizing stress, but not to other types of oxidants, and the thiol reductant DTT prevented death caused by TrxB2 depletion. DTT did not promote growth of TrxB2-depleted Mtb, likely because DTT is very labile in neural aqueous solution and it is therefore difficult to maintain a constant concentration over time. Alternatively, TrxB2 has a function beyond its enzymatic activity, which is required for optimal growth and cannot be replaced by DTT. Notwithstanding, these results indicate that the primary function of TrxB2 in Mtb is to detoxify thiol-specific oxidative stress. Its potential role in defending against H2O2, superoxide and nitrosative stress is likely redundant with other antioxidant systems.
Mycothiol, a low-molecular-weight thiol present in millimolar quantities in mycobacterial cells, is thought to function as the mycobacterial substitute for glutathione and serve as the major redox buffer system in Mtb [10]. Mtb mycothiol-deficient mutants have a dramatically reduced intracellular thiol concentration, require catalase for optimal growth in vitro and exhibit increased sensitivity to oxidants. However, they are viable in vitro and only slightly attenuated in immunecompetent mice [15,40]. In contrast, TrxB2 depletion caused rapid lytic death even in the absence of exogenous oxidative stress and death was only prevented by DTT, but not catalase, cysteine and glutathione. Furthermore, TrxB2-depleted Mtb was unable to establish and maintain infection in mice. These phenotypic differences between mutants of the two major mycobacterial thiol-reducing systems emphasize that the TrxB2-dependent system provides the dominant thiol-reducing source to maintain thiol redox homeostasis. Recently, upregulation of thioredoxin genes in mycothiol deficient Mtb has been observed supporting that the thioredoxin system can restore mycothiol [11,41]. Some genes involved in DNA and sulfur metabolism were also differentially expressed in both mycothiol and ergothioneine deficient Mtb [11], however, the majority of these was down regulated, while we found them induced in response to TrxB2 depletion. Thus, while some relationships exist between ergothionine, mycothiol and the thioredoxin system, they represent to a large degree systems with distinct activities in maintaining redox balance.
Depriving thiol-reducing power via TrxB2 depletion affected numerous essential processes, including sulfur and DNA metabolism pathways. The conversion of sulfate to sulfide by APS reductase (CysH) requires reducing potential from the thioredoxin system, which may explain why TrxB2 depletion induced extensive up-regulation of the genes involved in cysteine biosynthesis [25,42]. TrxB2 depletion also strongly induced three different mycobacterial DNA repair pathways and consistent with this caused hypersusceptibility to the genotoxic drug mitomycin C. Ribonucleotide reductase (RNR) requires reducing power from the thioredoxin system to catalyze the reduction of NTP to dNTP [43], but TrxB2 depletion did not lead to increased sensitivity to the RNR inhibitor hydroxyurea. This is possibly due to the presence of both class I and class II RNRs in Mtb while hydroxyurea only inhibits class I RNR [44]. It is also possible that other DNA biosynthesis and repair enzymes rely on the thioredoxin system, a hypothesis we are currently investigating.
Surprisingly, we found that TrxB2 depletion lysed replicating Mtb. We observed significant cell elongation preceding lytic death consistent with the observed down-regulation of cell division genes. TrxB2-depleted Mtb was also highly susceptible to the peptidoglycan glycosyltransferases inhibitors moenomycin and vancomycin, but not to inhibitors of peptidoglycan transpeptidation, mycolic acid synthesis and arabinogalactan synthesis. We speculate that some enzymes or regulatory proteins involved in transglycosylation may depend on the thioredoxin system to maintain their intracellular redox states and function. Inactivation of TrxB2 may impair transglycosylation and thereby contribute to bacterial lysis. This observation also suggests a connection between redox-homeostasis and cell-envelope integrity in Mtb. We can therefore not exclude that TrxB2 depletion caused increased permeability to the sensitized compounds, although TrxB2-depletion did not cause susceptibility to all high molecular weight antibiotics.
In summary, this work identified TrxB2 as the dominant thiol-reducing enzyme in Mtb and refined understanding of its physiological roles in defending against thiol-oxidative stress and maintaining growth-essential pathways. Our results establish the importance of TrxB2 in Mtb pathogenesis and validate the enzyme as a drug target. The leaky TetON mutants we developed will facilitate target-based whole cell screens for the identification of TrxB2 inhibitors and can help maintaining on-target activity during drug development. We expect this strategy of partial transcriptional silencing to be widely applicable and to facilitate chemical-genetic interaction studies for other growth-essential proteins in Mtb and other pathogens.
All animal experiments were performed following National Institutes of Health guidelines for housing and care of laboratory animals and performed in accordance with institutional regulations after protocol review and approval by the Institutional Animal Care and Use Committee of Weill Cornell Medical College (Protocol Number 0601-441A).
Wild type Mtb (H37Rv) and its derivative strains were grown in Middlebrook 7H9 medium supplemented with 0.2% glycerol, 0.05% Tween-80, 0.5% BSA, 0.2% dextrose and 0.085% NaCl or on Middlebrook 7H10 agar containing OADC (Becton Dickinson and Company) and 0.5% glycerol. For growth of the TrxB2 leaky mutants, the above media were supplemented with 400 ng/ml anhydrotetracycline.
To generate Mtb trxB2-DUC, we first transformed wild type Mtb H37Rv with an attL5-site integration plasmid expressing trxB2 and trxC under the control of P750 promoter to obtain a merodiploid strain; trxB2 and the first 4 bps of trxC (the OFR of trxB2 overlaps with the first 4 bps of trxC ORF) were then deleted from the merodiploid strain by allelic exchange as previously described [45,46]. After confirming deletion of the native copy of trxB2 by Southern blot, we performed replacement transformations of attL5 inserts to generate TrxB2-DUC [22]. In the TrxB2-DUC mutant, TrxB2 was expressed under the control of a TetOFF promoter and with a C-terminal DAS+4 tag. We also introduced a copy of trxC under the control of its native promoter to the attL5 site of TrxB2-DUC. The leaky TrxB2-TetON mutants were generated by replacement transformation of Mtb ΔtrxB2::P750-trxB2-trxC with plasmids containing trxB2 under the control of leaky tet promoters. A copy of trxC under the control of its native promoter was also introduced to the attL5 site of leaky TrxB2-TetON mutants.
We transformed Mtb ΔtrxB2::P750-trxB2-trxC mutant with zeocin resistant plasmids expressing trxB2 and trxC, trxB2, trxC or vector control. The transformants were selected on zeocin containing 7H10 agar. ΔtrxC was isolated from Mtb ΔtrxB2::P750-trxB2-trxC transformed with the plasmid expressing only trxB2.
The PBS starvation assay was set up as previously described [22]. Bacteria were grown in 7H9 medium to mid-log phase, washed three times with PBST, and suspended in PBST. After incubation for 10 d, atc was added to the cultures of TrxB2-DUC, and CFU were determined by plating on 7H10 plates.
We prepared cell lysates from mid-log phase culture by bead-beating cell pellets in lysis buffer (50 mM Tris HCl pH 7.4, 150mM NaCl and 2mM EDTA) containing protease inhibitor cocktail (Roche). We then centrifuged the lysates at 13,000 rpm for 20 min and sterilized the supernatant by passing through 0.22 μm Spin-X filters (Costar). 30–60 μg total protein were separated by SDS–PAGE and transferred to nitrocellulose membranes for probing with rabbit antisera against Mtb TrxB2, enolase (Eno), proteasome beta subunit (PrcB) and dihydrolipoamide acyltransferase (DlaT). Recombinant full-length Eno and TrxB2 were expressed with a C-terminal His tag, purified and used as antigen for immunization of rabbits.
Culture filtrates were prepared as follows. Mtb strains were grown in 7H9 medium with 0.2% glycerol, 0.05% Tween-80, 0.5% BSA, 0.2% dextrose and 0.085% NaCl until the culture reached an OD of 0.6 ~ 0.8. Cultures were then washed three times with PBS to remove BSA and Tween-80. We next suspended the pellet in 7H9 medium supplemented with 0.2% glycerol, 0.2% dextrose and 0.085% NaCl. After incubation, culture supernatant was harvested by centrifugation and filtration through 0.22 μm filters. Filtrates were concentrated 100-fold by using 3K centrifugal filter units (Millipore) and analyzed by immunoblotting with antisera against DlaT, Eno, PcrB and Ag85B (Abcam, ab43019).
We infected female C57BL/6 mice (Jackson Laboratory) using an inhalation exposure system (Glas-Col) with mid-log phase Mtb to deliver approximately 200 bacilli per mouse. Mice received doxycycline containing mouse chow (2,000 ppm; Research Diets) starting at the indicated time-points. Lungs and spleens were homogenized in PBS, serially diluted and plated on 7H10 agar to quantify CFU. Upper left lung lobes were fixed in 10% buffered formalin, embedded in paraffin and stained with hematoxylin and eosin.
For transcriptome analysis of TrxB2-depleted Mtb, we grew TrxB2-DUC in 7H9 medium to an OD of 0.5~0.6 and then added 400 ng/ml atc. Samples were collected at 6 hr and 24 hr later. Each experiment was performed with at least three independent cultures. To determine the impact of DTT, TrxB2-DUC was treated with atc, DTT (2 mM) or both for 24 hrs. Microarray analysis was performed as previously described [47]. Cultures were mixed at a 1:1 ratio with GTC buffer containing guanidinium thiocyanate (4 M), sodium lauryl sulfate (0.5%), trisodium citrate (25 mM), and 2-mercaptoethanol (0.1 M) and pelleted by centrifugation. Bacterial RNA was isolated and labeled using a Low Input Quick Amp Labeling Kit (Agilent) according to the manufacturer’s instruction. Custom-designed Mtb H37Rv whole genome microarray (GEO platform GPL16177) were used. Analysis and clustering were performed with Agilent GeneSpring software. One-way ANOVA was used to compare microarray data, with Benjamini–Hochberg correction for multiple hypothesis testing. All the data have been deposited in the GEO database with the accession numbers GSE72328, GSE72329, GSE72330 and GSE78894.
For oxidative stress, Mtb strains were grown to mid-log phase and washed twice in 7H9 medium. Bacterial single cell suspension was then prepared by centrifuging the cultures at 800 g for 10 min to remove clumps. We then adjusted the OD to 0.03, treated Mtb strains with H2O2, plumbagin, diamide or acidified nitrite and determined CFU by plating.
Mtb was grown to mid-log phase and diluted to an OD of 0.03 in 7H9 medium. Bacteria were then exposed to 1.5-fold serial dilution of antimicrobial compounds. Optical density was recorded after 14 days and normalized to the corresponding strains without drug treatment. Minimum inhibitory concentration is defined as the lowest concentration of a drug at which bacterial growth was inhibited at least 90%, as compared to the control containing no antimicrobial compounds. Ampicillin, auranofin, D-cycloserine, ebselen, ethambutol, faropenem, hydroxyurea, isoniazid, kanamycin, levofloxacin, meropenem, mitomycin C, moxifloxacin, piperacillin, rifampicin, streptomycin and vancomycin were purchased from Sigma Aldrich, St. Louis, MO. Moenomycin was from Santa Cruz Biotechnology. Bedaquilline was received as a gift from C. Barry.
One-way ANOVA was used for multiple group comparisons. Two-tailed unpaired Student’s t test was used for the analysis of differences between two groups. Statistical significance was defined as P < 0.05 unless otherwise stated. No statistical methods were used to predetermine sample size.
|
10.1371/journal.pcbi.1007077 | Multiscale model of integrin adhesion assembly | The ability of adherent cells to form adhesions is critical to numerous phases of their physiology. The assembly of adhesions is mediated by several types of integrins. These integrins differ in physical properties, including rate of diffusion on the plasma membrane, rapidity of changing conformation from bent to extended, affinity for extracellular matrix ligands, and lifetimes of their ligand-bound states. However, the way in which nanoscale physical properties of integrins ensure proper adhesion assembly remains elusive. We observe experimentally that both β-1 and β-3 integrins localize in nascent adhesions at the cell leading edge. In order to understand how different nanoscale parameters of β-1 and β-3 integrins mediate proper adhesion assembly, we therefore develop a coarse-grained computational model. Results from the model demonstrate that morphology and distribution of nascent adhesions depend on ligand binding affinity and strength of pairwise interactions. Organization of nascent adhesions depends on the relative amounts of integrins with different bond kinetics. Moreover, the model shows that the architecture of an actin filament network does not perturb the total amount of integrin clustering and ligand binding; however, only bundled actin architectures favor adhesion stability and ultimately maturation. Together, our results support the view that cells can finely tune the expression of different integrin types to determine both structural and dynamic properties of adhesions.
| Integrin-mediated cell adhesions to the extracellular environment contribute to various cell activities and provide cells with vital environmental cues. Cell adhesions are complex structures that emerge from a number of molecular and macromolecular interactions between integrins and cytoplasmic proteins, between integrins and extracellular ligands, and between integrins themselves. How the combination of these interactions regulate adhesions formation remains poorly understood because of limitations in experimental approaches and numerical methods. Here, we develop a multiscale model of adhesion assembly that treats individual integrins and elements from both the cytoplasm and the extracellular environment as single coarse-grained (CG) point particles, thus simplifying the description of the main macromolecular components of adhesions. The CG model implements sequential interactions and dependencies between the components and ultimately allows one to characterize various regimes of adhesions formation based on experimentally detected parameters. The results reconcile a number of independent experimental observations and provide important insights into the molecular basis of adhesion assembly from various integrin types.
| As the linker between cytoskeletal adhesion proteins and extracellular matrix ligands, integrins play a vital role in the formation of adhesions and profoundly influence different phases of cell physiology, such as spreading, differentiation, changes in shape, migration and stiffness sensing [1–5].
Integrins are large heterodimeric receptors, with a globular headpiece projecting more than 20 nm from the cell membrane, two transmembrane helices, and two short cytoplasmic tails that bind cytoskeleton adhesion proteins (see Fig 1A). In order to form adhesions, integrins undergo lateral diffusion on the cell membrane, switch conformation from bent to extended, and change chemical affinity for extracellular matrix (ECM) ligands, EIL. Integrins also assemble laterally, owing to interactions with talin [6,7], kindlin [8], or glycocalyx [9], and can grow nascent adhesions into mature adhesions [10–12]. Integrin diffusion, activation, ligand binding, and clustering occur at the individual protein scale, but their effects can also be reflected on the cellular scale, resulting in a multiscale biological process. Simulations of adhesion assembly based on all-atom approaches are too detailed and computationally demanding to capture adhesion formation from multiple integrins. Instead, highly coarse-grained, CG, approaches based on Brownian Dynamics can condense the description of individual proteins into a few interacting CG “beads” that can recapitulate the emergent dynamics of complex biological systems from its individual components (see, e.g., Refs [13–17] for the example of cytoskeleton networks).
Nascent adhesions are complex biological systems that form near the leading edge of protruding cells, appearing as spots of about 0.1 μm in diameter, with lifetimes of 2–10 min (Fig 1B) [18–22]. Unfortunately, the small size and short lifetime of nascent adhesions have made it challenging to study them experimentally. Among 24 different integrin isoforms, the αvβ3 and α5β1 integrins, have important, but potentially separate roles in the assembly of adhesions and the physiology of many cell types [23–32]. Nanoscale differences in physical properties between αvβ3 and α5β1 integrins can determine how nascent adhesions assemble [33], their organization [34–36], transmitted traction [37] and lifetime [38], on account of their different properties. For example, it has been reported that the rate of integrin activation, ka, determines the number of integrins per adhesion [21,39], while lateral clustering, or avidity, EII, increases the size of individual adhesions [40–42].
Single-protein tracking experiments combined with super-resolution microscopy and computational methods have helped extract physical properties of different integrin types. β-1 and β-3 integrins were found to have diffusion coefficients of 0.1 and 0.3 μm2/s, respectively [43]. β-1 integrins also maintain their active conformation longer than β-3 integrins. Free-energy energy differences between active and inactive states revealed activation rates for β-3 integrins about 10-fold higher that β-1 integrins [44,45]. The intrinsic ligand binding affinity, EIL, of β-1 integrins for soluble fibronectin is about 10–50 fold higher than β-3 integrins, spanning an overall range for the two integrins of 3–9 kBT [46]. β-1 integrins display a catch bond and adhesion strength-reinforcing behavior and are stationary within adhesions [47–49]. β-3 integrins, on the other hand, rapidly transit from closed to open conformations, break their bonds from ligands more easily under modicum tensions, and undergo rearward movements within adhesions [26,38]. How these differences in diffusion, rate of activation, ligand binding affinity, and bond dynamics reflect on the assembly of nascent adhesions and on the probability of adhesion maturation remains elusive.
In this paper, we show that mixed populations of β-1 and β-3 integrins localize to both nascent and mature adhesions, suggesting that there could be important interactions between the two types of integrins. To address this question, we have developed a highly CG model of adhesion formation, based on Brownian Dynamics (Fig 2) and study how nanoscale physical properties of different types of integrins interplay in the assembly of nascent adhesions. The CG model treats individual integrins as point particles within an implicit cell membrane and includes actin filaments as explicit semiflexible polymers (Fig 2A). By incorporating nanoscale physical properties of individual integrins, sequential interactions and feedback mechanisms between integrin, ligands and actin filaments (Fig 2B–2D), the model is used to characterize the formation of micrometer-size adhesions at the cell periphery in a multiscale fashion. Our calculations show that integrins with high EIL and enhanced bond lifetimes, such as β-1 integrins, facilitate ligand binding, transmission of traction stress, and engagement of actin networks. By contrast, integrins with low EIL and lower ligand bond lifetimes, such as β-3 integrins, are correlated with clustering, repeated cycles of diffusion and immobilization, and weak engagement of actin filaments. The architecture of actin filaments does not impact the amount of ligand binding and integrin clustering, but determines the probability of adhesions maturation, consistent with previous experimental findings [50]. Collectively, our data reveal important insights into adhesions assembly that are currently very challenging to obtain experimentally. The data supports the general view that cells, by controlling physical nanoscale properties of integrins via expression of specific types, can regulate structural, dynamical, and mechanical properties of adhesions.
Motivated by our recent work on integrin catch-bonds regulating cellular stiffness sensing [5], we sought to investigate how interactions between different integrins could affect adhesion formation. Immunostaining in Human Foreskin Fibroblasts (HFF) for actin and either β-1 or β-3 integrins revealed that both types of integrins localize in nascent adhesions at the cell leading edge and in mature adhesions at the end of actin stress fibers (Fig 1B). This suggests that potential interactions between the different adhesion populations could be important during adhesion formation. To address this question, we developed a computational model to investigate how the nanoscale properties of different integrins affect adhesion formation and stability.
Since β-1 and β-3 integrins differ in ligand binding affinity, EIL, and strength of pairwise interactions, EII [44,45], we use the CG model to test how variations in EII and EIL impact adhesion assembly in terms of the amount of integrin clustering, ligand binding, and spatial arrangement of adhesions. Different morphological arrangements of integrin adhesions are detected (Fig 3A–3C). For high EII and low EIL, clustering is promoted (Fig 3D), but only a few integrins are bound to ligands (Fig 3E), resulting in few large integrin clusters (Fig 3A). Conversely, for low EII and high EIL, only a few integrins cluster (Fig 3D) while ligand-binding is promoted (Fig 3E), resulting in many ligand-bound integrins and few small integrin clusters (Fig 3B). When EII and EIL have intermediate values, a mix of big clusters of integrins that are weakly bound to the substrate and smaller, ligand-bound clusters co-exist (Fig 3C). By systematically varying EII and EIL, morphological regions differing in size and number of ligand-bound integrins versus clusters were precisely identified. A region of few large clusters exists for EII >3 kBT and EIL < 3 kBT; a region of many small adhesions exists for EII < 3 kBT and EIL > 3 kBT; the rest of the parameter space shows co-existence of intermediate-size clusters and ligand-bound integrins (Fig 3D–3E). The fraction of ligand-bound integrins increases with EIL and is independent from EII. (Fig 3E). By contrast, clustering is not independent from EIL and is promoted when EIL is low (Fig 3D). In the model, when active, integrins can bind free ligands and cluster, when in close proximity of a ligand or another active integrin, respectively. Since the number of ligands is higher than the number of integrins, the probability for an integrin to find a free ligand is higher than that of finding an active integrin. Therefore, clustering increases less with EII when EIL is high than when EIL is low (Fig 3D). This indicates that integrin clustering and ligand binding are competing mechanisms.
Together, our results show that different arrangements of nascent adhesions can be achieved depending on EII and EIL. When we use high EII and low EIL, as for β-3 integrins, clustering is enhanced, and ligand binding reduced; when we use high EIL and low EII, as for β-1 integrins, clustering is reduced, and ligand-binding promoted. Thus, the competition between clustering and ligand binding can be determined by the integrin type. However, β-1 and β-3 integrins also differ in their rates of activation, which can lead to differences in this competition, by promoting clustering at high EIL. Therefore, we next aimed to understand how activation rates, combined with variations in EII and EIL, impact clustering and ligand binding.
Competition between integrin clustering and ligand binding can be determined by the difference in activation rate between β-1 and β-3 integrins. By varying ka from 0.005 s-1 to 0.5 s-1, our model shows that both clustering and ligand binding are promoted (Fig 4A and 4B). Using EII = 5 kBT and varying EIL from 3 kBT to 11 kBT, clustering is independent from EIL (Fig 4A), while overall ligand binding increases with EIL (Fig 4B). Clustering is mostly set by the strength of pairwise interactions between integrins, EII. It can be promoted by low EIL. and high ka, leading to a higher number of integrins able to diffuse and cluster (Fig 4A). Ligand binding is proportional to EIL at all ka. In experiments, variations in integrin activation rate are tied to variations in ligand binding affinity, making it unclear whether it is ka or EIL that determines organization of nascent adhesions. Our model shows that the rate of integrin activation set the level of the competition between ligand binding affinity and strength of pairwise interactions (Fig 4A).
Experimentally, Mn2+ or antibodies are typically used to modulate ligand binding affinity [51–54]. Both of these approaches, however, not only increase ligand binding affinity, but also the lifetime of the ligand bond. The increase of the ligand bond lifetime can be formally represented using a catch-bonds [55], where ligand unbinding rates decrease under tension and promotes stress transmission from the adhesions [56]. Therefore, we next used the model to test how variations in catch bond kinetics, combined with differences in the relative amount of β-1 and β-3 integrins, modulate ligand binding and stress transmission.
Since nascent adhesions transmit tension between the cytoskeleton and the ECM, we next asked how mixing integrins with different load-dependent bond kinetics impacts ligand binding and transmitted tension. The β-1 and β-3 integrins both behave as catch bonds that differ for unloaded and maximum lifetimes (Fig 2C). In the model, an increase in the percentage of β-1 integrins while keeping the rest as β-3 integrins, increases ligand binding from about 5% to 35% when using actin flow speeds below 15 nm/s (Fig 5A). The percentage of ligand-bound integrins is in direct proportion to the amount of β-1 integrins (Fig 5A). At actin flow speeds below 15 nm/s, traction stress and flow rate are positively correlated, while at higher flows they are inversely correlated (Fig 5B), in agreement with previous findings [60, 61]. Interestingly, variations in the relative fractions of the two integrin types do not affect the average tension on each integrin-ligand bond (Fig 5B). Below 10 nm/s actin flow, the minimum separation between ligand-bound integrins decreases from about 120 to 10 nm by increasing the fraction of β-1 integrins (Fig 4C). Stable adhesions, with minimum separation between ligand-bound integrins of 70nm, form with at least 20% β-1 integrins (Fig 5C). Together, our results show that the relative fractions of β-1 and β-3 integrins cooperate with actin flow to determine ligand binding and adhesion stability.
Interactions of adhesions with a cytoskeleton network play important role in several cell activities, including spreading and migration. The actin cytoskeleton exists in different architectures, depending on the cell location and function. Therefore, we next considered how the architecture of the actin cytoskeleton can impact the formation of adhesions. We incorporated in the model explicit actin filaments, using random, crisscrossed, and bundled architectures (Fig 6A–6C). The model assumes that ligand-bound integrins can interact with actin filaments, and that binding to actin increases integrin activation rate, as detected experimentally [63, 64]. Increasing the fraction of β-1 integrins, ligand binding increases independent of network architecture (Fig 6D). By contrast, integrin clustering remains at about 20–30% when a percentage of β-3 integrins is used. When only β-1 integrins are used, integrin clustering decreases of about 3-fold, independent from network architecture (Fig 6E). The number of ligand-bound integrins with a separation less than 70 nm is enhanced using a bundled network architecture (Fig 6F). This suggests that the probability of adhesion stability and ultimately maturation is higher with bundled architectures relative to both crisscrossed and random distributions of actin filaments (Fig 6F).
Collectively, our results indicate that the architecture of the actin cytoskeleton does not modulate the amount of ligand binding and integrin clustering. However, actin network architecture determines the physical distribution of ligand-bound integrins in adhesions, with bundled actin filaments increasing the probability of adhesion stability, consistent with previous experimental observations [50].
Since our experiments show that different integrin types exist in nascent and mature adhesions (Fig 1B), a computational model is here developed in order to understand if differences in nanoscale physical properties of integrins reflect on adhesions. This is largely untested by experimental approaches because it is very challenging to simultaneously distinguish between integrin types and isolate their nanoscale physical properties. The model is used to study how ligand binding affinity, rate of integrin activation, strength of pairwise interactions, bond kinetics, as well as the architecture of a network of actin filaments modulate integrin organization in adhesions and stress transmission. Our results collectively show that ligand binding and integrin clustering are competing mechanisms and that bundled actin networks favor adhesions stability, and ultimately maturation.
The model is developed through three consecutive stages of increasing complexity: (i) simulations of single-point integrins diffusing on a quasi-2D surface and switching between active and inactive states, binding ligands, and interacting laterally; (ii) incorporation of an implicit actin flow and integrin/ligand catch bonds kinetics; (iii) binding of integrins to semi-flexible actin filaments in either random, bundled, or crisscrossed architectures. At all stages, we distinguish between β-1 and β-3 integrins, by using either exact, experimentally detected physical parameters, realistic fold differences between the two, or estimates from previous free energy calculations.
For high EIL, many active integrins bind ligands and the fraction of integrins that can diffuse, and cluster, is reduced (Fig 3D–3E). Accordingly, this happens when the fraction of β-1 integrins is higher than that of β-3 integrins (Fig 5A), since β-1 integrins have higher ligand-binding affinity than β-3 integrins. By contrast, with many free diffusing integrins that have low EIL, and are less likely to bind ligands, the fraction of integrins that can encounter each other, and cluster is enhanced and reduces ligand binding (Fig 3D–3E). This happens when the fraction of β-3 integrins is higher than that of β-1 integrins (Fig 5A) and also results from the higher diffusion coefficient of β-3 integrins with respect to β-1 integrins [43,44]. The result that ligand binding and integrin clustering are competing mechanisms is consistent with a kinetic Monte Carlo model showing that thermodynamics of ligand binding and dynamics of integrin clustering interplay [46]. Our model reproduces this competing process over the same range of ligand binding affinities and strength of pairwise interactions.
The molecular mechanisms resulting in integrin lateral clustering remain controversial. However, several lines of evidence have suggested that β -3 integrins assemble clusters more easily than β-1 integrins. For example, activation of β -3 integrins induces formation of clusters with recruitment of talin [7], while β-1 integrins require recruitment of many more signaling components in order to form clusters, such as FAK [57]. β-3 integrins cluster in response to talin binding without a concomitant increase in affinity [24], while β-1 integrins cluster only when extended [58]. Moreover, previous studies in U2OS cells showed that β-3 integrins cluster on both β-3 and β-1 integrin ligands, while β-1 integrin clusters are present in adhesions only on β-1 ligands [59]. Our result that β-1 integrins are correlated with ligand-binding while β-3 integrins are mostly responsible for clustering is consistent with the observation that clusters of β-1 integrins are present only on β-1 ligands, possibly because, in this case, ligand binding and not pairwise interactions facilitate adhesions assembly. Further evidence that β-3 integrins assemble more easily than β-1 integrins is provided by the reported spatial and functional segregation of the two integrin types. β-1 integrins translocate from the cell periphery to the cell center to withstand higher tensions, whereas β -3 integrins remain at the cell edges to do mechanosensitive activities via dynamic breakage and formation of multiple bonds with the substrate and with one another [38].
Our model also shows that the number of integrins per cluster, computed as the fraction of clustered versus ligand bound integrins, is in the range of 2–15 particles, depending on EII and EIL (Fig 3F). This value is comparable to the experimentally estimated number of integrins in nascent adhesions, between 5–7 [10]. By varying ligand density in the model, the ratio between clustered and ligand bound integrin particles does not vary significantly (S1 Fig), suggesting that the average number of integrins per cluster in nascent adhesions is not modulated by ligand concentration, consistent with previous experimental observations [10].
In the presence of actin flow, the fraction of β-1 integrins is positively correlated with ligand binding (Fig 5A). Above a threshold actin flow, however, ligand binding is almost suppressed, independent from relative amounts of β-1 and β -3 integrins (Fig 5A), because of faster ligand unbinding from both integrins. This reduction in bound ligands corresponds to a drop in the average tension per integrin upon increasing actin flow (Fig 5B). The biphasic response of tension to actin flow was previously observed experimentally [60] and is consistent with models of adhesion clutch assembly and rigidity sensing [61]. By increasing the fraction of β-1 integrins, a reduction of lateral integrin spacing is observed with our model (Fig 5C). Previous studies on the lateral separation of integrins in adhesions reported that a minimum spacing of 70 nm is required to form stable adhesions [62]. This value corresponds in the model to a minimum of 20% β-1 integrins (Fig 5B). This value represents a prediction from our CG model that can be experimentally tested in the future. When only β-3 integrins are used in the model, their lateral separation, upon binding ligands, is about 120 nm (Fig 5C), supporting the notion that β-1 integrins are needed to form stable adhesions. This is consistent with the fast binding/unbinding dynamics of β-3 integrins previously observed in experiments [28].
By incorporating a positive feedback between actin filament engagement and integrin activation, as observed in [63,64], the competition between clustering and ligand binding is maintained in all actin architectures (Fig 6D–6E). This positive feedback represents the functional link between cytoskeleton and adhesions, where an increase in the probability of ligand binding results from binding actin, via an inside-out pathway [12,65]. Our model shows that the number of ligand-bound integrins with an average separation below 70 nm is enhanced with a bundled architecture (Fig 6F), suggesting that this configuration favors adhesion stability, and ultimately maturation [50]. When integrins bind a bundled network, they are likely to re-bind in close proximity because bundled filament architectures present filaments that are spatially closer than filaments of crisscrossed or random networks, forming a spatial trap for the receptors. Of interest for future studies is mimicking conditions of actin filament turnover, in order to understand how a dynamic cytoskeleton can interplay with integrin mixing in forming nascent adhesions. This will help understanding outside-in pathways, where, for example, adhesions formation modulates actin filaments polymerization. A further extension of the model will incorporate dynamic ligands, interconnected by a fibrous extracellular matrix that deform under tension. We will study how adhesions formation can change ligand localization and how this, in turns, affects adhesions morphology.
Previous computational studies of integrin dynamics range from all-atom simulations and enhanced sampling methods for understanding integrin activation at the level of individual molecules [66–68], to lower resolution coarse-grained [61,69–74], lattice-based [75,76], diffusion-reaction algorithms [77] and theoretical models [78] for multiple integrins. With respect to the previous lower resolution models of multiple integrins, our new model allows us to directly incorporate properties of different integrin types, as detected experimentally (Fig 1B). The particle-based implementation scheme of our model is similar to that of other software for modeling the cytoskeleton, such as Cytosim [79] and Medyan [80]. However, important differences exist. In contrast to Cytosim, an explicit implementation scheme is used here because our time step, combined with the limited number of simulated particles (a few hundreds), allows us to achieve time scales of a few minutes, that are relevant for adhesion assembly, without excessive computational cost. In addition, in contrast to Medyan, our model does not have a scheme for solving stochastic reaction-diffusion equations, but instead focuses on the mechanics of particle interactions and displacements under deterministic and Brownian forces.
To conclude, with our highly coarse-grained model based on Brownian Dynamics, we extend the scope of previous theoretical and computational studies of integrin-based adhesions formation, by testing how differences in nanoscale properties of β-1 and β-3 integrins impact ligand binding, clustering and transmission of traction stress. By coupling physical parameters (such as diffusivity) together with chemical (i.e., affinity and receptor pairwise interactions) and mechanical (bond kinetics) parameters, and by using an explicit actin cytoskeleton, our model shows that nascent adhesions assembly can be finely tuned by differences in nanoscale physical properties of integrins. The CG model ultimately demonstrates that nanoscale differences in integrin dynamics are sufficient to determine ligand binding and integrin clustering. By incorporating dynamics of individual integrins in an explicit way, our model provides results that are consistent with a number of previous independent experimental observations, revealing important insight into the molecular origins of adhesion organization and mechanics. Taken together, our modeling results support the general view that a cell can control integrin expression to determine morphological and dynamic properties of adhesions.
In order to characterize how nanoscale physical properties of integrins impact the assembly of nascent adhesions, we developed a highly coarse-grained computational model based on Brownian Dynamics. The model is agent-based in the way sequential dependencies regulate interactions between integrin/ligand, integrin/integrin, and integrin/actin. Integrins, ligands, and actin filaments are explicit particles, while the cell membrane is implicit. Solvent molecules mimicking cytoplasmic effects are replaced by stochastic forces, depending on cytoplasmic viscosity. Inactive integrins diffuse and, when active, can bind ligands and interact laterally. When integrins are bound to ligands, they can engage actin filaments. The interaction between integrins and actin filaments locally increases integrin activation rate, ultimately resulting in a positive feedback between actin binding and ligand binding [61,81].
In order to distinguish between β-1 and β-3 integrins, we examine the effect of varying integrin activation rates, motility, ligand binding affinity, clustering, and bond kinetics (Fig 2A–2C). By varying the relative amounts of β-1 and β-3 integrins, we analyze fractions of ligand-bound integrins, clustered integrins, and average tension on integrin-ligand bonds (Figs 3–5). Moreover, we study the effect of different actin filaments architecture on adhesions morphology (Fig 6).
The model is an extension of our mechanosensing model [5] but differs from it in several ways. First, each integrin exists in either active or inactive state, determined by activation and deactivation rates. Second, the model incorporates tunable parameters for integrin physical properties, allowing us to discriminate between integrin types. Third, explicit semiflexible actin filaments are included.
The computational domain includes two systems: a square bottom surface, of 1 μm per side, and a rectangular 3D domain above the surface, with dimensions 1 x 1 x 0.04 μm (Fig 2A). The bottom surface mimics the substrate; the lower side of the rectangular domain mimics the ventral cell membrane above the substrate, while its inside space represents a 40 nm thick cytoplasmic region where actin filaments diffuse beyond the ventral membrane (Fig 2A). The cell membrane is separated from the substrate by 20 nm, a dimension characteristic of active integrin headpiece extension (Fig 1A) [82]. Within the cell membrane, integrins diffuse in quasi-2D and are restrained in the vertical direction by a weak harmonic potential with spring constant 100 pN/μm, mimicking membrane vertical friction. In the cytoplasmic region, a repulsive boundary is used on the top surface, to avoid filaments crossing the boundary. Periodic boundary conditions are applied on all lateral sides of the domain, in order to avoid finite size effects.
The model considers a given number of ligands on the substrate, randomly distributed and fixed in space. We use a ligand density of 1000#/μm2, of the same order of that used in a previous model of adhesions assembly [72]. Integrin density on the cell membrane is ~100#/μm2 [5]. Integrins are single-point particles, that are initially randomly distributed and diffuse over the course of the simulations. Integrin diffusion coefficient is D = 0.1 μm2/s for β-1 integrins and D = 0.3 μm2/s for β-3 integrins [43]. Introducing volume exclusion effects between integrins, in the form of a weak repulsion between nearby particles (1 pN force), does not change the fraction of ligand-bound integrins, their average separation, the mean tension per integrin and its distribution (see S2A–S2D Fig). Increasing the magnitude of this repulsion (10 pN), however, affect the average separation of integrins (S2E–S2F Fig).
Semiflexible polymers represent actin filaments as spherical particles connected by harmonic interactions. Filaments have fixed length of 0.5 μm, corresponding to 6 beads separated by 0.1 μm equilibrium distance. The model of actin filaments is explained in detail in [83]. Actin filament beads are subjected to both stochastic and deterministic forces. Stochastic forces on the i-th bead are random in direction and magnitude in order to mimic thermal fluctuations and satisfy the fluctuation-dissipation theorem:
〈Fistochastic∙FistochasticT〉α,β=2(kBTμ/dt)I^α,β
(1)
with I^α,β being the second-order unit tensor [84] and μ being a friction coefficient equal in three directions.
Deterministic contributions come from bending and extensional forces on the filament beads. The bending force is computed as:
Fibend=−dEspringdri=kBTlpl0∑j=1N−1d(tj∙tj−1)dri
(2)
where lp = 10 μm is actin filament the persistence length, N is the number of beads in a filament (N = 6) and ti=(rj+1−rj)|rj+1−rj|.
The extensional force on filaments beads is computed as:
Fiextension=−dEextensiondri=k2∑j=1N−1d(|rj+1−rj|−l0)2dr
(3)
where l0 is the equilibrium length of 0.1 μm, k is the spring constant of 100 pN/μm.
Each spherical particle of a filament represents a binding site for integrin and each binding site can interact with multiple integrins.
In order to mimic hierarchical formation of nascent adhesions [10], the algorithm incorporates sequential interactions between integrins, ligands and actin filaments. First, we simulate a system composed of only integrins and ligands, in order to explore the ways in which integrins cluster and bind ligands in an actin-independent way. Then, we add actin filaments and study the effect of actin network architecture on adhesions formation.
Integrins switch between inactive and active states, with rates of activation and deactivation ka = 0.5 s-1 and kd = 0.0001 s-1, of the same orders of those previously estimated [44,45,72]. Activation probability corresponds to Pa = kadt, with time-step dt = 0.0001 s, as the time of the smallest simulated phenomena. This large time-step is allowed because the extensional stiffness of actin filaments, 100 pN/μm, is smaller than the real actin filaments stiffness, of about 400 pN/nm [85]. Upon activation, integrins can interact with free ligands, using a harmonic potential (with equilibrium separation 20 nm and spring constant 1 pN/μm), and cluster with other active integrins, depending on relative distances. Ligand binding occurs within a threshold distance of 20 nm, which reflects the extension of the open conformation of αIIbβ3 integrin away from the membrane [82]. Each integrin can bind only one ligand, and each ligand can bind only one integrin, mimicking binding sites specificity. Clustering occurs below a threshold of 30 nm, a value of the same order of the integrin-to-integrin lateral separation observed experimentally [82] and one order of magnitude lower than the minimum separation between individual adhesions [86].
The probability of integrin deactivation is Pd = kddt. Once inactive, integrin loses its connections with ligands and other integrins. Integrins unbind ligands with dissociation probabilities depending on their affinities: P=λeEILdt, using a prefactor λ = 1 s-1, for simplicity. They break later connections with probabilities inversely proportional to strength of pairwise interaction: P=λeEIIdt. For β-1 integrins, we use high affinity, ~9 kBT; for β-3 integrins we use lower affinity, 3–5 kBT.
Ligand-bound integrins can establish harmonic interactions with semiflexible actin filaments below 5 nm, approximating the size of the intracellular integrin tails [82]. Since the exact molecular composition of the layer between integrin and actin can contain up to ~150 different proteins [87], a detailed modeling representation is not possible. Therefore, interactions between integrin and actin are approximated by harmonic potentials with equilibrium distance of 3 nm and spring constant of 1 pN/μm. These interactions simplify the ~40 nm layer of adhesion molecules, including vinculin, talin and α-actinin, and is consistent with the level of details of the simulations, where harmonic interactions are used to connect particles within 20–100 nm.
Displacements of integrin and actin filament particles are governed by the Langevin equation of motion in the limit of high friction, thus neglecting inertia:
Fi−ξidridt+FiT=0
(4)
where ri is a position vector of the ith element, ζi is a drag coefficient equal in three directions, t is time, Fi is a deterministic force, and FiT is a stochastic force satisfying the fluctuation-dissipation theorem [88]. Fi is the sum of forces resulting from interactions of integrins with a ligand and/or other particles in the system, and actin flow in a direction parallel to the substrate. Positions of the various elements are updated at every time step using explicit Euler integration scheme:
ri(t+dt)=ri(t)+1ξi(Fi+FiT)dt
(5)
Since contraction forces are not needed for the assembly of nascent adhesions [5,10,89], our computational model only incorporates forces mimicking actin retrograde flow. In order to simulate actin flow and characterize distribution of traction stress at various flow rates, a constant force is applied on ligand bound integrins, along y (Fig 2B). Lifetime of the bond between integrin and ligand follows the catch-bond formalism (Fig 2C), using: for β-1 integrins an unloaded affinity of 2 s and a maximum lifetime of 15 s; for β-3 integrins an unloaded affinity of 0.5 s and a maximum lifetime of 3 s. The parameters for the catch bond kinetics are from previous experimental characterizations [38,47,56]. Curves of bond lifetime versus tension are shown in Fig 2C.
For β-1 integrins, we implemented an unbinding rate as a function of the force acting on the bond, F:
ku(F)=0.4e−0.04F+4E−7e0.2F
(6)
For β-3 integrins, we used unbinding rate:
ku(F)=2e−0.04F+4E−6e0.2F
(7)
The functional forms of the catch bonds were taken from a model that assumes a strengthening and a weakening pathway for the bond lifetimes, using a double exponential with exponents of opposite signs [90,91]. This model was also used for previous simulations of integrin-based adhesions [5].
To mimic promotion of integrin clustering upon ligand binding and actin filament engagement [81], we introduce a positive feedback between binding of integrin to a filament and integrin activation rate. In the model, integrins can bind a filament only if already bound to a ligand. Upon binding to actin, integrin activation rate is increased by 2 to 4% relative to its initial value. This assumption is motivated by recent evidence from TIRF experiments on T-cells, where it was demonstrated that actin binding and correct ligand positioning are needed for integrin activation [81]. The positive feedback between actin binding and integrin activation rate also represents conditions of inside-out signaling, with increased affinity for ligand binding induced by the cytoplasm [65]. We use the model with the positive feedback (schematics in Fig 2D) to test the effect of different actin architectures on ligand binding and clustering (Fig 6).
For bundled and crisscrossed actin filament architectures, we impose spatial restraints on filaments pairs. Bundled architectures have harmonic connections between beads of filament pairs that keep the filaments in parallel; crisscrossed architectures impose 90 deg angle between the axis of filaments pairs, that keep them almost perpendicular.
Human Foreskin Fibroblasts (HFF) were purchased from ATCC and cultured in DMEM media (Mediatech) supplemented with 10% Fetal Bovine Serum (Corning), 2 mM L-glutamine (Invitrogen) and penicillin-streptomycin (Invitrogen). HFFs were plated on glass coverslips incubated with 10 μg/mL fibronectin (EMD Millipore) for 1 hr at room temperature. Cells were fixed 1 hr after plating by rinsing them in cytoskeleton buffer (10 mM MES, 3 mM MgCl2, 1.38 M KCl and 20 mM EGTA) and then fixed, blocked and permeabilized in 4% PFA (Electron Microscopy Sciences), 1.5% BSA (Fisher Scientific), and 0.5% Triton X-100 (Fisher Scientific) in cytoskeleton buffer at 37° for 10 minutes. Coverslips were subsequently rinsed three times in PBS and incubated with either a β1 antibody (1:100; Abcam product #:ab30394) or β3 antibody (1:100; Abcam product #:ab7166) followed by AlexaFluor 488 phalloidin (1:1000; Invitrogen) and a AlexaFluor647 donkey anti-mouse secondary antibody (1:200; Invitrogen).
Cells were imaged using a 1.2 NA 60X Plan Apo water immersion lens on an inverted Nikon Ti-Eclipse microscope using an Andor Dragonfly spinning disk confocal system and a Zyla 4.2 sCMOS camera. The microscope was controlled using Andor’s Fusion software.
|
10.1371/journal.pgen.1006681 | CLAVATA1 controls distinct signaling outputs that buffer shoot stem cell proliferation through a two-step transcriptional compensation loop | The regulation of stem cell proliferation in plants is controlled by intercellular signaling pathways driven by the diffusible CLAVATA3 (CLV3p) peptide. CLV3p perception is thought to be mediated by an overlapping array of receptors in the stem cell niche including the transmembrane receptor kinase CLV1, Receptor-Like Protein Kinase 2 (RPK2), and a dimer of the receptor-like protein CLV2 and the CORYNE (CRN) pseudokinase. Mutations in these receptors have qualitatively similar effects on stem cell function but it is unclear if this represents common or divergent signaling outputs. Previous work in heterologous systems has suggested that CLV1, RPK2 and CLV2/CRN could form higher order complexes but it is also unclear what relevance these putative complexes have to in vivo receptor functions. Here I use the in vivo regulation of a specific transcriptional target of CLV1 signaling in Arabidopsis to demonstrate that, despite the phenotypic similarities between the different receptor mutants, CLV1 controls distinct signaling outputs in living stem cell niches independent of other receptors. This regulation is separable from stem cell proliferation driven by WUSCHEL, a proposed common transcriptional target of CLV3p signaling. In addition, in the absence of CLV1, CLV1-related receptor kinases are ectopically expressed but also buffer stem cell proliferation through the auto-repression of their own expression. Collectively these data reveal a unique in vivo role for CLV1 separable from other stem cell receptors and provides a framework for dissecting the signaling outputs in stem cell regulation.
| The proliferation of plant stem cells in above ground tissues is controlled by a suite of receptors in response to the CLAVATA3 peptide ligand. Receptor signaling in response to CLAVATA3 prevents over-proliferation of stem cells. It is unclear what the functional relationship is between the proposed CLAVATA3 receptors or if they impact common signaling outputs. Here I demonstrate that CLAVATA1 signals independently of the other receptors kinases to control distinct transcriptional outputs independent of stem cell proliferation. Stem cell proliferation is buffered by a two-step mechanism which transcriptionally regulates receptor levels in the stem cell niche. This mechanism helps explain the strict control of stem cell proliferation and could provide new avenues for improving plant growth.
| Co-operative receptor kinase function is a common feature in both animal and plant signaling systems. Receptor kinase mutants are frequently genetically additive in plants but the molecular mechanisms underlying this effect are often different. For instance, double mutants between the EF-TU RECEPTOR and FLAGELLIN SENSITIVE2 receptor kinases display enhanced susceptibility to bacterial infection above each single mutant [1], reflecting differences in pathogen derived ligands, followed by quantitative activation of common downstream outputs. On the other hand, additive genetic interactions among mutants in SOMATIC EMBRYOGENESIS RECEPTOR-LIKE KINASE family co-receptor kinases in response to specific ligands reflect quantitative redundancy as co-receptors [2]. Dissecting the molecular basis of redundancy in gene families in plants is often also complicated by unequal contribution from distinct genes and often requires in vivo analysis of signaling outputs or component interactions [3]. Balanced stem cell production in shoot (SAM) and floral meristems (FMs) is mediated by cell-to-cell signaling pathways initiated by the CLAVATA3 (CLV3) peptide ligand, a founding member of the CLE family of peptides [4]. Mutations in CLV3 lead to excess accumulation of stem cells in both SAM and FMs [5]. CLV3p is thought to be perceived by a series of overlapping receptor kinases which signal to dampen stem cell production. To date, much of the analysis of these receptor-ligand mutants has been at the gross morphological level. Mutations in the different proposed receptors vary considerably in their strength and genetic interactions. It is unclear if the morphological similarities, strength differences, or genetic interactions are due to co-operative cross talk, convergence on a common signaling output, or divergent pathways.
CLV3 is secreted from stem cells in the growing tip of the meristem, proteolytically processed and modified to a 13 amino acid diffusible glycopeptide (CLV3p) which diffuses broadly throughout the SAM [6–10]. Genetically, CLV3p perception is mediated by the transmembrane receptor kinase CLV1 [11–13], but also by a heterodimer of the receptor like protein CLV2 and the transmembrane pseudokinase CORYNE (CRN, [14–17]). Additionally, the receptor kinase RECEPTOR-LIKE PROTEIN KINASE 2 (RPK2) may also function in CLV3p perception [18]. All four mutants are resistant to exogenous CLV3p treatment to different degrees suggesting that they act as receptors for CLV3p in vivo. Previous data using overexpression in differentiated tobacco leaf cells has suggested that CLV1 could form higher order complexes with CLV2/CRN leading to the hypothesis that they may signal co-operatively in the SAM [19–22]. However conflicting results have been obtained by different groups and it is not clear if such complexes form in SAM tissues, or when receptors are expressed at endogenous levels in the appropriate cell types. While CLV3p has been reproducibly demonstrated to bind the CLV1 ectodomain [12, 23], differing results have been obtained for the ability of CLV2 to bind CLV3p [20, 23]. In addition, these studies have not tested for potential co-receptor binding of CLV3p. As such it is unclear how, or if, these receptors control stem cell proliferation co-operatively in vivo.
Strong alleles of clv1, such as clv1-4 and clv1-8, contain missense mutations in the LRR domain of CLV1 [11]. The residues effected in these mutants are highly conserved among related LRR-RKs and structural and biochemical studies with PXY/TDIF receptor ligand pair have shown that these residues direct ligand binding [24–28]. In contrast, clv1 null mutants are significantly weaker [29]. The weak stem cell defects in clv1 null mutant plants can partially be explained by the compensatory up-regulation of CLV1-related receptor kinases BARELY ANY MERISTEM1(BAM1), BAM2 and BAM3 [28]. In wild type plants, BAM receptor expression in the center of the SAM, where CLV1 is highly expressed, is undetectable. Consistent with this observation, triple mutants in bam1 bam2 bam3 have no defects in stem cell regulation on their own. In clv1 mutant SAMs BAM receptors are ectopically expressed in the center of the SAM and partially compensate for clv1. However it is unclear why null clv1 alleles are only weakly compensated by ectopic BAM expression. It is also unclear why strong clv1 alleles are phenotypically more severe. Previous work has suggested that strong clv1 mutant receptors may interfere with CRN/CLV2 signaling [16]. Alternatively, strong clv1 mutant receptors have been suggested to interfere with BAM signaling [30]. It is not clear how this relates to the feedback regulation of BAM expression by CLV1.
CLV1, and the other putative CLV3p receptors, are proposed to negatively regulate WUSCHEL (WUS) expression in the center of the SAM [13, 31]. WUS is a homeodomain transcription factor and de-repression of WUS in clv3 mutants is thought to drive excess stem cell proliferation [31, 32]. Despite this, the expression of WUS is robust and co-incident with CLV1 in wild type plants in the center of the SAM and WUS levels do not change dramatically at the cellular levels in response to loss of CLV1 signaling [11, 28, 32, 33]. Unlike WUS, CLV3p-CLV1 signaling fully represses BAM expression in the center of the SAM in wild type plants [28]. Plants expressing up to 300 fold higher levels of CLV3 have a wild type appearance, suggesting that repression of WUS is most effective outside of the physiological range of CLV3p concentration [34]. Interestingly, expression of CLV1 from the WUS promoter is necessary and sufficient to fully complement both clv1 null mutants and clv1 bam1 bam2 bam3 quadruple null mutants back to wild type levels of stem cell regulation [28]. As such, CLV1 operates exclusively in WUS expressing cells of the SAM, despite WUS being a target for transcriptional repression. It is not clear where in the SAM other proposed CLV3p receptors function or if WUS-mediated cell proliferation is linked to BAM transcriptional regulation by CLV1 in vivo.
Here I use quantitative genetics, and the highly specific transcriptional repression of BAM3 by CLV1 to demonstrate that CLV1 signals independent of CRN, CLV2 and RPK2 in response to CLV3p in vivo. In clv1 null mutants, ectopic BAM receptors compensate for CLV1 but also act in an additional feedback loop to dampen their own expression in the SAM and buffer stem cell proliferation. Strong alleles of clv1 specifically interfere with this process and have no impact on CLV2/CRN function. Despite their proposed ability to repress WUS, CRN/CLV2 function exclusively in WUS expressing cells of the SAM like CLV1. Consistent with this, WUS-induced stem cell proliferation is genetically separable from BAM3 regulation by CLV1. My data demonstrate that despite the qualitative phenotypic similarities, CLV1 signaling outputs diverge from other receptors, and from WUS, and support a model in which CLV1 is functionally independent of other proposed receptors in vivo.
In order to determine the functional relationship between the proposed CLV3p receptors I used previously published null alleles in CLV1, CLV2, BAM1, BAM2 and BAM3 in the Col-0 background [28, 35]. To date there are no null EMS generated alleles of CRN in any ecotype and no T-DNA insertions in the CRN coding sequence. I therefore used Cas9 to target CRN and create a null mutant. I created a gRNA targeting the signal sequence encoding region of the CRN gene and used the pCUT series of Cas9 vectors to create indels in the CRN gene in the Col-0 ecotype [36]. One of these, hereby referred to as crn-10, introduced a single thymine base between bases 20 and 21 in the CRN CDS. The mutation introduces a frameshift in the protein after amino acid 6 in the 33 amino acid signal sequence, leading to three in-frame stop codons four amino acids downstream. No other in frame methionine residues are present in, or before, the predicted CRN transmembrane sequence. Thus, the crn-10 allele retains 6 amino acids out of the original 402 in CRN and creates an early stop codon in the CRN signal sequence and deletes all downstream domains of CRN. crn-10 was segregated away from the Cas9 transgene for all subsequent work. crn-10 plants are qualitatively similar to other published crn alleles [16], with no new phenotypes noted. Wild type FMs give rise to stem cell populations that support stereotypical numbers of floral organs, culminating in the production of two central fused carpels. In clv mutant FMs, the enhanced rate of stem cell production results in increases in floral organ numbers, providing a quantitative measure of stem cell defects [29]. crn-10 plants displayed increased carpel number with a strength comparable to existing clv2 null alleles in Col-0 (Fig 1A). As such crn-10 behaves similarly to recessive non-null EMS alleles of crn such as crn-1 in the La-er ecotype [16]. crn-10 was fully complemented by a CRN-2xmCherry fusion protein expressed from the endogenous CRN promoter (44/44 lines, S1A Fig) as expected [17]. Previous work aimed at testing the hypothesis that CRN encoded a pseudokinase demonstrated that expression of a CRNK146E mutant protein, which further mutates the conserved active site lysine in CRN, fully complements crn-1 when expressed from the native CRN promoter [22]. At the time crn-1 was the only allele available and encodes a protein with a missense mutation in the CRN transmembrane domain (G70E) [16]. It is formally possible that if CRN homodimerized, the crn-1 and CRNK146E proteins could cross complement. I therefore transformed crn-10 with a pCRN::CRNK146E-2xmCherry transgene. Again I observed full complementation supporting the previous designation of CRN as a pseudokinase (26/26 lines, S1A Fig). Recent work has suggested that phosphorylation of CRN at serine 156 is important for function and that S156A substitutions fail to complement crn-1 when expressed from the 35S promoter [37]. In contrast, expression of either a S156A or a S156D CRN variant fully complemented the crn-10 null mutant when expressed from the CRN native promoter (26/27 and 24/24 line displaying full complementation for the S156A and S156D CRN variants respectively, S1A Fig). The reason for the different complementation results are unknown but could reflect either the use of different crn mutant plants or different transgene promoters in the complementation experiments.
I previously demonstrated that CLV1 function in WUS expressing cells is necessary and sufficient for stem cell regulation [28]. Both CLV2 and CRN are expressed broadly in inflorescence tissue (see S1B Fig, [15, 16]) but it is not clear if they function in WUS expressing cells of the SAM like CLV1. I therefore tested the ability of CLV2-myc and CRN-2xmCherry fusion proteins to complement their respective null mutants when expressed from either their native promoters or the WUS promoter. Like CRN, expression of CLV2-myc from the native CLV2 promoter in clv2 null mutant plants (rpl10-1, [35]) fully complemented stem cell defects in the majority of lines (Fig 1). Both CRN and CLV2 expression from the WUS promoter also fully complemented their respective null mutant plants in the majority of T1 lines (Fig 1). Fully complemented lines contained no flowers with more than two carpels, a level of complementation equivalent to complementation of the clv1 bam1 bam2 bam3 quadruple provided by the pWUS::CLV1-2xGFP transgene as previously reported [28]. Collectively these data indicate that like CLV1, CRN and CLV2 function exclusively in WUS expressing cells of the SAM and FM.
My data demonstrate that CRN, CLV2 and CLV1 all function exclusively in WUS-expressing cells in the center of the meristem, however, this observation does not address if they act together at the biochemical level to converge on similar signaling outputs. In wild type plants CLV1 represses the expression of the related BAM receptors in the center of the SAM in response to CLV3p [28]. I therefore asked if CRN and CLV2 participated in CLV1-mediated repression of BAM3 in the SAM center. For simplicity, BAM3 expression was analyzed since BAM1, BAM2 and BAM3 are all targets of CLV1 in the SAM center, but BAM1 and BAM2 display expression in the SAM epidermis and floral primorida which is clv1-independant [28]. I introgressed the previously characterized pBAM3::Ypet-N7 transgenic line [28] from Col-0 into the null alleles of clv2 and crn, and isolated homozygous transgenic lines in each mutant background. For rpk2, a CRISPR null (rpk2-cr) was generated directly in the homozygous pBAM3::Ypet-N7 wild type transgenic line, and segregated away from the Cas9 transgene for analysis (see Materials and Methods). The pBAM3::Ypet-N7 reporter generates a tandem Ypet fusion protein that is targeted to the nucleus. I then compared the expression of the BAM3 reporter in the L3-L5 cells of the SAM center (see S2 Fig for image calibration). CLV1 is expressed strongly in L3-L5 cells in wild type, clv3, and clv1-8 meristems and expression in these cells is sufficient to account for all stem cell regulation by CLV1 [25]. Consistent with previous imaging, BAM3 reporter expression was undetectable in the center of the SAM in wild type plants, but was robustly detectable in clv3 mutants and strong alleles of clv1 (Fig 2A). Similar results were found in FMs (Fig 2A), consistent with CLV3p-CLV1 repression of BAM3 expression in both meristems [28]. In contrast, BAM3 was not expressed in the center of SAMs or FMs in either clv2 or crn mutants (Fig 2A, see S2 Fig for imaging calibration). BAM3 is expressed in phloem lineage cells independent of CLV3-CLV1 signaling [28, 38]. In all plants examined, BAM3 reporter expression was observed in the phloem linage cells of the vasculature outside of the SAM, consistent with previous work [28], demonstrating that lack of signal in the SAM was not due to reporter silencing in any one line (for example see S4 Fig). Similarly, no ectopic expression of BAM3 was observed in the L3-L5 cells from SAMs or FMs of rpk2 null mutant plants Fig 2A, see Materials and Methods for construction of rpk2 null mutant) [18]. These data demonstrate that CLV1 signals to repress BAM3 in response to CLV3p independent of CLV2, CRN, and RPK2.
I sought to test this observation genetically by creating higher order receptor mutants. Owing to their repression by CLV1 in the center of the SAM, BAM receptors do not normally participate in stem cell regulation leading to an invariant two carpels per flower in bam1 bam2 bam3 triple mutants as in wild type Col-0 plants. However, when ectopically expressed in the center of FMs and SAMs in clv1-101 null mutants, BAM receptors partially compensate for the lack of CLV1 and correspondingly clv1-101 bam1 bam2 bam3 mutants greatly enhance carpel numbers of clv1-101 null mutants and display massive SAM over-proliferation during vegetative growth (Fig 2B and 2C, [28]). crn null mutants are phenotypically weaker than clv1-101 null mutants (Fig 2C). However, unlike clv1-101, crn null mutants were strictly additive with bam1 bam2 bam3 triple mutants in FMs and crn-10 bam1 bam2 bam3 plants were identical to crn alone (Fig 2C). In addition, crn-10 bam1 bam2 bam3 displayed a bam1 bam2 bam3 vegetative SAM phenotype, and lacked the unregulated SAM over-proliferation seen in clv1-101 bam1 bam2 bam3 plants (Fig 2B) [39]. This observation is consistent with CRN being dispensable for CLV1 mediated regulation of BAM expression and signaling in vivo.
I previously assessed BAM3 reporter expression in strong alleles of clv1 [28], using the clv1-8 allele in Col-0 which contains a D295N mutation implicated in CLV3p binding [11]. Strong alleles of clv1 are weakly dominant negative and the molecular basis of this remains unclear [29]. Previous work has suggested strong clv1 mutant receptors could be interfering with CRN [16], or perhaps BAM1 and BAM2 [30]. To test these possibilities I first examined BAM3 expression in the clv1-101 null allele in Col-0. BAM3 reporter expression was considerably lower in L3-L5 cells of clv1-101 null SAMs compared to both clv3 null and clv1-8 strong alleles (Fig 3 and Fig 2A, S3 Fig). In some clv1-101 null plants, BAM3 expression was nearly undetectable in the SAM center. Expression of BAM3 in FMs was more reliably detected in clv1-101 null plants, but never approached the levels seen in clv3 null and clv1-8 strong alleles (Fig 3, Fig 2A). Unlike the low levels of BAM3 expression in clv1-101 null plants, BAM3 reporter expression in clv2, crn or rpk2 was not observed (Fig 2A, S3 Fig). Therefore, while BAM3 expression is de-repressed in the center of the SAM in both null and strong clv1 alleles, the level of de-repression of BAM3 is higher in the strong clv1-8 allele. This result implies that in clv1-101 null mutants, unknown receptor(s) signaling is still effective at repressing BAM3 expression. This reduced repression is not due to co-operative receptor function with CRN, as crn clv1-101 double null mutants contained levels of the BAM3 reporter equivalent to the clv1-101 null alone (Fig 3A).
Since BAM receptors are ectopically expressed in clv1-101 null SAMs and partially compensate for clv1 they are capable at some level of signaling like CLV1 [28]. I therefore reasoned that ectopic BAM receptors might be dampening their own expression in the center of the SAM in clv1-101 null mutants. To test this I generated bam1 bam2 bam3 and clv1-101 bam1 bam2 bam3 quadruple mutant pBAM3::Ypet-N7 transgenic lines. Consistent with the previous observation that bam1 bam2 bam3 are dispensable for CLV1 function [28], BAM3 was fully repressed in the center of either SAMs or FMs in bam1 bam2 bam3 mutant plants. In contrast, the weak expression of the BAM3 reporter in the center of SAMs and FMs in clv1-101 null plants was greatly enhanced in clv1-101 bam1 bam2 bam3 quadruple mutants (Fig 4A). This result demonstrates that ectopically expressed BAM receptors in the center of clv1-101 null mutant SAMs signal to dampen their own expression. The level of BAM3 reporter de-repression in clv1-101 bam1 bam2 bam3 plants was comparable to, and occasionally stronger, than in clv1-8 alleles [28]. Previous work has suggested that clv1 missense receptors could interfere with BAM function [30] or perhaps CRN [16]. To test these hypotheses, I generated crn-10 clv1-8 double mutants and clv1-8 bam1 bam2 bam3 mutants. Consistent with the BAM3 imaging, crn-10 clv1-8 mutants were additive with respect to carpel number compared to single mutant plants (Fig 3B, S5 Fig), indicating that clv1-8 receptors do not interfere with CLV2/CRN signaling genetically. If clv1-8 receptors were interfering with the function of unknown receptors, and not BAM receptors, then clv1-8 bam1 bam2 bam3 should be as strong, or stronger, than quadruple null mutants. However, I found that clv1-8 bam1 bam2 bam3 were comparable to clv1-101 bam1 bam2 bam3 quadruple null mutants (Fig 4B). These data support the hypothesis that clv1-8, and presumably other strong missense clv1 receptors, function exclusively by interfering with the signaling of ectopic BAM receptors in the SAM. This result is consistent with the BAM3 reporter imaging demonstrating that CRN is dispensable for CLV1 signaling, and also implies that like CLV1, BAM receptors signal in vivo in a CLV2/CRN independent manner.
CLV1 signaling strongly represses BAM expression, but has a considerably weaker effect on WUS in physiological ranges of CLV3p [11, 28, 33, 34]. The current model in the field postulates that upregulation of WUS drives the excess cell proliferation in clv class meristems. I sought to test if WUS up-regulation was causally connected to BAM de-repression in the SAM of clv1 mutants. To do this I repeated the experiments in Schoof et al [31] and used the CLV1 promoter to express WUS in the SAM of Col-0 pBAM3::Ypet-N7 plants. In those experiments, ectopic expression of WUS using a two-component inducible system drove SAM enlargement and resulted in flowers with extra carpels, phenocopying clv1 mutants. Of the 36 T1 CLV1p::WUS lines generated in the Col pBAM3::Ypet-N7 background, only 11 plants displayed increases in SAM size and stem fasciation. Despite having enlarged SAMs reminiscent of clv1 SAMs (Fig 5B), no de-repression of BAM3 was observed in the SAM or FMs of any of the enlarged meristem plants examined (N = 10, Fig 5A). This indicates that WUS-induced over-proliferation of the stem cell niche is genetically separable from BAM3 transcriptional regulation in the CLV1 pathway.
Despite being cloned nearly 20 years ago, we have little understanding about of how CLV1 signals in planta or its relationship to other proposed CLV3p receptors. Here I demonstrate, using BAM3 repression as a readout, that CLV1 signals to control at least some transcriptional outputs independent of CLV2/CRN and RPK2 but fully dependent on CLV3 (Fig 6). CLV2/CRN have no effect on BAM3 reporter expression by themselves, or in combination with CLV1, and genetically do not participate in BAM feedback compensation. As such, despite the different receptor mutants having similar qualitative stem cell defects and resistance to CLV3p, CLV1, RPK2 and CLV2/CRN are functionally separate and converge on distinct signaling outputs in vivo. This result implies that CLV2/CRN are dispensable for CLV3p mediated perception by CLV1, and that putative CLV2/CRN/CLV1 complexes seen in tobacco overexpression studies are dispensable for CLV1 signaling in vivo. This is consistent with previous data showing that CLV1 traffics from the plasma membrane to the lytic vacuole in response to CLV3p in a CLV2-independent manner [10], and consistent with additive genetic interactions with clv2, crn and clv1 [16]. As such, every readout for CLV1 function would suggest that CLV2/CRN are dispensable for CLV3p perception and signaling in vivo. Previously it was suggested that strong clv1 receptors could interfere with CLV2/CRN function in vivo [16], however I found that there is a significant enhancement of the strong clv1-8 allele in clv1-8 crn-10 double mutants. Genetic analysis and BAM3 repression analysis suggests that the strength of the clv1-8 mutant receptor can be accounted for solely by interfering with ectopic BAM receptors in SAM, supporting previous studies [30].
Despite their independence from CLV1, clv2 and crn mutants are resistant to ectopic CLE peptide-mediated SAM or RAM termination in several species of plants, and have not been identified in genetic screens for other peptide responses to date. This suggests that there is either a tight relationship between CLV2/CRN and CLV3/CLE ligand function, or that CLV2/CRN impact a developmental process which superficially resembles CLV3p function. Interestingly, while CLV2 and CRN mutants display resistance to CLV3/CLE peptide induced root termination [16, 40], they do not display conspicuous root growth or patterning defects in the absence of peptide ligand [16]. I previously demonstrated that mutants that do not function in the CLV3 pathway, but have a higher rates of SAM cell proliferation and an expanded SAM, display resistance to CLE peptide mediated SAM termination [28]. This suggests that CLE peptide resistance per se is insufficient to determine gene function in CLV3p signaling or perception. Based on this, it is formally possible that CLV2/CRN do not function in CLE mediated perception in vivo, as has been suggested by peptide binding assays [23]. The function of CLV2/CRN in CLE mediated signaling, if any, remains enigmatic, but current data demonstrates that CLV1 ligand perception and binding, stability, endomembrane trafficking, and signaling are all independent of CLV2/CRN in planta.
clv1 null mutants are phenotypically weak, despite having ectopic BAM expression in the SAM center. However, like CLV1, BAM receptors also repress BAM expression in the center of the SAM (Fig 6). As such, clv1 null mutant SAMs contain low levels of BAM expression, potentially explaining why ectopic BAM expression is not sufficient to fully compensate for clv1. In the organizing center, CLV1 is proposed to also repress WUS expression. Despite this, CLV1 and WUS expression is largely co-incident and expression of either CLV1 or CRN/CLV2 in WUS-expressing cells is sufficient to account for all functions in stem cell regulation. WUS induced SAM over-proliferation can be uncoupled from CLV1-mediated BAM repression. As such, clv1 and clv3 SAMs are functionally different from WUS-induced clv-like SAMs at some level. WUS itself has been proposed to repress CLV1 [33], but the significance of this is unclear due to the co-incident expression patterns of both genes, the lack of de-repression of BAM3 in enlarged SAMs ectopically expressing WUS, and the fact that uncoupling CLV1 expression from the native CLV1 promoter in SAM cells has no phenotypic consequence other than to complement clv1 null mutants [10, 28–30, 39]. In addition, CLV1 mediated repression of WUS is quantitatively different from that for BAM receptors. In the center of wild type SAMs BAM gene expression is nearly undetectable and becomes robustly detectable in clv1 or clv3 mutants. In contrast, WUS is robustly detected in wild type SAMs [28, 31]. Overexpression of CLV3 represses WUS, however plants expressing up to 300 fold more CLV3 are wild type in appearance [34]. Thus, at endogenous levels CLV3p strongly suppresses BAM expression in a CLV1-dependent manner, but has less of an effect on WUS, which requires considerably higher and potentially non-physiological levels of CLV3p for full repression. Understanding the regulation of BAM expression by CLV1 could lead to new insights into this signaling pathway.
Plant growth and transgenic plant selection was performed as described in [28]. All clv1, bam1, bam2, bam3 and clv2 alleles are in the isogenic Col-0 background and have been characterized previously. Genotyping of plants from crosses were performed using appropriate primers selecting for mutant alleles or T-DNA insertions [28, 35]. Carpel counts were performed as described with the exception of comparisons using crn-10 or clv2. In these lines early termination of the SAM and the cessation of flower production was observed after 5 flowers on average as noted before for clv2 mutations in the Col-0 background [35]. Therefore, I counted flowers 6 and beyond for all comparisons using clv2 or crn-10 alleles for all genotypes in those experiments. The transient termination phenotype of crn-10 was not altered by mutations in clv1, clv3 or in bam1 bam2 bam3 triple null mutants. In bam1 bam2 bam3 clv1 plants stems elongation is highly distorted as is SAM production as described in [28], making inferences about floral primordia order difficult. As such, all flowers were counted for comparisons between quadruple mutants.
Confocal imaging was performed using an inverted Zeiss 710 confocal. Briefly, an inverter adaptor (LSM Tech, Etters, PA, USA) was used to allow upright imaging of shoot meristems when attached to the Zeiss 710. Details on the configuration of the inverter are available on request and will be published elsewhere. Meristem staging, dissection and mounting were performed as described in [28], with each presented photo being a mean of eight scan cycles. For each experiment a minimum of 6 meristems were imaged and all imaging experiments were repeated with different plant populations two to four times. Imaging settings for the Ypet channel were kept constant across all experiments, but the gain on the red channel for propidium iodide was altered to account for staining differences as necessary. Image settings were calibrated to capture the dynamic linear range in most plants. At these image settings YFP signal in clv1 bam1 bam2 bam3 quadruple mutant SAMs was occasionally saturating in some nuclei (S2 Fig) but no signal was detectable in wild type, clv2, crn or bam1 bam2 bam3 SAM at these settings (S2 Fig, Figs 2–5).
The CRN promoter binary vectors used were described in [17]. For the CLV2 promoter binary, 1250 bp and 588 bp of the 5' and 3' promoter and UTR regions were amplified from Col-0 and fused together using recombinant PCR to introduce a unique BamH1 site and cloned into pBJ shuttle vector. A Gateway cloning cassette was inserted into the BamH1 site and the entire promoter cassette was transferred into the pMOA33 binary vector as a Not1 fragment [41]. Col-0 CLV2 CDS was amplified using primers that allowed cloning into the pENTRD (Invitrogen) vector and fused with an in frame MYC epitope to the C-terminus. This was then recombined into either pMOA33 CLV2p or pMOA33 WUSp. The pMOA33 WUSp vector was described as in [28]. For the generation of pMOA33 CLV1p::WUS, a pENTRD WUS CDS clone was recombined into the pMOA33 CLV1p vector used previously [10, 28].
The pCUT vector system was used to generate the crn-10 allele and rpk2-cr allele used here [36]. Briefly, this vector series co-expresses nuclear targeted Cas9 from the UBQUITIN10 promoter and gRNAs from the U6 promoter. A 20 base pair gRNA target site (bold), including upstream G and downstream PAM (underlined) was selected gaagcaaagaagaagaagaaatgg near the initiator methionine codon in the CRN genomic signal sequence encoding region. This target site was used to generate a gRNA that was cloned into pCUT3 as described in [36]. Kanamycin resistant plants were selected in the T1. In a couple of T1 plants, branches were observed with flowers which all contained elevated levels of carpels relative to wildtype. Seeds were collected specifically from one of these branches and sequencing in the next generation revealed that these branches arose as somatic bi-allelic sectors containing either an A or T insertion (bold) at the same location downstream of the initiator ATG (uppercase) (ATGaagcaaagaagaagaagTaaatgg) leading to equivalent truncated CRN proteins with only 7 amino acids remaining in the signal sequence of CRN (MKQRRRRKWMstop). Plants with the homozygous T insertion event were selected as this provides resistance to Hph1 digestion using the dCAPs primers CRN-10 F gtagaagcagcaatgaagcaaagaagaaggtg and CRN-10 R gttgaagttgtggataagtg [42], and segregated away from the pCUT3 transgene. Complementation analysis was performed using vectors described in [17]. For rpk2-cr mutant creation, a tandem array of two different U6 promoter gRNA cassettes targeting the RPK2 CDS were gene synthesized by Invitrogen and cloned into pCUT3 as described in [36]. The RPK2 gRNA target sites chosen were aagattactgctcctggtttgg and tcatggctcttaacattagtgg, with the PAM sequence in bold. This vector was transformed directly into the Col-0 pBAM3::Ypet-N7 line. In the T2 generation from a select T1 line, multiple plants displaying an rpk2 phenotype were identified based on male sterility and extra carpels [18, 43]. Imaging was performed on 10 plants with an rpk2 phenotype. One line lacking the pCUT3 vector was identified by PCR and backcrossed to Col-0 to maintain as a heterozygous owing to the male sterility of rpk2 mutants [43]. This line was then imaged again in the next generation to confirm BAM3 expression patterns. This line, termed rpk2-cr, contains a +A insertion between nucleotides 229 and 230 and an additional +T insertion between nucleotides 279 and 290 in the RPK2 CDS. This results in the production of an RPK2 protein that contains a stop codon directly after amino acid 76 (serine76) downstream of the signal sequence resulting in the deletion of all LRR repeats and downstream transmembrane, juxtamembrane and kinase domains [43].
|
10.1371/journal.pntd.0004814 | Community Knowledge, Perceptions, and Practices Associated with Urogenital Schistosomiasis among School-Aged Children in Zanzibar, United Republic of Tanzania | On the Zanzibar islands, United Republic of Tanzania, elimination of urogenital schistosomiasis is strived for in the coming years. This qualitative study aimed to better understand community knowledge, perceptions, and practices associated with schistosomiasis among school-aged children on Unguja and Pemba islands, in order to inform the development of behavior change interventions contributing to eliminate urogenital schistosomiasis.
In 2011, we conducted 35 children’s discussion groups, 41 in-depth interviews with parents and teachers, and 5 focus group discussions with community members in Zanzibar. Using a modified-grounded theory approach, we transcribed and coded the narrative data followed by thematic analysis of the emergent themes.
Urogenital schistosomiasis is a common experience among children in Zanzibar and typically considered a boys’ disease. Children engage in multiple high-risk behaviors for acquiring schistosomiasis because of poor knowledge on disease transmission, lack of understanding on severity of disease-associated consequences, and lack of alternative options for water related activities of daily living and recreational play. Local primary school teachers had little to no training about the disease and no teaching tools or materials for students.
Conducting activities in open natural freshwater contaminated by S. haematobium larvae compromises the health of school-aged children in Zanzibar. The perception of urogenital schistosomiasis as a minor illness rather than a serious threat to a child’s well-being contributes to the spread of disease. Understanding community perceptions of disease along with the barriers and facilitators to risk reduction behaviors among children can inform health promotion activities, campaigns, and programs for the prevention, control, and elimination of urogenital schistosomiasis in Zanzibar.
| On the Zanzibar islands, United Republic of Tanzania, elimination of urogenital schistosomiasis, a disease caused by infection with a blood fluke (Schistosoma haematobium), locally known as kichocho, is strived for in the coming years. This study used qualitative research methods of focus groups and in-depth interviews with adults, and group discussions with school-aged children to explore (i) knowledge and perceptions of kichocho transmission, (ii) specific behaviors among children in Zanzibar that put them at risk for acquiring infections with the kichocho parasite, (iii) symptoms of kichocho along with personal health-seeking behaviors and treatment strategies, and finally (iv) ideas for preventing kichocho in children. We found that there was little available formal education about disease transmission, which contributed to myths and misperceptions about routes of transmissions, causes, and treatment of kichocho. School-aged children regularly exposed themselves to contaminated natural, open freshwater bodies through recreational and domestic activities of daily life. Kichocho was often wrongly viewed as an infection with an intestinal worm of little significance, rarely associated with severe health consequences, and with little to no disease stigma. Local primary school teachers had little to no training about the disease and no teaching tools or materials for students. The findings add valuable insights into how current knowledge, perceptions, and practices impede optimal disease prevention and control and highlight the necessity for a community tailored behavioral intervention to interrupt transmission of urogenital schistosomiasis.
| Schistosomiasis is a debilitating disease that affects poor and deprived population groups, especially in rural Africa [1]. The global impact is enormous with more than 200 million people infected with blood flukes of the genus Schistosoma [2]. Urogenital schistosomiasis, caused by S. haematobium, can include acute illness such as blood in urine (hematuria) and anemia in children, while fibrosis of the bladder and ureter, and kidney damage can occur as infections persist [3,4]. Bladder cancer can be a complication in adults and female genital schistosomiasis may be associated with increased risk of human immunodeficiency virus (HIV) infection [5–9].
Over the past decades, programs to reduce the morbidity caused by schistosomiasis have been implemented in many endemic countries and the number of people who received treatment with praziquantel has increased annually [10]. In order to eliminate schistosomiasis as a public health problem and to interrupt transmission in areas where morbidity control has been achieved, the World Health Organization (WHO) and other institutions and stakeholders are advocating the intensified use of integrated schistosomiasis control approaches [7,11–14]. Health education and interventions based on social and behavioral science triggering behavior change, in addition to regular preventive chemotherapy with praziquantel, are likely to become a key component of future elimination efforts [12,13,15].
Behavior change in humans requires close interaction with the at-risk population [16]. To change risk behaviors, enhance the interruption of disease transmission, and finally eliminate schistosomiasis particularly in children, innovative, community-tailored approaches are needed. Understanding the community is critical to creating effective behavioral interventions promoting the adoption of protective behaviors and reducing risk behaviors [16,17]. Community participatory processes are fundamental to understanding the community’s current knowledge, perceptions, attitudes, and behaviors as well as motivators and barriers to behavior change [18]. Community participation can also create ownership of public health initiatives, which is often viewed as fundamental for the success of population-based health outcomes [16,18]. People are experts about the communities they live in and have many different ways of knowing and gathering information.
On the Zanzibar islands, United Republic of Tanzania, elimination of urogenital schistosomiasis is strived for in the coming years. In the frame of a randomized operational research trial implemented on Unguja and Pemba islands from 2012 till 2017, the impact of behavior change interventions in addition to biannual praziquantel treatment on the prevalence and intensity of S. haematobium infections is assessed [19–21]. Here we present results of the qualitative formative research that was implemented in 2011 to inform, along with future participatory community co-design workshops, the development of a community-tailored behavioral change intervention that might help to eliminate urogenital schistosomiasis in Zanzibar [19].
In 2011, the National Centre for Emerging Zoonotic Diseases (NCEZID) of the Centers for Disease Control and Prevention (CDC) received and approved the qualitative formative research protocol for Zanzibar (NCEZID Tracking Number: 103111BP) to go forward to the CDC Human Research Protection Office (HRPO) and Institutional Review Board (IRB) for review. The HRPO and IRB determined that the project activities were exempt under regulation 45 CFR 46.101(b)(2) and issued a written waiver. The full study protocol of the “Study and implementation of schistosomiasis elimination in Zanzibar (Unguja and Pemba islands) using an integrated approach” received additional ethical approval from the Zanzibar Medical Research Ethics Committee in Zanzibar, United republic of Tanzania (reference no. ZAMREC 0003/Sept/011), the “Ethikkommission beider Basel” (EKBB) in Basel, Switzerland (reference no. 236/11) and the IRB of the University of Georgia, in Athens, Georgia, United States of America (project no. 2012-10138-0). The study is registered at the International Standard Randomized Controlled Trial Number Register (ISRCTN48837681). The data collection was conducted with support from the CDC in Atlanta, Georgia, United States of America. The CDC HRPO and IRB approved the informed consent process conducted with all participants, who took part in student group discussions and interviews. Due to a limited ability of participants to read and write the informed consent was available in both English and Kiswahili, the local language, and read aloud by trained bilingual research staff. Participants provided a verbal consent, with the consent acknowledged with the signature on the informed consent document of a witness present at the time [22,23]. Research staff reviewed the consent procedure and all consent forms to ensure compliance with the process. In case of children below the age of 18 years, their parents or legal guardians provided written informed consent for their participation.
This qualitative inquiry was conducted including school-aged children, parents, teachers, and community leaders from seven small administrative areas, called shehias, on the islands of Unguja and Pemba from July till September 2011. The islands of Unguja and Pemba have an estimated combined population of around 1.3 million people and the main industries are spices, raffia, and tourism [24,25]. More than 99 percent of Zanzibar's population is Muslim. Urogenital schistosomiasis constituted a considerable public health problem on both islands in the past century [26–28], but regular treatment of the at risk population with praziquantel reduced S. haematobium prevalences and intensities [29–31]. In 2012, the baseline survey of the “Study and implementation of schistosomiasis elimination in Zanzibar (Unguja and Pemba islands) using an integrated approach” revealed an overall prevalence of 4.3% and 8.9% in schoolchildren from Unguja and Pemba, respectively [20].
The field team consisted of a senior social scientist from the CDC and seven Kiswahili and English speaking research assistants from the Ministry of Health, Department of Neglected Tropical Diseases, and the Ministry of Education, Department of Health Education, in Unguja, and three research assistants from the Public Health Laboratory—Ivo de Carneri in Pemba. The local team had little or no previous experience with the application of qualitative research methods. Research assistants were trained in research ethics and qualitative data collection methods by the senior social scientist and served as the primary data collectors and logistic coordinators setting up focus groups and interviews within the communities.
We used purposive sampling to recruit a homogeneous study sample of school-aged children, who might engage in risky behaviors. Such risk behaviors include swimming, fishing, bathing, washing clothes, or performing other domestic chores in ponds, lakes, streams, and rivers that are potentially contaminated with S. haematobium larvae. For this initial informative research study it was decided that the easiest way to reach school-aged children was through government supported public primary schools. Schools in the selected shehias were identified with the assistance of the staff from the Ministry of Health and from the Ministry of Education. Students in grades Standard 1 to Standard 7 were recruited through the headmaster of each school. Individual teachers and parents were also recruited through public primary schools and community leaders were recruited through local social networks.
The qualitative inquiry was conducted through 35 children’s group discussions (GD), 5 focus group discussions (FGD) with community leaders, and 41 in-depth interviews (ID) with teachers and parents. The children’s GD included 6–8 children of the same sex, facilitated by a Kiswahili speaking research assistant using a simple topic guide to lead the discussion (S1 Topic Guide). Children were provided paper and crayons and were first asked to draw anything they wanted to draw, followed by a discussion of their picture. Then they were asked to, “Please draw me a picture about everything you know about kichocho (the local term for schistosomiasis).” These drawings encouraged more robust discussions [32]. Students drew pictures and then described their drawings about the disease, risk behaviors, and prevention ideas. A note taker managed an audio recorder and took written notes in support of the group facilitator. FGD with community leaders followed a similar format without the drawings. Individual IDIs were conducted with parents and teachers using the audio recorders without the note takers. Examples of qualitative open-ended questions used with the adults are shown in Fig 1. The study was conducted in five shehias on Unguja (Chaani, Dole, Kilombero, Mwera, and Uzini) and in two shehias on Pemba (Chambani and Kizimbani). The shehias were selected among the 15 behavioral study shehias on each island [19], based on their location on the island and previous knowledge about urogenital schistosomiasis in the area [29,31,33,34]. Data were collected until saturation was obtained [35,36].
All data collection tools including project overview, informed consents, and topic guides were translated from English into Kiswahili, pretested, and modified to adapt to local linguistic and cultural nuances by the research team [37]. Topic guides explored i) knowledge and perceptions of schistosomiasis transmission, ii) specific behaviors among children in Zanzibar that put them at risk for acquiring S. haematobium infections, iii) symptoms of schistosomiasis along with personal health-seeking behaviors and treatment strategies, along with iv) ideas to prevent schistosomiasis in children. Additionally, we explored other cultural factors, gender-roles, influential communication channels, and decision-making processes [37]. Additional probes allowed for deeper exploration of the topics that emerged, supporting additional areas of interest. Following group discussions and interviews, participants received a small thank you gift for their time and participation. Group discussions conducted in schools and FGD conducted in community settings were approximately 1.5 hours in length, IDIs took approximately 45–60 minutes. Data were collected using audio recorders (Olympus 70, Olympus Corporation, Tokyo, Japan). Additional field notes were handwritten during the interviews and reviewed during debriefing sessions to verify accuracy of the interview session [38]. We did not collect personal identifiers.
Due to the paucity of behavioral information on urogenital schistosomiasis in Zanzibar we chose a qualitative approach to better understand cultural practices associated with activities of daily living linked with contact with local natural open freshwater bodies. In this study, we examined barriers to schistosomiasis prevention and control related to urination practices of children along with recommendations for improving such practices and reducing disease threats. We used a modified grounded theory approach with an emergent qualitative thematic analysis allowing the hypothesis to be generated from the data [39].
Narrative data were transcribed into English with review following translation to ensure accurate translation and local meanings. Transcripts were entered into Atlas-ti (ATLAS.ti Scientific Software Development GmbH, Berlin, Germany) as a Word document (Microsoft Corporation, Redmond, WA, USA) to facilitate text searching, data coding and analysis. Data analysis began with the first discussions and interviews allowing for emerging, unexpected, and/or inconsistent issues to be explored in subsequent discussions and interviews [36,39]. The coding structure evolved inductively with the codes from the narrative data of earlier interviews informing subsequent coding of the following interviews supplemented with field notes from the interviewer and note taker [32,38–40]. Due to time constraints and ongoing data collection tasks, the primary author, a social scientist experienced in qualitative research, was the primary data coder with verification of interpretive codes by the research assistants. Open, axial, and selective coding was used to analyze the GD, FGD, and IDI narratives [36,38–41]. A coding frame was developed through open coding, a word-by-word analysis used to identify, name, and categorize explanations and descriptions of the day-to-day reality of participants as related to schistosomiasis. Consensus on the coding frame was obtained through discussions with the local qualitative research assistants, who were from Zanzibar and conducted the original interviews. Axial coding, the process of relating codes to each other, via a combination of inductive and deductive thinking, was used for analysis of specific emergent themes, across themes, and for the relationships between themes [40,41]. Over the course of data collection, emergent themes became redundant, suggesting that all major themes had been identified and saturation reached [42]. An analysis matrix served as a framework for the resulting findings. Narrative excerpts from an analysis framework matrix are shown in Table 1.
The trustworthiness of our data was derived from standardization of methods and documentation for auditability, triangulation of the data, and verification of data findings with local staff members [43]. A standardized implementation document guided the training and implementation of the qualitative methodology with all procedures, topic guides, informed consents, timelines, interview schedules, data collection strategies, data management, and analysis strategies written out [38,43]. Triangulation of data was derived through the multiple data collection methods (GD, FGD, and IDI), multiple perspectives (younger and older girls and boys as well as adult women and men), and multiple venues (school-based, private home-based, and public venues). Findings were verified amongst local staff of the Ministry of Health in Unguja and the Public Health Laboratory Ivo-de-Carneri in Pemba by corroborating results with similar findings across other settings [44,45].
As shown in Table 2, narrative data were collected from 27 children’s GD, 5 FGD with community leaders, and IDI with 21 teachers and 16 parents on Unguja Island. We also conducted 8 children’s GD and 4 IDI with teachers on Pemba Island to verify similarities among children on the two islands. Group discussions included boys and girls from both islands, who were 6–17 years old and attended grades of Standard 1–7. FGD included 16 male and 14 female community leaders, aged 23–72 years. They were teachers, farmers, leaders of women’s groups, religious leaders, school coaches, religious schoolteachers, petty traders, and small business owners, as well as housewives and traditional leaders. In-depths interviews were conducted with 2 male and 19 female teachers from Unguja and 2 female and 2 male teachers from Pemba. Teachers were 26 to 56 years of age. Eight fathers and eight mothers from Unguja were interviewed. Parents had a median age of 41years with a range of 24 to 72 years of age.
Qualitative research, alone or in mixed methods, has been used to better understand the experiences of people affected or at risk for numerous neglected tropical diseases such chagas disease, filariasis, and schistosomiasis [46–48]. Results of our informative in-depth, qualitative investigation of schistosomiasis among school-aged children suggested that despite previous initiatives related to urogenital schistosomiasis control and prevention in Zanzibar [29,44,49], people’s knowledge about disease symptoms, transmission, and prevention were poor. Our findings identified several barriers to optimal disease prevention and control.
First, we observed that school-aged children regularly exposed themselves to contaminated natural, open freshwater bodies through recreational and domestic activities of daily living with little knowledge about routes of schistosomiasis transmission, which is in line with findings from previous studies in Zanzibar, Tanzania, Zimbabwe, and Western Kenya [50–52].
Second, S. haematobium infection was often viewed as an infection with an intestinal worm of little significance, not typically associated with severe health consequences, and little to no disease stigma. This is in contrast to reports from previous research in Nigeria, where individuals with schistosomiasis disease were stigmatized by others [53]. The Health Belief Model posits that perceived seriousness along with perceived susceptibility, perceived benefits, and perceived barriers are critical constructs used to explain and influence changes in health behaviors [54,55]. It also specifies that if individuals perceive a negative health outcome to be severe and perceive themselves to be susceptible to those negative outcomes, they are more likely to adopt positive protective behaviors [17,54,55]. Drawing upon the constructs of this behavioral theory supports shifting the context of schistosomiasis to that of a blood fluke, with serious health consequences such as bladder cancer and infertility, rather than the current perception of a less severe “worm.” Elaborating on the perceived seriousness of the infection, whether through medical information or increased awareness of the serious effects of the disease on a person’s life, is critical to address in a behavioral intervention [17,54,56]. There is evidence that theory-based, behavioral interventions can increase effectiveness among a variety of public health issues [57–60]. Synthesis of behavioral intervention research and non-regulatory interventions most often advocates the application of behavioral theory as an integral step in intervention design and evaluation [55,61].
Third, many people described abdominal pain, blood in the urine (hematuria), pain or burning during urination (dysuria), and commonly genital itching as symptoms of infection. However, as observed in studies conducted elsewhere in sub-Saharan Africa [62], these symptoms were also perceived as sexually transmitted infections that indeed may appear similar to symptoms of urogenital schistosomiasis. A person with a sexually transmitted infection may be reluctant to seek treatment due to shame and stigma [53,62]. Therefore, correcting the misperception that schistosomiasis is a sexually transmitted disease, while at the same time supporting the need to seek treatment for any and all similar symptoms, could be an important component of a schistosomiasis educational campaign to improve treatment seeking.
Fourth, first line treatment for a few people in Zanzibar, similar to mainland Tanzania [62], was often home remedies and occasional use of locally available herbalists followed by more conventional treatments when those earlier ones had failed. Lack of decentralized, locally available drugs and cost of transportation were also identified as barriers to seeking more conventional drug treatments. The decentralization of drug treatment to the local level as well as increasing knowledge about free drug treatment accessible through mass drug administration campaigns could improve treatment seeking among infected individuals. Further research into understanding any underlying barriers to treatment seeking behaviors should be explored [63].
Fifth, little available formal education about disease transmission contributed to myths and misperceptions about routes of transmissions, causes, and severity of disease, treatment, and ultimately prevention of disease. Schoolteachers and Koran school (Madrassa) teachers, viewed as influential people in children’s lives lacked formal scientific training, teaching materials, and other resources to be able to educate students about schistosomiasis. Teachers reported a need for a teacher’s training with a standardized, detailed syllabus to teach children about schistosomiasis during school sessions. Trainings could be set up similar to the Lushoto Enhanced Health Education Project that introduced interactive teaching methods into mainland Tanzanian study schools and demonstrated a feasible and effective intervention capable of changing schistosomiasis knowledge and health seeking behaviors among children [64]. The inclusion of religious teachers as change agents could maximize exposure of a schistosomiasis educational program to a broader community because they often engage children who may not attend government schools. Trained school and religious teachers could instill a perception of perceived seriousness of disease as well as perceived susceptibility of disease among children engaging in risky behaviors. Teachers could also identify and address the barriers to change and promote perceived benefits of reducing risky behavior to children.
Educating through schools could encourage students to act as change agents through peer education, role modeling, and shifting social norms of acceptable behavior [65,66]. Peer education, defined as “the teaching or sharing of health information, values and behaviors by members of similar age or status,” is widely used in the field of health promotion and education recently, such as the prevention of HIV/acquired immune deficiency syndrome (AIDS), smoking, and alcohol and drug use [67–71]. Peer education is focused on sharing information and experiences along with trust between the people in the similar context and learning from each other. Peer education, has been noted as a feasible method for transferring schistosomiasis knowledge from students to parents [65,66].
Sixth, most adults, and some children recognized the difficulty of extinguishing the behavior of urinating in the ponds and streams. It was seen as a private behavior and often associated with urgent need. Children and adults described educational, behavioral, and structural interventions to prevent kichocho in children. Community members often described the need for the community to work together to prevent kichocho in children suggesting the importance of a participatory approach to intervention development and implementation.
Previous research has noted that top down approaches to community interventions have been perceived by some community members as not in their best interest or being a poor fit for the socio-cultural context within the community [72–74]. The lack of attention to an individual’s social, cultural, religious, environmental, and physical context often results in a poor understanding of why an intervention is valuable and ultimately in an inadequate adoption of the desired positive behaviors and practices by community members [72–74]. This may explain why despite years of community administered preventive chemotherapy, the perception of schistosomiasis in Zanzibar was that of a commonplace, minor illness, rather than a serious threat to a child’s wellbeing. Administering preventive chemotherapy without addressing the local circumstances of community members with tailored communication and educational efforts can lead to not only misunderstandings but also to potentially poor treatment compliance [75,76]. Understanding community perceptions along with the social, religious, economic and environmental context of schistosomiasis risk and risk reduction behaviors among children can inform behavior change interventions that are relevant and provide meaning to the vulnerable populations in Zanzibar [74,77,78]. A recent evaluation of a comic-strip medical booklet Juma na Kichocho associated with a schistosomiasis health education campaign in 16 primary schools in Zanzibar reported disappointing findings [44,77]. The authors recognized that changing the behaviors of children could not be done by an isolated school curriculum but needed to consider their everyday realities of daily living [77]. The information garnered from our qualitative inquiry will allow for the ideas and problem solving solutions of community members to be incorporated into a behavioral intervention that is germane to others in their communities. Increasingly, there is a commitment to bringing a community perspective into research and implementation of interventions along with a growing body of evidence that public health and health-promotion interventions based on social and behavioral science theories are more effective than those without a theoretical foundation [55,79,80]. Drawing upon the constructs of perceived seriousness, perceived susceptibility, perceived benefits, and perceived barriers from the Health Belief Model complemented by a social ecological model that addresses multiple levels of the community could provide a functional framework for designing, implementing, and evaluating a health promotion program for the prevention and control of schistosomiasis tailored to the context of community members, particularly school-aged children [17,55,74,79,81].
There were several limitations to this inquiry. Given that we used a purposive, convenience sample, the findings may not be representative of all members of the communities in which the inquiry took place, and results are not generalizable. The triangulation of data suggests that there were similarities across behaviors of school-aged children attending the government primary schools on both islands where we conducted the student discussion groups. The behaviors we assessed appeared to be generally practiced among children across the shehias on Unguja and Pemba and the lessons learned could be used to tailor messages for future schistosomiasis control programs for primary school aged children. That being said, these findings are not generalizable to children attending private schools or not attending school at all. Further investigation is needed to explore the schistosomiasis knowledge, attitudes, perceptions, and practices of students in private schools and of students who do not attend school to assess if they are similar to those from children who participated in our student discussion groups. There may have been information bias during IDI, GD, and FGD as interview subjects may have provided answers that they believed the interviewers expected or wanted to hear. Additionally, bias may have been introduced due to only having a single coder, even though data interpretations and language translations were substantiated with local research assistants.
Conducting recreational and domestic activities of daily living in water contaminated with S. haematobium larvae compromises the health of school-aged children in Zanzibar. An important objective of this study was to facilitate improved design of an educational and control program. Urogenital schistosomiasis, characterized as a minor illness typically of boys, along with the lack of formal school-based and community-wide education about disease transmission, symptoms, and treatment can contribute to undiagnosed disease and a lack of treatment among both girls and boys. Understanding community perceptions of disease along with the barriers and facilitators to risk reduction behaviors among children can inform behavior change activities and health promotion programs augmented with chemotherapies for an integrated approach in support of the prevention, control, and elimination of urogenital schistosomiasis in Zanzibar and elsewhere.
|
10.1371/journal.pgen.1000228 | Condensin II Resolves Chromosomal Associations to Enable Anaphase I Segregation in Drosophila Male Meiosis | Several meiotic processes ensure faithful chromosome segregation to create haploid gametes. Errors to any one of these processes can lead to zygotic aneuploidy with the potential for developmental abnormalities. During prophase I of Drosophila male meiosis, each bivalent condenses and becomes sequestered into discrete chromosome territories. Here, we demonstrate that two predicted condensin II subunits, Cap-H2 and Cap-D3, are required to promote territory formation. In mutants of either subunit, territory formation fails and chromatin is dispersed throughout the nucleus. Anaphase I is also abnormal in Cap-H2 mutants as chromatin bridges are found between segregating heterologous and homologous chromosomes. Aneuploid sperm may be generated from these defects as they occur at an elevated frequency and are genotypically consistent with anaphase I segregation defects. We propose that condensin II–mediated prophase I territory formation prevents and/or resolves heterologous chromosomal associations to alleviate their potential interference in anaphase I segregation. Furthermore, condensin II–catalyzed prophase I chromosome condensation may be necessary to resolve associations between paired homologous chromosomes of each bivalent. These persistent chromosome associations likely consist of DNA entanglements, but may be more specific as anaphase I bridging was rescued by mutations in the homolog conjunction factor teflon. We propose that the consequence of condensin II mutations is a failure to resolve heterologous and homologous associations mediated by entangled DNA and/or homolog conjunction factors. Furthermore, persistence of homologous and heterologous interchromosomal associations lead to anaphase I chromatin bridging and the generation of aneuploid gametes.
| Some of the processes that ensure proper chromosome segregation take place upon the chromosomes themselves. The chromosomes of Drosophila males undergo an interesting and relatively enigmatic step before entering meiosis, where each paired homologous chromosome becomes clustered into a discrete region of the nucleus. In this article, we provide evidence that improper chromosomal associations are resolved and/or prevented during this “chromosome territory” formation. This was uncovered through the study of flies mutant for Cap-H2, which have abnormal territory formation and improper chromosomal associations that persist into segregation. Another important process that chromosomes undergo in meiosis is the pairing and physical linking of maternal and paternal homologs to one another. Linkages between homologs are essential to ensure their proper segregation to daughter cells. In contrast to meiosis in most organisms, linkages between homologs in male Drosophila are not recombination mediated. Here, we provide evidence that Cap-H2 may function to remove Drosophila male specific linkages between homologous chromosomes prior to anaphase I segregation. When chromosomal associations persist during segregation of Cap-H2 mutants, the chromosomes do not detach from one another and chromatin is bridged between daughter nuclei. The likely outcome of this defect is the production of aneuploid sperm.
| There are several critical steps that chromosomes must undergo as they transition from their diffuse interphase state to mobile units that can be faithfully transmitted to daughter cells. In the germline, faulty segregation leading to the creation of aneuploid gametes is likely a leading cause of genetic disease, miscarriages, and infertility in humans [1].
Some steps that promote proper segregation are universal to all cell types undergoing cell division. Chromosomal “individualization” is necessary to remove DNA entanglements that likely become introduced naturally through movements of the threadlike interphase chromatin [2]. Topoisomerase II (top2) contributes to individualization with its ability to pass chromosomes through one another by creating and resealing double strand breaks [3]. The necessity of top2's “decatenation” activity to chromosome individualization becomes clear from fission yeast top2 mutants and vertebrate cells treated with a top2 inhibitor, where mitotic chromosomes appear associated through DNA threads [4],[5]. Another step that occurs prior to chromosome segregation is chromosome “condensation,” entailing the longitudinal shortening from the threadlike interphase state into the rod like mitotic chromosome [2]. Condensation is necessary due to the great linear length of interphase chromosomes that would be impossible to completely transmit to daughter cells.
Because chromosome individualization and condensation appear to occur concurrently, it has been inferred that both are promoted by the same catalytic activity. In support of this idea, the condensin complexes have been implicated in chromosome individualization [6] and condensation [7], suggesting a molecular coupling of both processes. The condensin I and II complexes are thought to be conserved throughout metazoa, each utilizing ATPases SMC2 and SMC4, but carrying different non-SMC subunits Cap-H, Cap-G, Cap-D2 or Cap-H2, Cap-G2, and Cap-D3, respectively [7]–[9]. In vitro, condensin I is known to induce and trap positive supercoils into a circular DNA template [10]–[12]. Current models to explain condensin I chromosome condensation highlight this activity as supercoiling may promote chromatin gathering into domains that can then be assembled into a higher order structure [13]. Condensin complexes may also promote condensation and individualization through cooperating with other factors, such as chromatin-modifying enzymes [14]–[17] and top2 [15], [18]–[22]. While the effect of condensin mutations or RNAi knockdown on chromosome condensation is variable depending on cell type and organism being studied, in most if not all cases, chromatin bridges are created between chromosomes as they segregate from one another [7]. This likely represents a general role of the condensin complex in the resolution of chromosomal associations prior to segregation.
While the second cell division of meiosis is conceptually similar to mitotic divisions where sister chromatids segregate from one another, the faithful segregation of homologous chromosomes in meiosis I requires several unique steps. It is essential for homologous chromosomes to become linked to one another for proper anaphase I segregation [23] and most often this occurs through crossing over to form chiasmata [24]. As recombination requires the close juxtaposition of homologous sequences, homologs must first “identify” one another in the nucleus and then gradually become “aligned” in a manner that is DNA homology dependent, but not necessarily dictated by the DNA molecule itself. Eventually, the homologous chromosomes become “paired,” which is defined as the point when intimate and stable associations are established. The paired state is often accompanied by the laying down of a proteinaceous structure called the synaptonemal complex between paired homologous chromosomes, often referred to as “synapsis” [25],[26]. Importantly, the recombination mediated chiasmata can only provide a linkage between homologs in cooperation with sister chromatid cohesion distal to the crossover [27].
Drosophila male meiosis is unconventional in that neither recombination [28] nor synaptonemal complex formation occur [29], yet homologous chromosomes still faithfully segregate from one another in meiosis I. Two proteins have been identified that act as homolog pairing maintenance factors and may serve as a functional replacement of chiasmata. Mutations to genes encoding these achiasmate conjunction factors, MNM and SNM, cause homologs to prematurely separate and by metaphase I, they can be observed as univalents that then have random segregation patterns. It is likely that MNM and SNM directly provide conjunction of homologs as both localize to the X–Y pairing center (rDNA locus) up until anaphase I and an MNM-GFP fusion parallels this temporal pattern at foci along the 2nd and 3rd chromosomes [30]. While MNM and SNM are required for the conjunction of all bivalents, the protein Teflon promotes pairing maintenance specifically for the autosomes [31],[32]. Teflon is also required for MNM-GFP localization to the 2nd and 3rd chromosomes [30]. This suggests that Teflon, MNM, and SNM constitute an autosomal homolog pairing maintenance complex.
A fascinating aspect of Drosophila male meiosis is that during prophase I, three discrete clusters of chromatin become sequestered to the periphery of the nuclear envelope's interior. Each of these “chromosome territories” corresponds to one of the major chromosomal bivalents, either the 2nd, 3rd or X–Y [33]–[36]. A study of chromosomal associations within each prophase I bivalent demonstrated that the four chromatids begin in close alignment. Later in prophase I, all chromatids seemingly separate from one another, but the bivalent remains intact within the territory [36]. It has therefore been proposed that chromosome territories may provide stability to bivalent associations through their sequestration into sub-nuclear compartments [36].
Here we document that Drosophila putative condensin II complex subunits, Cap-H2 and Cap-D3, are necessary for normal territory formation. When they are compromised through mutation, chromatin is seemingly dispersed throughout the nucleus. We propose that the consequence of this defect is failure to individualize chromosomes from one another leading to the introduction and/or persistence of heterologous chromosomal associations into anaphase I. This underscores the role of chromosome territory formation to prevent ectopic chromosomal associations from interfering with anaphase I segregation. Cap-H2 is also necessary to resolve homologous chromosomal associations, that like heterologous associations, may be mediated by DNA entanglements and/or persistent achiasmate conjunction as anaphase I bridging is rescued by teflon mutations. This highlights condensin II mediated chromosome individualization/disjunction in meiosis I and its necessity to the creation of haploid gametes.
Faithful chromosome segregation is necessary to organismal viability, therefore it is not surprising that in Drosophila, homozygous lethal alleles exist in the following condensin subunits: SMC4/gluon, SMC2, Cap-H/barren, and Cap-G [19],[37],[38]. It has however been reported that one mutant Cap-D3 allele, Cap-D3EY00456 (Figure S1) is homozygous viable, yet completely male sterile [39]. We have confirmed the necessity of Cap-D3 to male fertility as both Cap-D3EY00456 homozygous and Cap-D3EY00456/Cap-D3Df(2L)Exel6023 males were completely sterile when mated to wild-type females. Furthermore, males trans-heterozygous for strong Cap-H2 mutations (Figure S1) were also male sterile as no progeny were derived from crosses of Cap-H2Z3-0019/Cap-H2Df(3R)Exel6159, Cap-H2TH1/Cap-H2Df(3R)Exel6159, and Cap-H2TH1/Cap-H2Z3-0019 to wild-type females. A third allele, Cap-H2Z3-5163 (Figure S1), was found to be fertile as a homozygote and in trans-combinations with Cap-H2Z3-0019, Cap-H2Df(3R)Exel6159, and Cap-H2TH1 alleles.
To determine whether the primary defect leading to loss of fertility in Cap-H2 mutant males is pre or post copulation, Cap-H2Z3-0019 homozygous mutant and heterozygous control siblings were engineered to carry a sperm tail marker, don juan-GFP, and aged in the absence of females to allow sperm to accumulate in the seminal vesicles. In contrast to Cap-H2Z3-0019 heterozygous control males where the seminal vesicles fill with sperm, those from Cap-H2Z3-0019 homozygous males were seemingly devoid of sperm as no DAPI staining sperm heads or don juan-GFP positive sperm tails were detectable (Figure 1A and 1B). The lack of mature sperm in the seminal vesicles confirmed that sterility in Cap-H2 mutant males is attributed to a defect in gamete production.
To test whether a Cap-H2 mutant allelic combination that is male fertile, Cap-H2Z3-0019/Cap-H2Z3-5163, has a decreased fertility, males of this genotype and heterozygous controls were mated to wild-type females and the percent of eggs hatched was quantified. There was no significant difference in male fertility between Cap-H2Z3-0019/Cap-H2Z3-5163 and Cap-H2Z3-5163/+ males (Figure 1C). However, the introduction of one mutant allele of another condensin subunit, SMC408819, to the Cap-H2 trans-heterozygote led to a substantial decrease in fertility relative to the SMC408819/+; Cap-H2Z3-5163/+ and SMC408819/+; Cap-H2Z3-0019/+ double heterozygous controls (Figure 1D). This suggests that Cap-H2 is functioning in the Drosophila male germline as a member of a condensin complex along with SMC4 during gametogenesis.
Given the well-documented roles of condensin subunits in promoting chromosome segregation [7], we reasoned that a possible cause of fertility loss in Cap-H2 and Cap-D3 mutants is through chromosome missegregation in the male germline. Male gametogenesis begins with a germline stem cell division. While one daughter maintains stem cell identity, the gonialblast initiates a mitotic program where 4 synchronous cell divisions create a cyst of 16 primary spermatocytes that remain connected due to incomplete cytokinesis. These mature over a period of 3.5 days, undergo DNA replication, and subsequently enter meiosis [34]. To test whether chromosome segregation defects occur during gametogenesis of Cap-H2 mutants, i.e. during the mitotic divisions of the stem cell or gonia or from either meiotic divisions, genetic tests were performed that can detect whether males create an elevated level of aneuploid sperm. In these “nondisjunction” assays, males are mated to females that have been manipulated to carry a fused, or “compound”, chromosome. Females bearing a compound chromosome and specific genetic markers are often necessary to determine whether eggs had been fertilized by aneuploid sperm. Importantly, in nondisjunction assays, fertilizations from aneuploid sperm generate “exceptional” progeny that can be phenotypically distinguished from “normal” progeny that were created from haploid sperm fertilizations.
Sex chromosome segregation was monitored as previously described for mutants in the ord gene [40], with males bred to carry genetic markers on the X and Y chromosomes. These y1w1/y+Y; Cap-H2Z3-0019/Cap-H2Z3-5163 and corresponding Cap-H2 heterozygous controls males were crossed to females bearing compound X chromosomes (C(1)RM, y2 su(wa)wa). As shown in Table 1, no significant amount of exceptional progeny were generated from Cap-H2 mutant males. It is important to point out that the lack of significant sex chromosome segregation defects found in these nondisjunction assays with a likely weak Cap-H2 male fertile mutant may be misleading. In fact, sex chromosome segregation defects are observed cytologically in stronger Cap-H2 mutant backgrounds that could not be tested with nondisjunction assays because of their sterility (see below).
Fourth chromosome segregation was assayed as described previously for teflon mutants [32], with males carrying one copy of a 4th chromosome marker mated to females bearing compound 4th chromosomes (C(4)EN, ci ey). As with the sex chromosome segregation assays, 4th chromosome segregation did not differ substantially between the Cap-H2Z3-0019/Cap-H2Z3-5163 and heterozygous control males (Table 2). The possibility remains that this hypomorphic Cap-H2 allelic combination is not strong enough to reveal 4th chromosome segregation defects. Like sex chromosomes, 4th chromosome segregation abnormalities were observed cytologically in stronger male sterile mutants (see below).
Effects on second and third chromosome segregation were assayed with the use of females carrying either compound 2 (C(2)EN, b pr) or compound 3 (C(3)EN, st cu e) chromosomes. Interestingly, both the 2nd and 3rd chromosomes had a heightened sensitivity to Cap-H2 mutation as Cap-H2Z3-0019/Cap-H2Z3-5163 males created an elevated level of exceptional progeny (Tables 3 and 4). In both cases, the exceptional class most over represented were those from fertilization events involving sperm that lacked a 2nd (nullo-2) or 3rd (nullo-3) chromosome.
Nullo progeny can be created from defects in either meiotic division. For example, the reciprocal event of incorrect cosegregation of homologs during meiosis I is one daughter cell completely lacking that particular chromosome. Similarly, nullo sperm can be created from meiosis II defects where sister chromatids cosegregate. To address whether meiotic I and or II segregation defects occur, males in the 2nd chromosome assays were bred to be heterozygous for the 2nd chromosome marker brown (bw1). If both 2nd homologous chromosomes mistakenly cosegregate in meiosis I, then a normal meiosis II will generate diplo-2 sperm that are heterozygous for the paternal male's 2nd chromosomes (bw1/+). Additionally, a normal meiosis I followed by a faulty meiosis II where sister chromatids cosegregate would generate diplo-2 sperm homozygous for the paternal male's 2nd chromosomes (bw1/bw1 or +/+). There was a trend toward an elevated level of the bw1/+ exceptional class from both Cap-H2Z3-0019/Cap-H2Z3-5163 and Cap-H2Z3-0019/+ males. This suggested meiosis I nondisjunction that possibly occurs even in Cap-H2 heterozygous males. Furthermore, there may also be a slight increase in meiosis II nondisjunction as the bw1/bw1 class is elevated in the Cap-H2 trans-heterozygous and heterozygous males.
The Cap-H2 allelic combination utilized in these genetic nondisjunction assays is likely weak in comparison to others where males are completely sterile. Therefore, the elevated frequency of exceptional progeny from 2nd and 3rd chromosome assays relative to the sex and 4th may only represent a heightened sensitivity of these chromosomes rather then a role for Cap-H2 specifically in 2nd and 3rd chromosome segregation. In fact, defects in sex and 4th chromosome segregation were observed in stronger male sterile Cap-H2 mutants (see below). One possible explanation for a major autosome bias in our nondisjunction assays may be related to the greater amount of DNA estimated for the 2nd (60.8 Mb) and 3rd (68.8 Mb) relative to the X, Y, and 4th chromosomes (41.8, 40.9, and 4.4 Mb, respectively) [41]. Thus, perhaps larger chromosomes require more overall condensin II function to promote their individualization or condensation and are therefore more sensitive to Cap-H2 dosage. While plausible, if sensitivity to Cap-H2 mutation were purely due to chromosome size, it is difficult to explain why a more significant level of XY nondisjunction did not occur given that they are ∼70% the size of the 2nd and 3rd.
An alternative hypothesis involves the fact that 2nd chromosome conjunction may occur at several sites or along its entire length [42], whereas XY bivalent pairing is restricted to intergenic repeats of the rDNA locus [43],[44]. This suggests that more total DNA is utilized for conjunction of the 2nd chromosome relative to the sex bivalent. Assuming the 3rd and 4th chromosomes maintain homolog pairing like the 2nd, then the relative amount of DNA utilized in conjunction is as follows: 3rd>2nd>4th>XY. Given that this closely parallels the trend of sensitivity to Cap-H2 mutation in the nondisjunction assays, it suggests that chromosomes which utilize more overall DNA in pairing/pairing maintenance activities require a greater dose of functional Cap-H2 for their proper anaphase I segregation. This points toward a role for Cap-H2 in the regulation of homolog conjunction/disjunction processes. We next addressed this hypothesis through cytological analyses of meiotic chromosome morphology in Cap-H2 mutant backgrounds.
In prophase I stage S2 (Figure 2A), nuclei appear to commence the formation of chromosome territories. By mid-prophase I stage S4, territory formation is more evident (Figure 2B) and in late prophase I, stage S6 nuclei exhibit three discrete chromosome territories seemingly associated with the nuclear envelope (Figure 2C) [35]. Each of the three chromosome territories corresponds to the 2nd, 3rd, and sex chromosomal bivalents and are thought to have important chromosome organizational roles for meiosis I [33]–[36]. In male sterile mutants of the genotype Cap-H2Z3-0019/Cap-H2TH1, chromosome organizational steps throughout prophase I are defective, as normal territory formation is never observed in 100% of S2, S4, and S6 stages (n = 100 nuclei of each stage). Instead, chromatin is seemingly dispersed within the nucleus (Figures 2D–2F). Male sterile Cap-D3EY00456 mutants mimic these defects (Figure 2G–2I), suggesting that Cap-D3 and Cap-H2 function together within a condensin II complex to facilitate territory formation. No prophase I defects were observed in Cap-H2Z3-0019/Cap-H2Z3-5163 males, although subtle morphological changes may be difficult to detect.
To establish possible roles for Cap-H2 and Cap-D3 in prophase I chromosome organization, it is important to outline the two general processes that must occur for proper territory formation. One is to gather or condense bivalent chromatin into an individual cluster. The second is to sequester each bivalent into a discrete pocket of the nucleus. Condensin II may perform one or both tasks, for example, perhaps chromatin is dispersed throughout the nucleus in the Cap-H2/Cap-D3 mutants because of faulty condensation. Alternatively, or in addition to, sequestration of chromatin into territories may be a primary defect in Cap-H2/Cap-D3 mutants.
During late prophase I of wild-type primary spermatocytes, chromosomes from each territory condense further and appear as three dots corresponding to the 2nd, 3rd and sex bivalents. This stage, referred to as M1 of meiosis I, may be morphologically abnormal in strong Cap-H2 mutants because it was not detected in our studies (n>50 testes). This is likely because these mutants fail to form normal chromosome territories. Proceeding further into meiosis, metaphase I is signified by the congression of the three bivalents into one cluster at the metaphase plate (Figure 3A). Despite not forming normal chromosome territories and possibly never reaching normal M1 chromosomal structure, there were no unusual features detected in Cap-H2 male sterile metaphase I figures (Figure 3D). Although subtle changes to chromosome morphology would not be detectable, it can be concluded that by metaphase I, gross chromosomal condensation occurs at least somewhat normally in Cap-H2 strong mutant males. This raises the interesting possibility that a gradual prophase I chromosome condensation is catalyzed by condensin II components in the course of chromosomal territory formation and culminates at M1. Next, a second condensation step to form metaphase I chromosomes occurs, which is only partially dependent or completely independent of condensin II components. Perhaps condensin I or some other factor is the major player for metaphase I chromosome assembly or compensates for condensin II loss.
In contrast to metaphase I, anaphase I is clearly not normal in Cap-H2 mutants, where instead bridges are often found between segregating sets of chromosomes (Figure 3E, 3F, and 3H). The frequency of these bridges occurs in a manner that matches other phenotypic trends, found in 30.4% of the anaphase I figures for sterile Cap-H2Z3-0019/Cap-H2TH1 males (n = 102 anaphase I figures), 11.5% for Cap-H2Z3-0019/Cap-H2Z3-5163 males that are fertile yet undergo 2nd and 3rd chromosome loss (n = 78), and never in the wild-type (n = 90, Figures 3B, 3C, 3G). As with territory formation, Cap-H2 is likely functioning along with Cap-D3 because in two cysts observed from Cap-D3EY00456 homozygous males, 7 of 20 anaphase I figures were bridged (Figure 3I). This anaphase I bridging most likely represents a failure to resolve chromosomal associations prior to segregation as chromatin appears to be stretched between chromosomes moving to opposing poles.
To gain further insight into why anaphase I bridges are created in Cap-H2 and Cap-D3 mutants, a chromosome squashing technique was employed that enables the visualization of individual anaphase I chromosomes (Figure 4A). With this method, the 4th chromosomes are easily identified because of their dot like appearance. Centromere placement enables the identification of the sex chromosomes, where on the X it is located very near the end of the chromosome (acrocentric) and on the Y is about a quarter of the length from one end (submetacentric). The 2nd and 3rd chromosomes are indistinguishable from one another because of their similar size and placement of the centromere in the middle of the chromosome (metacentric). Whereas bridged anaphase I figures were never observed in wild-type squashed preparations (n = 14; Figure 4A), bridging occurred in 40.5% of those from Cap-H2Z3-0019/Cap-H2TH1 mutant males (n = 42; Figure 4B–4F).
The chromosome squashing method was utilized to determine the nature of anaphase I bridges, and interestingly, it was concluded that bridging exists between both homologous and heterologous chromosomes (Figure 4). Of the total anaphase I figures from Cap-H2Z3-0019/Cap-H2TH1 testes, 21.4% appeared to have anaphase I bridging that existed between homologous chromosomes (Figure 4B and 4C). A FISH probe that recognizes 2nd chromosome pericentromeric heterochromatin was used to distinguish 2nd and 3rd chromosomes and demonstrates that linkages in Figure 4B (inset) are between the 3rd chromosomes, perhaps at regions of shared homology. Furthermore, despite not finding 4th chromosome segregation defects in nondisjunction assays (Table 1), the 4th chromosome was bridged in 4.8% of anaphase I figures (Figure 4C). This suggests that chromosome 4 becomes sensitive to further loss of Cap-H2 function in the stronger Cap-H2Z3-0019/Cap-H2TH1 mutant background.
Persistent associations between homologous chromosomes in anaphase I may be explained by a failure to individualize paired homologs from one another prior to anaphase I entry. It is probable that DNA entanglements normally exist between paired homologous chromosomes as they are likely raveled around one another rather then simply aligned side by side in a linear fashion. Therefore, individualization failure in Cap-H2 mutants may allow entanglements to persist into anaphase I. Cap-H2 may mediate homolog individualization in prophase I, where bivalents do not appear to condense properly in Cap-H2 mutants (Figure 2). Another plausible scenario is that Cap-H2 functions to antagonize achiasmate homolog conjunction mediated by teflon, MNM, and SNM at some point prior to anaphase I entry.
The other 19% of anaphase I figures that were bridged (n = 42) in the Cap-H2Z3-0019/Cap-H2TH1 mutant involve heterologous chromosomes (Figure 4D) and cases where bridging is so substantial that its chromosomal nature could not be determined (Figure 4E and 4F). The observed X–Y linkage in Figure 4D is consistent with the XY pairing site, or “collochore,” and occurs in wild-type preparations [45]. The other linkage is an atypical heterologous association occurring between the Y and one of the major autosomes (2nd or 3rd). We speculate that the substantially bridged images in Figure 4E and 4F are comprised of associations between heterologous and/or homologous chromosomes. Figure 4F is particularly interesting because the 4th and sex chromosomes appear to have segregated normally, yet the major autosomes remain in an unresolved chromosomal mass. This pattern fits the trend of the nondisjunction studies, where the 2nd and 3rd chromosomes had a heightened sensitivity to Cap-H2 mutation.
Because the 4th chromosome naturally tends to be separated from other prometaphase I to anaphase I chromosomes, it was often easily observed to be involved in heterologous chromosomal associations (Figures 4G, 4H, and 5A'). These appear as threads and occurred in 42.5% of metaphase and anaphase I figures (n = 40; Figures 4G and 5A'). Interestingly, 4th-to-heterolog threads were also observed in the wild-type, although at a lower frequency of 19% (n = 21, Figure 4H).
Persistent associations between heterologous chromosomes such as that observed in figure 4D and inferred to exist within 4E and 4F may be traced to failed territory formation in Cap-H2 mutant prophase I. Perhaps interphase chromosomes are naturally entangled with one another and the Cap-H2/Cap-D3 mediated nuclear organization steps that occur during territory formation effectively detangle and individualize them into discrete structures. Alternatively, Cap-H2/Cap-D3 mediated chromosome territory formation may act to prevent the establishment of heterologous entanglements. These are plausible scenarios given that failed territory formation in Cap-H2/Cap-D3 mutants seemingly leads to persistent intermingling of all chromosomes. Such an environment could provide a likely source of heterologous chromosomal associations. Heterologous associations involving the 4th chromosome may also be entanglements that persist and/or were initiated through failure in territory formation. These cannot however be completely attributed to loss of Cap-H2 function because they were observed in the wild-type (Figure 4H).
The anaphase I bridging in Cap-H2 mutant males is one likely source for their elevated amount of nullo-2 and nullo-3 sperm (Tables 3 and 4). Chromatin stretched between daughter nuclei may occasionally lead to the creation of sperm lacking whole chromosomes or variable sized chromosomal regions. Bridged anaphase I images in Figure 4 represent likely scenarios where chromosome loss would occur and furthermore, visualization of the post-meiotic “onion stage” from Cap-H2 mutants is consistent with chromosome loss. With light microscopy, white appearing nuclei within the onion stage are nearly identical in size to the black appearing nebenkern, which represents clustered mitochondria (Figure 4I, arrows). In onion stages from Cap-H2Z3-0019 homozygotes, micronuclei are often observed which may be the manifestation of chromatin lost through anaphase I bridging (Figure 4I, arrowheads).
The associations that create anaphase I bridging between chromosomes moving to opposing poles may also be capable of causing improper cosegregation of homologs. In fact, 9.5% of squashed anaphase I figures (n = 42) are of asymmetrically segregating homologs that were never observed in the wild-type (n = 14). These are consistent with failure in homolog disjunction and subsequent cosegregation to one pole (Figure 5A–5D). These may also be the consequence of associations between heterologous chromosomes that lead to one being dragged to the incorrect pole. As an expected outcome of cosegregation in meiosis I, aneuploidy in prophase II and anaphase II figures was also observed (Figure 5E–5G). Such events likely explain the slight increase in diplo-2 sperm that were heterozygous for the male's 2nd chromosomes (bw1/+ in Table 3). The also provide a likely source for the elevated amount of nullo-2 and nullo-3 sperm (Tables 3 and 4).
While the prevalence of meiotic anaphase I bridging is likely a major contributor to the observed 2nd and 3rd nondisjunction, it cannot be ruled out that the preceding stem cell and gonial mitotic divisions are also defective and lead to aneuploid sperm. This exists as a formal possibility, yet aneuploid meiotic I cells were not observed in squashed Cap-H2 mutant anaphase I figures where all chromosomes could be distinguished (n = 10). This suggests that pre-meiotic segregation is unaffected. Similarly, anaphase II defects could have contributed to the elevated nullo-2 and nullo-3 sperm and perhaps the slight increase in bw1/bw1 progeny that would have been generated from meiosis II nondisjunction (Table 3). In fact, anaphase II bridging was observed in 8.7% of Cap-H2Z3-0019/Cap-H2TH1 anaphase II figures (n = 69, Figure 6), 2.1% of those from Cap-H2Z3-0019/Cap-H2Z3-5163 males (n = 47), and never in the wild-type (n = 66). Anaphase II defects may occur because of a specific role of Cap-H2 in meiosis II, or alternatively, anaphase II bridging could be attributed to faulty chromosome assembly or individualization in meiosis I.
The protein Teflon is implicated in the maintenance of Drosophila male meiosis I autosome conjunction as teflon mutants lose autosomal associations prior to anaphase I [31]. To investigate whether persistent associations between homologous chromosomes in anaphase I of Cap-H2 mutants (Figure 4B and 4C) are Teflon dependent, teflon mutations were crossed into a Cap-H2 mutant background and the frequency of anaphase I bridging was assessed. While 30.4% of anaphase I figures from Cap-H2Z3-0019/Cap-H2TH1 males were bridged (n = 102), bridging existed within only 10.8% of anaphase I figures from tefZ2-5549/tefZ2-5864; Cap-H2Z3-0019/Cap-H2TH1 males (n = 74, p<1×10−6, X2) (Figure 7A). Furthermore, in squashed preparations anaphase I bridging was decreased from 40.5% in Cap-H2Z3-0019/Cap-H2TH1 males (n = 42) to 25.6% in the tefZ2-5549/tefZ2-5864; Cap-H2Z3-0019/Cap-H2TH1 double mutants (n = 43, p<0.05, X2).
The ability of teflon mutations to rescue Cap-H2 mutant anaphase I bridging suggests that Cap-H2 functions to antagonize Teflon mediated autosome conjunction. This may entail deactivation of an achiasmate conjunction complex consisting of MNM, SNM, and perhaps Teflon, at some point prior to the metaphase I to anaphase I transition. Consistent with this hypothesis, the percent of anaphase I figures where homologous chromosomes appeared to be bridged were decreased from 21.4% in the Cap-H2Z3-0019/Cap-H2TH1 mutants (n = 42) to 9.3% in tefZ2-5549/tefZ2-5864; Cap-H2Z3-0019/Cap-H2TH1 males (n = 43, p<0.1, X2, Figure 7B).
As an important alternative to Cap-H2 functioning to antagonize an achiasmate homolog conjunction complex, it may be that wild-type Teflon exacerbates DNA associations between chromosomes. For example, perhaps Teflon linked homologs are now particularly prone to becoming entangled. Under this scenario, teflon mutations may decrease the opportunity for DNA entanglements to be introduced between homologs because of their spatial distancing from one another during late prophase I to metaphase I. Given the formal possibility of both models, we conclude that Cap-H2 functions to either remove teflon dependent conjunction and/or to resolve chromosomal entanglements between homologs.
The remaining bridged anaphase I figures from squashed preparations in tefZ2-5549/tefZ2-5864; Cap-H2Z3-0019/Cap-H2TH1 males were uninterpretable making it impossible to assess whether Cap-H2 mutant heterologous anaphase I bridging was also rescued by teflon mutation. However, 4th-to-heterolog threads were greatly suppressed by teflon mutations, decreasing from 42.5% (n = 40) to only 6% (n = 50, p<0.00001, Figure 7C). This is a surprising result given that Teflon has been described as a mediator of associations between homologous chromosomes. One plausible explanation is that Teflon can exacerbate heterologous chromosomal associations. This may occur when Teflon establishes autosomal conjunction in a prophase I nucleus where territory formation had failed. Cap-H2 may also antagonize a Teflon mediated autosomal conjunction complex that might mistakenly establish conjunction between heterologs when territories do not form.
As described above, completely male sterile Cap-D3 and Cap-H2 allelic combinations exist and Cap-H2 mutant males lack mature sperm in their seminal vesicles (Figure 1A and 1B). One possible explanation for this result is that chromosome damage created during anaphase bridging in the Cap-H2 mutants causes spermatogenesis to abort. This scenario seems less likely because tefZ2-5549/tefZ2-5864 rescued Cap-H2Z3-0019/Cap-H2TH1 anaphase I bridging to levels near that of fertile Cap-H2 mutants, yet tefZ2-5549/tefZ2-5864; Cap-H2Z3-0019/Cap-H2TH1 males were still found to be completely sterile. This points toward another function for Cap-H2 in post-meiotic steps of spermatogenesis.
Figure 8 illustrates a working model of condensin II in Drosophila male meiosis to resolve both heterologous and homologous chromosomal associations. We speculate that these associations likely consist of DNA entanglements that naturally become introduced between interphase chromosomes due to their threadlike nature (Figure 8A). The studies herein identified a function for condensin II during prophase I, when paired homologous chromosomes become partitioned into discrete chromosomal territories [33]–[36]. We propose that condensin II either promotes this partitioning, by actively sequestering bivalents into different regions of the nucleus, or functions to perform prophase I chromosome condensation. It is important to stress that in both scenarios, the role of condensin II mediated territory formation is to ensure the individualization of heterologous chromosomes from one another (Figure 8B, large arrows). When sequestration into territories and/or condensation of the bivalents do not take place, i.e. in the condensin II mutants, individualization does not occur, heterologous entanglements persist into anaphase I, and chromosomes may become stretched to the point where variable sized chromosomal portions become lost (Figure 8E). Persistent heterologous entanglements may also lead to one chromosome dragging another to the incorrect pole (not shown).
Despite what appears to be failed chromosome condensation in prophase I of Cap-H2 mutants, by metaphase and anaphase I no obvious defects in chromosome condensation were observed (Figures 3, 4, 5, and 7). This suggests that sufficient functional Cap-H2 is present in this mutant background to promote metaphase/anaphase I chromosome condensation. Alternatively, perhaps another factor fulfills this role and/or compensates for condensin II loss. This parallels Cap-G mutants, where embryonic mitotic prophase/prometaphase condensation was abnormal, yet metaphase figures appeared wild-type [38]. In Drosophila, mutant and RNAi knockdown studies of condensin complex subunits in mitosis lead to a range of phenotypes, from complete failure in condensation [46] to seemingly normal axial shortening, but failure in chromatid resolution [37],[39]. The variable phenotypes produced from these studies may reflect differences in cell type specific demand for condensin subunit dosage/activity.
Anaphase I figures of Cap-H2 mutants also revealed persistent entanglements between homologous chromosomes that may be at regions of shared homology. We suggest that the paired state of homologs initiates or introduces the opportunity for DNA entangling between homologs and that condensin II functions to resolve these prior to segregation. A likely scenario is that this occurs during prophase I, where chromosome condensation appears abnormal in Cap-H2 and Cap-D3 mutants. Perhaps condensin II mediated prophase I condensation functions to individualize intertwined homologous chromosomes prior to segregation (Figure 8B, small arrows). It is also plausible that condensin II homolog individualization continues up until anaphase I.
We have found that mutations in teflon, a gene required for autosomal pairing maintenance, were capable of suppressing anaphase I bridging in Cap-H2 mutant males. Specifically, both homologous and heterologous chromosomal bridging were decreased in the teflon/Cap-H2 double mutant. This may occur because Teflon is capable of exacerbating DNA entanglements, if for example persistent homolog conjunction provides more opportunity for entanglements between homologs to be introduced. Teflon may also exacerbate entanglements between heterologous chromosomes. This might be especially true in a Cap-H2 mutant background with failed territory formation, as Teflon mediated autosomal conjunction may augment the extent of entangling.
It is also plausible that Cap-H2 acts as an antagonist of Teflon mediated autosomal conjunction. Perhaps autosomal homologous associations persist into anaphase I of Cap-H2 mutants because a homolog conjunction complex was not disabled prior to the metaphase I to anaphase I transition. However, Cap-H2 as an antagonist of Teflon cannot explain persistent heterologous associations into anaphase I, unless Teflon is capable of mistakenly introducing conjunction between heterologous chromosomes. The opportunity for this might exist in a Cap-H2 mutant prophase I nucleus where heterologs continue to intermingle because of failed territory formation.
An interesting result in our course of studies was the heightened amount of chromosome 2 and 3 nondisjunction in weaker male fertile Cap-H2 allelic combinations, whereas the sex and 4th chromosomes were unaffected. This is reminiscent of mutants from several other genetic screens that only affected the segregation of specific chromosomes or subsets [32], [47]–[51]. However, given that sex and 4th chromosome segregation defects were observed in the stronger male sterile Cap-H2 mutant background, we propose that condensin II functions upon all chromosomes, yet the 2nd and 3rd require the greatest functional Cap-H2 dose for their proper segregation. This sensitivity of the 2nd and 3rd chromosomes may be due to their greater total amount of DNA utilized in homolog pairing and pairing maintenance activities. For example, perhaps longer stretches of paired DNA are more prone to entanglements or require more achiasmate conjunction factors and therefore necessitate higher levels of Cap-H2 individualization or disengagement activity. As an interesting corollary to support this theory, weak teflon mutations only lead to 4th chromosome missegregation, while the other autosomes segregate normally [31]. This suggests that the 4th chromosomes are more sensitive to Teflon dosage because of their fewer sites of conjunction.
The majority of the data provided in this manuscript were on our studies of mutant Cap-H2 alleles, however, we found that a homozygous viable Cap-D3 mutant also failed to form normal chromosomal territories and exhibited anaphase I chromosome bridging. This provides support that these two proteins are functioning together within a condensin II complex. It is important to point out however, that to date there is no data in Drosophila to support that these proteins physically associate with each other or with other condensin subunits, namely SMC2 and SMC4 (a Drosophila Cap-G2 has yet to be identified with computational attempts) [8].
At this point in our studies of putative condensin II subunits in disjunction of achiasmate male homologous chromosomes, we cannot distinguish between possible scenarios that Cap-H2 and Cap-D3 act to disentangle chromosomes through individualization activity, that they function as antagonists of Teflon dependent achiasmate associations, or a combination of both activities. The fact that Teflon mutations do rescue Cap-H2 anaphase I bridging defects is an especially intriguing result as it points toward a molecular mechanism for Cap-H2 as an antagonist of achiasmate associations. While three genes have been found to promote achiasmate conjunction (teflon, MNM, and SNM), no factors have been identified that act to negatively regulate conjunction and allow homologs to disengage at the time of segregation. Interestingly, one conjunction factor, SNM, is orthologous to the cohesin subunit Scc3/SA that appears to be specialized to engage achiasmate homologs [30]. Condensin has been shown to antagonize cohesins in budding yeast meiosis [52] and mitotic human tissue culture cells [53]. This raises the possibility that a conserved molecular mechanism exists for condensin II as a negative regulator of SNM in Drosophila male meiosis. The investigation of Teflon, MNM, and SNM protein dynamics in a Cap-H2 mutant background will be an important set of future studies to help decipher the function of Cap-H2 in achiasmate segregation mechanisms.
Homologous chromosomal individualization in meiosis I has been previously documented as a condensin complex catalyzed activity in C. elegans as homologs remained associated in hcp-6/Cap-D3 mutants even in the absence of recombination and sister chromatid cohesion [6]. Here we demonstrated that condensin subunits are also required to individualize heterologous chromosomes from one another prior to anaphase I. As discussed above, this is likely through condensin II mediated chromosome organizational steps that occur during prophase I territory formation. This suggests that Drosophila males carry out territory formation to disfavor associations between heterologs, while also enriching for interactions between homologs. This model is particularly interesting as it may point toward an adaptation of Drosophila males to ensure meiotic I segregation in a system lacking a synaptonemal complex and recombination.
To visualize sperm head (DAPI) and tail (don juan-GFP) content in the seminal vesicles, males were restricted from females for ten days, then testes were dissected and fixed as previously described for whole mounted ovaries [54]. Meiotic microtubules were detected with rat anti-alpha tubulin antibodies (Serotec, MCA78G and MCA77G) at 1∶40 each and a FITC-conjugated donkey anti-rat secondary (Jackson ImmunoResearch, #112-095-167) at 1∶200. Immunofluorescence was conducted following protocols 5.2 and 5.6 from ref [55], with the addition of two extra final PBS washes, the second to last containing 100 ng/ul DAPI. DAPI stained chromosome squashes were prepared as detailed in protocol 1.9, method #3 w/o steps necessary for immuno-detection from ref. [56]. Testes were opened to release cells while in fixative on a siliconized coverslip prior to lowering a non-siliconized slide and squashing. Subsequent FISH to anaphase chromosome spreads was conducted as detailed in protocol 2.9 in ref. [57]. An (AACAC)6 oligonucleotide end labeled with terminal deoxytransferase (Roche 03333566001) and reagents provided in the ARES Alexa Fluor 546 DNA labeling kit (Invitrogen A21667) were utilized to fluorescently detect 2nd chromosome pericentromeric heterochromatin. All imaging was performed with a Zeiss Laser Scanning Microscope, LSM 510 Meta, and the acquisition software LSM 510 Meta, version 4.0. Images in figure 1 were captured with a Plan-Apochromat 20×/0.8 objective at an image bit depth of 8 bit. All other images were acquired with a Plan-Apochromat 63×/1.4 Oil DIC objective at an image bit depth of 8 bit. Appropriate filters and dichroic mirrors for fluorochromes DAPI, Alexa Fluor 546, and FITC were used where applicable.
To test for male fertility, 10 mutant males were crossed to 20 wild-type (Oregon R) virgin females and monitored frequently for the presence of larvae. To score fertility over time of the Cap-H2 trans-heterozygous and heterozygous control males, 10, 1–4 day old males were placed with 30, 1–5 day old virgin females in containers with grape juice agar plates and wet yeast. Flies were transferred to new plates every 24 hours for 4 days, but on the 4th, 8th, and 12th days, only males were kept and placed with a new batch of 1–5 day old virgin females. This scheme was carried out over a period of 16 days and in triplicate. For the SMC4; Cap-H2 double mutant studies, the strategy is as detailed above, except only 20 virgin females were used for each brood. To score hatch rates, the percent of eggs that hatched (n = 200 total eggs/plate) was scored from randomly selected regions of each plate 48 hours after parents were removed.
Five Cap-H2Z3-0019/Cap-H2Z3-5163; spapol/+ males were crossed to fifteen C(4)EN, ci ey females at 25°C on standard fly food. As controls, the same experimental design was carried out with Cap-H2Z3-0019/TM6B, Hu; spapol/+ or Cap-H2Z3-5163/TM6B, Hu; spapol/+ males. Males and virgin females were 2–3 days old and the experimental cross was done in replicate, while the controls were only performed once. Parents were twice flipped into a new bottle after three days and then discarded from their final bottle after three days. Progeny were scored on the 13th, 15th, and 18th day after parents were placed into the bottle. Because the 4th chromosome in these females is attached, they produce eggs that either carry the compound C(4)EN, ci ey chromosome (diplo-4) or no 4th chromosome (nullo-4). The fertilization of nullo-4 eggs by normal haploid sperm creates nullo-4/+ and nullo-4/spapol progeny. Both of these will develop into very small flies (Minute) from only carrying one 4th chromosome, with the latter also spapol. When normal haplo-4 sperm fertilize C(4)EN, ci ey/0 eggs, C(4)EN, ci ey/+ or C(4)EN, ci ey/spapol progeny are produced. These both appear wild-type from the wild-type alleles of ci and ey on the paternal 4th and wild-type spapol on the C(4)EN chromosome. There are two exceptional classes from male chromosome missegregation events that are detectable with this assay. The first is when nullo-4 sperm fertilize C(4)EN, ci ey/0 eggs to produce ci ey offspring. The second are sperm diplo-4 and homozygous for spapol fertilizing nullo-4 eggs to create spapol offspring. The following exceptional classes go undetected with this assay because they are phenotypically wild-type: +/+, spapol/+, spapol/spapol sperm that fertilize C(4)EN eggs or +/+, spapol/+ and all triplo-4 and tetra-4 sperm possibilities that fertilize nullo-4 eggs. Therefore, the % 4th chromosome nondisjunction is likely an underestimate. This assay was adapted from that described in ref. [32].
Ten males, that were 2–3 days old, were crossed to 17 virgin females that were 0–3 days old at 25°C. Males each carried a Y chromosome with an X translocation containing the wild-type yellow gene. Females carried an attached X chromosome: C(1)RM, y2 su(wa)wa. In this assay, the viable offspring from sperm bearing the normal sex chromosome content, either one X or one Y, will be y1w1/nullo-X (yw, XO male) or y+Y/C(1)RM, y2 su(wa)wa (y+, XXY female) (nullo-X/Y and triplo-X are lethal combinations). If exceptional classes of sperm are created that are diplo-X, XY, XXY, or lack either sex chromosome entirely (nullo-X or nullo-Y), then yellow white females, white males, white females, or yellow females will be produced, respectively. With this assay it cannot be determined whether offspring carry an extra Y chromosome. This experiment is adapted from that detailed in ref. [40].
The line C(2)EN, b pr carries second chromosomes that are fused, referred to as “compound” chromosomes, that segregate together as a unit and therefore gametes are created that are either nullo-2 or diplo-2. Because any chromosome 2 content other than diplo-2 is lethal, viable offspring only occur from the fertilization of nullo-2 eggs by diplo-2 sperm or diplo-2 eggs by nullo-2 sperm. Therefore, if any offspring are created when crossing males to C(2)EN, b pr virgin females, then chromosome mis-segregation had occurred in the generation of male gametes. The males used in this experiment were heterozygous for a mutant allele of brown (bw1) that is an insertion of a 412 retrotransposable element into the brown gene. In this assay, there are four classes of sperm that can successfully fertilize eggs from C(2)EN bearing females that can then develop into adult flies: nullo-2, diplo-2 (bw1/bw1), diplo-2 (bw1/+), and diplo-2 (+/+). Progeny from nullo-2 sperm fertilizing diplo-2 eggs have the b pr phenotype. Those from bw1/bw1 sperm fertilizing nullo-2 eggs have the bw phenotype. Progeny from bw1/+ and +/+ sperm fertilizing nullo-2 eggs both appear wild-type. To distinguish between these two wild-type phenotypic classes, a PCR test was developed that could detect the presence of the bw1 mutant allele by utilizing the 412 element insertion in the brown gene. Thus, with forward primer tattatctgagtgagttttctcgag that anneals to the 412 element and reverse primer ttcacccacatcatcctcat that anneals to the brown gene, a 874 bp PCR product is generated only from bw1/+ and never from +/+ flies. Furthermore, with forward primer ggtgatctgcaattagggat and the same reverse primer as above (ttcacccacatcatcctcat), an ∼571 bp fragment amplifies from the wild-type brown locus within both bw1/+ and +/+ flies, and serves as a positive control. Wild-type in these assays was the parental line from the Z3-0019 and Z3-5163 backgrounds [58] crossed to Oregon R (bw1/+; st1/+). Similarly, Cap-H2 heterozygous males were generated from a cross to Oregon R. Ten 1–3 day old males were crossed to twenty 1–5 day old virgin C(2)EN, b pr females at 25°C. This was replicated 19 times for the bw1/+; Cap-H2Z3-0019/Cap-H2Z3-5163 males, 12 for bw1/+; Cap-H2Z3-0019/+, 20 for bw1/+; Cap-H2Z3-5163/+, and 15 for bw1/+; st1/+. The parents were kept in the original vial for a total of 5 days, flipped to a new vial for 5 more days, and then discarded. The progeny were scored on the 13th, 15th, and 18th day after parents were placed together into a vial.
Like the second chromosome, any chromosome 3 content other than diplo-2 is lethal, so viable offspring only occur from the fertilization of nullo-3 eggs by diplo-3 sperm or diplo-3 eggs by nullo-3 sperm. This experiment was therefore set up in the same way as the 2nd chromosome nondisjunction tests, except that C(3)EN, st cu e females were used, three replicates were performed, parents were kept in vials for 3 days and flipped twice, and these crosses were done at room temperature (21–23°C). In this assay, there are four classes of sperm that can successfully fertilize eggs from C(3)EN bearing females that can then develop into adult flies: nullo-3, diplo-3 (heterozygous for paternal 3rd chromosomes), diplo-3 (homozygous for one of the paternal 3rd chromosomes) and diplo-3 (homozygous for the other paternal 3rd chromosome). The Cap-H2Z3-0019 chromosome is marked with ru, h, st, sr, e, and ca, while the Cap-H2Z3-5163 chromosome is marked with only st. Using Cap-H2Z3-0019/Cap-H2Z3-5163 males as an example, the following describes how nullo-3 and the three different diplo-3 progeny classes were distinguished. Progeny from nullo-3 sperm fertilizing C(3)EN, st cu e, eggs have the st cu e phenotype. Those from diplo-3, Cap-H2Z3-0019/Cap-H2Z3-0019, sperm fertilizing nullo-3 eggs would be ru h st sr e ca. The progeny from diplo-3, Cap-H2Z3-0019/Cap-H2Z3-5163 and Cap-H2Z3-5163/Cap-H2Z3-5163, sperm fertilizing nullo-3 eggs both develop into st animals. These were distinguished by crossing to ru h st Cap-H2Z3-0019 st e ca/TM6B, Hu Tb e ca flies and scoring F2 progeny.
The percentage of bridged anaphase I figures where chromosomes are oriented such that their identity is unambiguous is low. Additionally, anaphase I chromosomes quickly decondense upon entry into telophase I, reducing the overall frequency of anaphase I figures where chromosomes can be observed. Thus, the stronger Cap-H2Z3-0019/Cap-H2TH1 allelic combination was analyzed to increase the likelihood of visualizing interpretable bridged figures. Bridges were scored as homologous when they appeared to connect morphologically similar chromosomes, based on size and centromere location (see text) that appeared to be segregating away from one. It was concluded that the 4th chromosome was involved in a heterologous association during meiosis I when a DAPI staining thread extended to another non-4th chromosome. Thus, images were only scored when this thread clearly was connected to a heterolog, or the other 4th was present and it was clear that it did not participate in the thread. In the wild-type figures where 4th chromosome threads were observed, it could not be concluded whether the thread extended to another 4th or a heterolog. The data in figure 7C for the wild-type may therefore be an overestimate of 4th-to-heterolog threads because threads may actually connect homologs.
bw; st Z3-0019/TM6B, Hu Tb e ca and bw; st Z3-5163/TM6B, Hu Tb e ca were obtained from Charles Zuker [58] and were identified in a previously detailed genetic screen [59]. A recombinant chromosome of the Z3-0019 line, ru h st Cap-H2Z3-0019 sr e ca/TM6B, Hu Tb e ca was used for all experiments herein. The Cap-H2TH1 allele was found on the Df(3L)W10 bearing chromosome during the course of complementation studies that will be described elsewhere. The deficiency Df(3L)W10 was recombined away from the Cap-H2TH1 bearing chromosome and instead ru h st Cap-H2TH1 Sb[sbd-2]/TM6B, Hu Tb e ca was utilized in these studies. The stocks SMC4k08819, spapol, C(2)EN, b pr, C(3)EN, st cu e, Df(3R)Exel6159, Cap-D3EY00456, and Df(2L)Exel7023 were obtained from the Bloomington stock center. John Tomkiel provided the following stocks: cn tefZ2-5549 bw/CyO, cn tefZ2-5864 bw/CyO, and y w sn; C(4)EN, ci ey. The don juan-GFP/CyO and C(1)RM, y2 su(wa)wa were received from Terry Orr-Weaver.
|
10.1371/journal.ppat.1000093 | Quorum Sensing Coordinates Brute Force and Stealth Modes of Infection in the Plant Pathogen Pectobacterium atrosepticum | Quorum sensing (QS) in vitro controls production of plant cell wall degrading enzymes (PCWDEs) and other virulence factors in the soft rotting enterobacterial plant pathogen Pectobacterium atrosepticum (Pba). Here, we demonstrate the genome-wide regulatory role of QS in vivo during the Pba–potato interaction, using a Pba-specific microarray. We show that 26% of the Pba genome exhibited differential transcription in a QS (expI-) mutant, compared to the wild-type, suggesting that QS may make a greater contribution to pathogenesis than previously thought. We identify novel components of the QS regulon, including the Type I and II secretion systems, which are involved in the secretion of PCWDEs; a novel Type VI secretion system (T6SS) and its predicted substrates Hcp and VgrG; more than 70 known or putative regulators, some of which have been demonstrated to control pathogenesis and, remarkably, the Type III secretion system and associated effector proteins, and coronafacoyl-amide conjugates, both of which play roles in the manipulation of plant defences. We show that the T6SS and a novel potential regulator, VirS, are required for full virulence in Pba, and propose a model placing QS at the apex of a regulatory hierarchy controlling the later stages of disease progression in Pba. Our findings indicate that QS is a master regulator of phytopathogenesis, controlling multiple other regulators that, in turn, co-ordinately regulate genes associated with manipulation of host defences in concert with the destructive arsenal of PCWDEs that manifest the soft rot disease phenotype.
| Many Gram-negative bacteria use a population density-dependent regulatory mechanism called quorum sensing (QS) to control the production of virulence factors during infection. In the bacterial plant pathogen Pectobacterium atrosepticum (formerly Erwinia carotovora subsp. atroseptica), an important model for QS, this mechanism regulates production of enzymes that physically attack the host plant cell wall. This study used a whole genome microarray-based approach to investigate the entire QS regulon during plant infection. Results demonstrate that QS regulates a much wider set of essential virulence factors than was previously appreciated. These include virulence factors similar to those in other plant and animal pathogens that have not previously been associated with QS, e.g., a Type VI secretion system (and its potential substrates), shown for the first time to be required for virulence in a plant pathogen; and the plant toxin coronafacic acid, known in other pathogens to play a role in manipulating plant defences. This study provides the first evidence that Pectobacterium may target host defences simultaneously with a physical attack on the plant cell wall. Moreover, the study demonstrates that a wide range of previously known and unknown virulence regulators lie within the QS regulon, revealing it to be the master regulator of virulence.
| Quorum sensing (QS) is a population density-dependent regulatory mechanism, utilising freely diffusible chemical signal molecules, which controls a wide range of phenotypes in many different bacteria [1]. The best-studied QS systems are those utilising N-acyl-homoserine lactone (AHL) signal molecules, synthesised by LuxI homologues. AHL concentration increases with bacterial population growth until, at high cell density, a threshold level of signal is reached. This is detected by AHL binding to receptor proteins, LuxR-family transcriptional regulators, resulting in altered gene expression [2]. QS plays an essential role in the pathogenesis of many bacterial pathogens of both plants and animals. Amongst the best studied AHL QS systems are those of the soft rotting enterobacterial plant pathogens Pectobacterium atrosepticum (Pba) and Pectobacterium carotovorum subsp. carotovorum (Pcc; formerly Erwinia carotovora subsp. atroseptica and E. c. subsp. carotovorum respectively) [3]. These pathogens cause disease primarily through the coordinate and prolific production of a variety of plant cell wall degrading enzymes (PCWDEs), which are secreted to the extracellular environment through the Type I (protease) and Type II (pectinases and cellulases) secretion systems [4]. However, they also possess a Type III secretion system (T3SS) with cognate effector (DspA/E) and helper/harpin proteins (HrpN/HrpW), which is required for full virulence [5]. While the role of the T3SS in the soft rotting pathogens remains to be elucidated, in the closely-related E. amylovora, DspA/E has been reported to interact with leucine-rich repeat receptor-like protein kinases (LLR-RLKs) of apple plants, implying a role in the manipulation of host defences [6]. QS in pectobacteria has been reported to regulate PCWDEs [7], the Type III secreted harpin HrpN [8], and other virulence factors, including Nip and Svx [9]–[11], a very small number of virulence regulators (expR, rsmA and virR) [12]–[14], and the antibiotic carbapenem [15]. These are controlled by the AHL, N-(3-oxohexanoyl)-L-homoserine lactone (OHHL), synthesised by ExpI. Different strains of pectobacteria possess up to three homologues of LuxR [16] including: VirR, which plays a central role in the repression of QS-regulated virulence factors [12]; CarR, which regulates the production of carbapenem [15]; and ExpR, which activates transcription of the global repressor, rsmA, in the absence of AHL [13].
Until now, studies on QS in pectobacteria have largely been in vitro and have examined its role in the regulation of targeted virulence factors, particularly PCWDEs. Such virulence factors are thought to operate as part of a necrotrophic mode of action (where the invading organism causes death of host tissue and colonises dead substrate). As a consequence, this group of pathogens have been termed “brute force” in line with this physical attack on plant cell walls. This is in contrast to pathogens such as Pseudomonas syringae, which are hemibiotrophic (requiring living host tissue as part of the infection process, during which they actively manipulate host defences) and, due to their ability to manipulate plant defences as part of the infection process, have been termed “stealth” pathogens. In the pectobacteria, it has been hypothesised that QS acts to delay the onset of PCWDE production until sufficient numbers of cells are present to overcome plant defences, which are induced by the formation of cell wall breakdown products [17],[18]. However, in previous work we showed that premature addition of OHHL to potato plants infected with low numbers of Pba induced early disease development [19], suggesting that this hypothesis may be an over-simplification of a more complex process. In addition, the full genome sequence of Pba strain SCRI1043 (Pba1043) has revealed many additional putative virulence determinants, including coronafacoyl-amide conjugates and homologues of the hemolysin-co-regulated protein (Hcp) and Rhs accessory element VgrG [20]. In Pseudomonas syringae, coronafacoyl-amide conjugates promote disease development and, together with the T3SS, may act to suppress salicylic acid-based defences as part of this process [21],[22]. Hcp and VgrG have been associated with virulence in animal pathogens and are potential effector proteins delivered through a Type VI secretion system (T6SS) [23]–[26]. Hcp and VgrG homologues were recently detected in the secretome of Pba1043 and over-expression of hcp1 increased Pba virulence, suggesting that this and other hcp family members are virulence determinants [27]. The presence of such determinants in Pba suggests that the pectobacteria may also act in a stealth-like manner by manipulating resistance during the infection process. However, whether these determinants are produced and act independently, or together with PCWDEs as part of a coordinated assault on the plant, is unknown. We developed a whole genome microarray for Pba1043 and report its use to study gene expression from an expI mutant of Pba1043 grown in planta, to determine global effects of QS on gene regulation during potato infection, with particular emphasis on the relationship between PCWDEs and possible stealth mechanisms.
The expI gene and ExpI product, OHHL, are required for full virulence in Pba and Pcc [7],[12],[28]. The virulence of an expI (ECA0105) mutant was significantly reduced on both potato stems and tubers and was restored following complementation with the expI gene in trans (Fig. S1). To confirm that virulence could be restored in planta by the presence of OHHL, the expI mutant strain was inoculated at low cell densities into an OHHL-producing transgenic potato plant [19], where virulence was restored compared to inoculation on a non-transgenic control plant (Fig. 1A). This supports previous work where the presence of OHHL in these transgenic plants induced early disease development from low cell densities (102 cells per inoculation site) of both WT and expI mutant strains [19].
A luminescence-based assay was used to monitor OHHL production during growth of the expI mutant and wild type strains in potato tubers (Fig. 1B). Both strains grew at comparable rates over a 120 h infection time course and reached similar population levels. Although the wild type and expI mutant strains would be expected to show differences in growth in natural condition during the course of disease development, the relatively short infection time (120 h) and the method of inoculation (onto the cut tuber surface), may account for the results observed. In the wild type, the level of OHHL rose sharply over the first 16 hours in line with log phase growth, before reaching a plateau at a concentration of approximately 80 µg/ml. At this plateau, bacterial cell density was approximately 5.0×106 cfu/ml, which was similar to previous reports [15]. In the expI mutant, OHHL production remained at background levels. Based on the above data, time points at 12 and 20 hours post inoculation (hpi), i.e. just prior to and just following maximum OHHL synthesis in planta, were selected to study transcriptional changes during QS (Fig. 1B).
Differential expression of genes (pelA [ECA4067], pelC [ECA4069], celV [ECA1981], prtW [ECA2785], pehA [ECA1095], ECA2220, svx [ECA0931] and nip [ECA3087]) previously shown to be under QS control [9],[10],[12] was investigated using quantitative real-time PCR (qRT-PCR) at 12 and 20 hpi in the expI mutant and wild type strains. In all cases, significant up-regulation of these genes was observed in the wild type only (Table S1). cDNA from the wild type and expI mutant at 12 and 20 hpi was hybridised to the Pba microarray. 1167 coding sequences (CDSs) (approx. 26% of the genome) showed statistically significant differences (P≤0.05) in expression between the expI mutant and wild type (Table S2). 498 CDSs showed reduced transcript abundance (421 at 12 hpi, 169 at 20 hpi, 92 at both time points) and 687 CDSs exhibited increased transcript abundance (551 at 12 hpi, 180 at 20 hpi, 44 at both time points) in the expI mutant compared to the wild type. Microarray comparison of mutant and wild type cDNAs from cells in buffer solution prepared for tuber inoculation following overnight growth in LB to stationary phase (zero time-point), was consistent with there being no overall transcriptional difference (P≤0.05) between the strains prior to plant inoculation (data not shown). Only 16% of CDSs within the horizontally-acquired islands [20] showed differential gene expression, suggesting that such CDSs are less likely to have been incorporated into the QS regulon than those on the chromosome backbone. qRT-PCR was used to study a number of genes in the expI mutant and wild type to examine differential gene expression, either to verify changes observed in the microarray or to examine the effects of a mutation in expI on additional genes (Table S1). Importantly, qRT-PCR analysis of selected genes after growth of the expI mutant and wild type in vitro revealed the same pattern of expI-dependence as observed in vivo, and these changes could be fully complemented by the addition of exogenous OHHL (Fig. 2).
The role of QS in pathogenesis of pectobacteria has been intensively studied in vitro, particularly for its ability to co-ordinately up-regulate PCWDEs [7]–[10],[12],[14],[28]. Previous work based on enzyme plate assays observed that all major groups of PCWDEs, including pectate lyases (Pel), cellulases (Cel), protease (Prt), pectin lyase (Pnl) polygalacturonase (Peh) and pectin methyl esterase (Pme) were under QS control [8]. In this study, we found that genes encoding all these groups showed lower transcript abundance in the expI mutant compared to the wild type at both 12 and 20 hpi (Table 1, Table S1). The major pectate lyases PelA, PelB (ECA4068) and PelC (ECA4069), as well as CelV (ECA1981), a putative cellulase ECA2220, PrtW (ECA2785) and PehA (ECA1095) have previously been associated with QS in pectobacteria, either through transcriptional or proteomic analyses [7],[10],[12],[28]. However, in addition the transcription (using microarray and/or qRT-PCR analyses) of genes encoding other PCWDEs and their isoforms, including “minor” pectate lyases (PelZ [ECA4070], Pel-3 [ECA1094], PelB and PelW [ECA2402]), CelB (ECA2827) and CelH (ECA3646), PehN (ECA1190), PmeB (ECA0107) and Pnl (ECA1499) was found to be expI-dependent. These results confirm previous observations of QS regulatory control in vitro and validate our in planta approach.
Other genes previously shown to fall under QS control in vitro, including svx, nip and a gene of unknown function (ECA3946), as well as three regulators (expR [ECA0106], rsmA [ECA3366] and virR [ECA1561]) involved in the production of PCWDEs [9], [10], [12]–[14], also showed reduced transcript abundance in the expI mutant compared to the wild type strain. This again justifies our approach in assessing the genome-wide effects of QS regulation during the potato interaction. While 1167 genes, representing a variety of processes, were found to be differentially expressed in the microarray experiment (Table S2), we focus predominantly on those that display reduced transcript levels in the mutant (as these are presumably induced directly or indirectly by QS), and which also have a known or putative role in virulence (Table 1).
To successfully cause disease Pba must secrete a multitude of PCWDEs and other proteins, many of which are under QS control. We observed that both Type I and Type II secretion systems (T1SS and T2SS, respectively), which can be considered as ‘accessory virulence factors’, are modulated by QS (Table 1). Prior to this study the secretion systems responsible for the delivery of these virulence factors had not been reported as QS-regulated, and this observation indicates a novel facet to QS control of pathogenesis in pectobacteria.
The T2SS is well characterised in pectobacteria and is responsible for secretion of many key virulence factors, e.g. Pel, Cel and Svx [10]. The T2SS of pectobacteria is encoded by a cluster of 15 out genes (ECA3098-3110 and ECA3113-3114) [20],[29], of which six (outMLHGFD) (by microarray analysis) exhibited reduced transcript abundance levels in the expI mutant (Table 1). Analysis of these and seven other out genes by qRT-PCR confirmed that all were expressed at a lower level in the expI mutant (Table S1), implying that expression of the Out T2SS is up-regulated by QS in vivo. Similar QS modulation of out expression was also demonstrated by both qRT-PCR (Table S1) and the use of an outD-gusA reporter fusion in vitro (Fig. 3). In the latter experiment, expression of outD (ECA3109) was reduced in the expI mutant and restored to wild type levels by the exogenous addition of OHHL, confirming QS modulation of out gene expression.
Regulation of the major secreted protease, PrtW, is QS-dependent in Pba [10]. Secretion of Prt by the PrtDEF T1SS is well-characterised in Dickeya dadantii (formerly Erwinia chrysanthemi) [30] and, by analogy, PrtW is expected to be secreted by the T1SS encoded by the neighbouring prtDEF (ECA 2781-2783) genes in Pba. To support this, the microarray data indicated that transcription of the T1SS genes prtDF was reduced in the expI mutant. Use of qRT-PCR confirmed that expression of all three T1SS genes, prtDEF, was significantly reduced in the expI mutant compared to the wild type (Table S1). QS-dependence of the T1SS and T2SS is a logical accompaniment to the simultaneous QS-dependent induction of their substrates, presumably allowing the systems to cope efficiently with the greatly increased quantity of these substrates. Examples of QS-modulated secretion have been reported previously in other pathogens, e.g. the Xcp T2SS of Pseudomonas aeruginosa and the Lip T1SS of Serratia marcescens [31],[32], although this is the first time that QS-dependant secretion systems have been described in pectobacteria.
As well as physically attacking the plant cell wall through the action of PCWDEs, in Pba1043 the Type III secretion system (T3SS) is also necessary for full virulence [5]. The T3SS is found in many Gram-negative pathogens of both animals and plants and is used to translocate effector proteins into host cells, where they manipulate host defences. Helper proteins (or harpins) are secreted to the extracellular environment, and may assist in effector translocation [33]. We observed that expression of the T3SS structural, putative effector and helper genes, and Type III-associated regulators were all modulated by QS. In Pba1043, and other pectobacteria, the T3SS is encoded by the hrp cluster, composed of around 40 CDSs. These CDSs encode components of the structural apparatus, as well as the putative effector DspA/E [ECA2113], and helpers HrpN [ECA2103] and HrpW [ECA2112]. The Pba1043 hrp cluster also contains a group of CDSs (ECA2104-ECA2110), which includes a number of lipoproteins, that appear to be absent in closely-related species [5]. ECA2104 shows homology to vgrG and is described below. In the microarray experiment, two CDSs hrpE [ECA2097], associated with the Type III structural apparatus, and a putative lipoprotein (ECA2108) exhibited decreased transcript abundance in the expI mutant (Table 1). qRT-PCR analysis of these and an additional 17 CDSs subsequently confirmed that CDSs encoding the Type III structural apparatus, the putative effector dspE, helpers hrpN and hrpW, regulators hrpL, hrpS and hrpY, and all CDSs between ECA2104 and ECA2110 were significantly reduced in the expI mutant compared to the wild type, predominantly at 12 h (Table S1). Either positive or negative QS regulation of the T3SS has been observed in other pathogens, e.g. Pseudomonas aeruginosa [34], Vibrio harveyi [35], enteropathogenic E. coli [36], Ralstonia solanacearum [37], and QS regulation of hrpN has been shown in Pba [8]. However, this is the first published evidence that QS plays a role in regulating the entire T3SS and its effectors in the enterobacterial plant pathogens, indicating that co-ordinated physical (PCWDEs) and stealth (T3SS) attacks may be necessary for successful disease development.
Recently, a novel T6SS was described and implicated in pathogenicity in Vibrio cholerae and P. aeruginosa [23],[38]. In V. cholerae, the system is encoded by the VAS locus, genes VCA0107-VCA0123. This locus is one member of a group of conserved gene clusters that are conserved in several pathogens. In both V. cholerae and P. aeruginosa, the T6SS is required for secretion of HcpA and VgrG proteins, although whether these represent putative effectors or simply secreted components of the secretion machinery is not yet clear [23],[38]. In Pba1043, the locus ECA3445–ECA3427 is predicted to encode a VAS-like T6SS and its putative substrates, since these genes encode proteins very similar to those encoded by VCA0107-VCA0123 and is similarly arranged on the chromosome. Microarray analysis indicated that 11 of the 18 genes were expressed at significantly lower levels in the expI mutant (Table 1), and so transcription of the T6SS also appears to be QS-dependent. The modulated genes included Pba homologues of VCA0120, VCA0116 and VCA0110, which are required for Type VI secretion in V. cholerae and/or P. aeruginosa [23],[38]. Moreover, the expression of several predicted T6SS substrates, i.e. encoded by hcpA and vgrG-like genes, was also found to be QS-dependent (Table 1). There are seven hcpA homologues in Pba, three of which (ECA4275[hcp1], ECA3428[hcp2] and ECA2866[hcp3]) are highly similar [20],[27]. ECA3428 and ECA4275 are sufficiently similar that it was not possible to design probes specific to each locus. Nevertheless, the probe detecting expression of both these genes showed decreased transcript abundance in expression in the expI mutant, indicating QS-dependent regulation. Expression of ECA2866 and four other homologues (ECA0456[hcp4], ECA3672, ECA0176 and ECA4277) was also decreased (Table 1). A combination of microarray analysis and qRT-PCR indicated reduced transcript abundance in the expI mutant of all five vgrG homologues, ECA2867, ECA3427, ECA2104, ECA4142 and ECA4276 in the Pba1043 genome (Table S1). Previous work showed that Hcp1-4 and a VgrG homologue (ECA3427) were found in the secretome of Pba1043. Over-expression of Hcp1 increased virulence, suggesting that this and related proteins are virulence factors in pectobacteria [27]. The VAS-like T6SS genes, ECA3445-ECA3427, appear to constitute an operon that may extend for a further seven CDSs (ECA3426-ECA3420). Of these, expression of six was reduced in the expI mutant (Table S2), raising the possibility that they may encode T6SS-dependent effectors.
As the T6SS is clearly important for virulence in other pathogens, and a predicted substrate (Hcp1) affects virulence in Pba, we investigated whether the putative T6SS plays a role in virulence in Pba. Mutants in ECA3438 and ECA3444, when tested in potato stem and tuber virulence assays, both showed significantly reduced virulence compared with the wild type (Fig. 4). In tuber tests, complementation of the mutants in trans was shown to return virulence to wild type levels (Fig. 4B). Our results indicate, for the first time in any pathogen, a role for QS in the regulation of the T6SS and its putative substrates. It also demonstrates that the T6SS in Pba plays a role in pathogenesis, which appears to act in conjunction with PCWDE, the T3SS and other virulence determinants during the QS process.
Microarray analysis revealed the QS-dependent differential expression of at least 79 CDSs with either known or putative regulatory functions (Table S2). Twelve CDSs, five of which showed enhanced (hexA [ECA3030], kdgR [ECA2425], phoP[pehR] [ECA2445], rdgA [ECA2435] and rsmA) and seven of which showed reduced (aepA [ECA1022], expA [ECA2882], expR, hexY [ECA0809], hor [ECA1931], rexZ [ECA4123] and virR) transcript abundance in the expI mutant, are known to regulate PCWDEs production and are required for full virulence in pectobacteria [4], [12]–[14],[39] (Table 1). However, only three (expR, rsmA and virR) have previously been shown, in vitro, to fall under QS control [12]–[14]. Three CDSs (hrpL [ECA2087], hrpY [ECA2089] and hrpS [ECA2090]) involved in the regulation of the T3SS in pectobacteria and other phytopathogens [40] also showed decreased transcript abundance in the expI mutant. As all 15 of these CDSs are QS-dependent, this places QS at the apex of a regulatory hierarchy controlling both PCWDEs and the T3SS with its cognate effector proteins. Other QS-controlled regulators are also likely to be important during interaction with the plant (see below). Although QS is central to pathogenesis, elucidating the hierarchical relationships between “subordinate” regulators presents a particular challenge due to the lack of data on such relationships in this particular strain. Several virulence regulators in pectobacteria are known to operate though the Rsm system, which plays a major role in controlling virulence [14]. While not investigated as part of this work, it is highly likely that at least some of the regulators identified in this study operate through this system. Nevertheless, we have still been able to add considerable new information to existing regulatory models [41] and propose an extended model for virulence in the pectobacteria (Fig. 5).
In addition to regulators previously characterised in pectobacteria, differential expression of 18 further CDSs were found that are similar to a diverse range of transcriptional regulators in other bacteria (Table S2). These include CDSs with putative regulatory functions in nitrogen signal transduction and assimilation (citB [ECA2578], glnB [ECA3254], nac [ECA4483]), hydrogenase activity (hypA [ECA1235]), oxygen sensing (fnr [ECA2207]), defence against superoxides and other stress responses (ohrR [ECA3168], phoB [ECA1110], recX [ECA3368], rseB [ECA3282], rseC [ECA3281]), motility (flgM [ECA1700], fliZ [ECA1740]) and survival in soil (sftR [ECA4305]) (Table S2). Three of these additional regulators (fliZ, ohrR and rscR) have been implicated in virulence in other bacterial pathogens (Table 1) [42]–[44]. However, it does not necessarily follow that homologous regulatory proteins in bacteria are responsible for regulation of homologous processes [45].
Many CDSs encoding putative regulators of unknown function were shown to be regulated by QS. These CDSs thus represent novel candidates for virulence factors. Expression of one such CDS, ECA1562, subsequently named virS, was enhanced in the expI mutant at 12 and 20 hpi and is thus proposed to be repressed by QS (Table 1). VirS is a predicted TetR-family transcriptional regulator whose target(s) is unknown, although its closest reported homologue is a TetR family regulator, TvrR, implicated in virulence in the plant pathogen Pseudomonas syringae pv. tomato [46]. virS is located adjacent to the gene encoding a key QS-controlled regulator, VirR (ECA1561, [12]). However, inactivation of virS does not affect transcription of virR (data not shown). In order to determine whether virS plays a role in virulence, a defined virS mutant was constructed and tested in stem and tubers virulence assay. The virS mutant showed significantly reduced lesion formation compared with the wild type (Fig. 4) and is thus a novel virulence factor in Pba. In tuber tests, complementation of the mutant in trans returned virulence to wild type levels (Fig. 4B). The precise role of virS in planta is under investigation.
The microarray data revealed a small reduction in expression of genes cfa2 (ECA0607) and cfa8A (ECA0601) in the expI mutant compared to the wild type (Table 1). These genes are of particular interest as they are part of a cluster responsible for the synthesis of coronafacic acid (CFA) which, in Pseudomonas syringae, is a component of the phytotoxin coronatine [47]. We showed previously that mutations in this cluster (cfa6 [ECA0603] and cfa7 [ECA0602]) significantly reduce pathogenicity of Pba1043 on potato stems [20]. Transcriptional changes in cfa2, cfa6 and cfa7, compared to a QS up-regulated (pelA) control, were thus examined at 12 and 20 hpi using qRT-PCR. At both time-points, pelA, the cfa genes, and the cfl (ECA0609) gene (involved in the formation of coronafacoyl conjugates by ligation of amino acids to CFA) showed reduced expression in the expI mutant, indicating that they are all under QS control (Table S1).
Salicylic acid (SA) and jasmonic acid (JA) are signalling molecules that play major roles in the activation of plant defences against pathogen attack [48]. CFA and its amino acid conjugates appear to act as structural and functional analogues of JA and its conjugates [49]. Recent work by Uppalapati et al. [21] showed that Pseudomonas syringae DC3000 mutants lacking CFA and/or coronatine were impaired in their ability to persist in tomato plants at the later stages of infection, and that the ability to persist coincided with the activation of JA-based, and concomitant suppression of SA-based, defences. It is hypothesised that, through this suppression of SA-mediated defences, coronafacoyl conjugates may aid P. syringae to enter the necrotrophic phase of infection and promote disease symptoms. It would appear therefore that Pba, through QS, synthesises CFA and coronafacoyl conjugates co-ordinately with multiple PCWDEs, the T3SS and T6SS in a synchronised assault on the plant as it progresses from biotrophy to necrotrophy. Although the effect of Pba-encoded CFA conjugates on plant defences has yet to be determined, such a two-pronged attack may be necessary for Pba to establish disease. It will be interesting to determine whether QS plays a similar role in P. syringae and related pathogens.
QS regulation in pectobacteria was observed originally in vitro through dramatic impacts on PCWDE production, pathogenesis and (in Pcc) carbapenem antibiotic production [16]. The microarray analysis of global gene expression in planta presented here indicates a far broader physiological impact of QS, uncovering effects on the expression of many other genes associated with pathogenesis, and on other physiological processes not necessarily connected to plant pathogenesis. As QS is AHL concentration-dependent, its impact is likely to be greatest towards the latter stages of infection, where large quantities of PCWDEs are induced to attack plant cells and the characteristic soft rot disease symptoms occur [4]. Correspondingly, we find that production of the T1SS and T2SS, which are involved in the secretion of PCWDEs to the extracellular environment, are also under QS control.
A very small number of virulence regulators have previously been shown to fall under QS control. Our study has added over 70 other regulators to this list, including the major known virulence regulators associated with PCWDE production in pectobacteria. An important inference from these microarray analyses is that the QS control system occupies a critical position in the regulatory hierarchy and that multiple downstream regulators, some which may operate through the Rsm system [14], are under QS control. Furthermore, QS is seen to have both positive (activation) and negative (repression) effects on its downstream targets. Our knowledge of the hierarchical chain of command in control of the complex regulatory systems of PCWDE and other virulence factors is fragmentary, in part because the current literature describes experimental data derived from multiple strains of Pba, Pcc, and Dickeya spp. It may not therefore be completely legitimate to assume that the identified regulators play conserved roles in these different bacterial strains [45]. Nevertheless, while accepting this caveat, our results are consistent with the notion that, within the infected potato plant, QS acts as a key “master regulator” sensory system in this phytopathogen (Fig. 5). Additionally, regulators associated with virulence in other bacteria, and many novel putative regulators have also been identified; including VirS, which has been associated with virulence in this study.
In addition to the T1SS and T2SS, we have identified a T6SS in Pba and shown, for the first time in a plant pathogen, that it has a role in virulence. Moreover, we have described the first example of a QS-controlled T6SS in any pathogen. The precise functional roles of Hcp, VgrG and other possible Type VI substrates is unknown, but their proposed functions as effector proteins may be important for manipulating host defences whilst PCWDEs mount a simultaneous physical attack on plant cell walls. This does appear to be the case for both the T3SS and associated effectors, and coronafacoyl-amide conjugates, which are similarly QS-dependent, and consequently may suppress or otherwise manipulate defences. This has important implications for the infection process in pectobacteria, as it suggests that these pathogens do not infect merely by “brute force”, where the action of PCWDEs alone is sufficient to overwhelm plant defences and break down plant cell walls towards the end of infection. It seems increasingly likely that, in conjunction with PCWDEs, the production of virulence determinants that actively suppress plant defences, may be necessary to facilitate the transition from biotrophy to necrotrophy during disease development.
Pba1043 [20], and strains with mutations in expI, ECA3438, ECA3444 and virS were used in this study. The expI mutant was derived from phage M1-mediated transduction of expI::mTn5gusAgfp from mutant MC3 into the wild type strain [10]. Mutants ECA3438 and ECA3444 were isolated from a mutation library of Pba1043 [5]. For inactivation of virS, 1085 bp of virS and surrounding regions were PCR-amplified using primers SC51 (ATTTGGATCCGTTGTTCCTGTTCTGTCG) and SC52 (TATATCTAGAGTTTACTGAGCAAGCGACG) and cloned into pBluescript-II KS+ using BamHI-XbaI sites. The KnR cassette from pACYC177 (NEB) was cloned into the NsiI site in the middle of virS. The resulting virS::KnR fragment was then cloned into the suicide vector, pKNG101 [50], generating the marker-exchange plasmid. The plasmid was introduced into Pba1043 by conjugation and transconjugants, resulting from integration of the suicide plasmid into the chromosome by homologous recombination, were selected by ability to grow on minimal medium containing 0.2% glucose+streptomycin. Following overnight growth in the absence of antibiotic selection, exconjugants, in which resolution of the plasmid from the chromosome leaving only the disrupted allele had occurred, were selected by ability to grow on minimal medium containing kanamycin+10% sucrose as sole carbon source and inability to grow on streptomycin. The disruption of the locus was confirmed by PCR analysis and DNA sequencing. All strains were maintained on Luria Bertani (LB) agar supplemented with kanamycin (50 µg/ml) and, unless stated otherwise, were cultured in 10 ml LB broth at 27°C overnight with aeration. Mutations were transduced into a clean Pba1043 background using phage M1 [51]. Pathogenicity tests were performed both on potato stems and tubers [19]. Approx. 102 and 104 cells per inoculation site were used for stems and tubers, respectively. Complementation of mutant strains was carried out in trans following cloning of ECA3444, ECA3438 and expI into plasmid pGEM-T (Promega, Southampton, UK) and virS into pQE80-L (Qiagen, Crawley, UK) with their own ribosome binding sites. OHHL-producing transgenic potato plants used for in planta OHHL complementation are as described [19]. GENSTAT for Windows was used for statistical analyses [20].
Wild type and expI mutant strains were grown in 10 ml LB broth at 27°C with aeration to stationary phase (approx. 1.0×109 cells/ml) and re-suspended in 10 mM Mg SO4 prior to inoculation into sterilized potato tubers (2×107 cells/ml into cv Maris Piper). The tubers were then wrapped in cling film and placed in a tray with wetted tissue to retain high humidity before incubation at 19°C in the dark. At 12 and 20 hours post inoculation (hpi), the bacterial cells were isolated from the tuber by scraping infected tissue into sterilised water. Starch was removed by centrifugation twice at 1000 rpm for 1 min. The bacterial cells in the supernatant were transferred to RNA stabilization buffer containing 1% phenol (pH 4.3, v/v) and 20% ethanol (v/v) and incubated on ice for at least 30 min. Total RNA was isolated using the SV Total RNA Isolation System (Promega) as described by the manufacturer and quantified using a NanoDrop ND-100 spectrophotometer (NanoDrop Technologies, Wilminton, DE). The quality of RNA was analyzed using an Agilent Bioanalyzer 2100 electrophoresis system (Agilent Technologies Inc., West Lothian, UK). In total 12 µg RNA was reverse transcribed and cDNA labelled [52]. 60-mer oligonucleotide probes were designed to Pba CDSs and used, together with controls, to generate 11K custom arrays with 99.5% genome coverage (Agilent, Inc., Santa Clara, CA, USA) [53]. Microarrays were carried out in triplicate for each time point [52],[53]. All microarray images were visually assessed for quality prior to feature extraction, whereby standard probe QC standards were applied (see further information in ArrayExpress-http://www.ebi.ac.uk/microarray-as/aer/). Features flagged as poor were removed prior to importing into Genespring software. Box plots and principle components analysis of whole datasets were used to assess array to array variation. Any outlying microarrays were repeated as necessary.
Microarray data were analysed using GeneSpring software (version 7.2) and normalized using the Lowess algorithm (Agilent Technologies Inc.). Gene expression was considered to be different between the wild type and expI mutant strains for a probe if there was at least 1.5 fold change in normalised hybridisation score, and that change showed a statistically significant (Student's t test: P value <0.05) difference in their normalized data. Microarray data were submitted to the ArrayExpress repository (http://www.ebi.ac.uk/microarray-as/aer/), submission E-TABM-384, including details of the SCRI Pectobacterium atrosepticum 11k array (submission A-MEXP-942).
qRT-PCR was performed using recA as an endogenous control to validate differences in expression of genes identified from the microarray experiment, and to test additional genes in the Pba genome. RNA samples were analysed in triplicate. 5 µg total RNA was used to synthesize cDNA and 1 µl diluted template DNA (1:10) was used in a reaction of 25 µl containing 1x SYBR Green PCR Master Mix (Qiagen) and 10 pmol of the appropriate primers. qRT-PCR data were analysed using the Relative Expression Software Tool [REST] 2005 (Corbett Life Science, Cambridge, UK). Selected mutants were complemented by the addition of OHHL (Sigma) in DMSO (or with DMSO alone as a control) to PMM media at a final concentration of 5 µM, and strains grown to a final cell density of 5×107 cfu/ml. After 18h incubation, bacteria were harvested and RNA extracted, purified and quantified as previously described. Differential expression was considered statistically significant if the t-test P-value was <0.05. To analyse gene expression of the out gene cluster in vitro, cultures were grown in Pel Minimal Medium (PMM) at 27 °C, RNA samples were prepared and qRT-PCR analysis performed as described in Burr et al., [12]. The outD-gusA strain, MC4, was as described by Corbett et al., [10], an outD-gusA/expI double mutant was generated by generalised transduction (data not shown), and β-glucuronidase (GusA) activity was measured throughout growth in Pel minimal broth (PMB) as described [10]. All primers used are described in Table S1.
Bacterial cells for counting were collected in 10 ml sterile water prior to dilution and plating as described [54]. Bacterial cells used for measuring OHHL levels were taken from these samples prior to dilution. OHHL levels were analysed using E. coli JM109 carrying a bioluminescence reporter vector (pSB401) [55]. Each sample of 100 µl was aliquoted into three wells of a sterile black 96-well microtitre plate, and 100 µl of the sensor strain (grown to an OD600 of 1.0) was added to each well. The microtitre plate was incubated at 37°C for 3 hours and the luminescence from each well was measured using a SpectraMax M5 luminescence plate reader at the default setting with an integration time of 1 second (Molecular Devices Corp., Sunnyvale, CA). A series of OHHL standards was used both as a positive control and to determine the level of OHHL.
ECA0105 (expI), YP048233; ECA0106 (expR), YP048234; ECA0176, YP048830; ECA0456 (hcp4), YP048574; ECA0601 (cfa8A), YP048718; ECA0602 (cfa7), YP048719; ECA0603 (cfa6), YP048720; ECA0607 (cfa2), YP048724; ECA0609 (cfl), YP048726; ECA0809 (hexY), YP048920; ECA0931 (svx), YP049040; ECA1017 (pmeB), YP049124; ECA1022 (aepA), YP049129; ECA1094 (pel-3), YP049200; ECA1095 (pehA), YP049201; ECA1110 (phoB), YP049216; ECA1190 (pehN), YP049296; ECA1235 (hypA), YP049341; ECA1499 (pnl), YP049604; ECA1561 (virR), YP049663; ECA1562 (virS), YP049664; ECA1700 (flgM), YP049801; ECA1740 (fliZ), YP049840; ECA1931 (hor), YP050028; ECA1981 (celV), YP 050075; ECA2087 (hrpL), YP050182; ECA2089 (hrpY), YP050184; ECA2090 (hrpS), YP050185; ECA2097 (hrpE), YP050792; ECA2103 (hrpN), YP050198; ECA2104 (vgrG), YP050199; ECA2108, YP050203; ECA2105-2110, YP050200-050205; ECA2112 (hrpW), YP050207; ECA2113 (dpsA/E), YP050208; ECA2207 (fnr), YP050300; ECA2220, YP050313; ECA2402 (pelW), YP050497; ECA2425 (kdgR), YP050520; ECA2435 (rdgA), YP050530; ECA2445 (pehR), YP050539; ECA2553, YP050644; ECA2578 (citB), YP050669; ECA2724 (rscR), YP050815; ECA2781 (prtF), YP050872; ECA2782 (prtE), YP050873; ECA2783 (prtD), YP050874; ECA2785 (prtW), YP050876; ECA2827 (celB), YP050918; ECA2866 (hcp3), YP050957; ECA2867 (vgrG), YP050958; ECA2882 (expA), YP050973; ECA3030 (hexA), YP051120; ECA3087 (nip), YP051177; ECA3098-3114, YP051188-051204; ECA3100 (outM), YP051190; ECA3101 (outL), YP051191; ECA3105 (outH), YP051195; ECA3106 (outG), YP051196; ECA3107 (outF), YP051197; ECA3109 (outD), YP051199; ECA3168 (ohrR), YP051257; ECA3254 (glnB), YP051343; ECA3281 (rseC), YP051370; ECA3282 (rseB), YP051371; ECA3366 (rsmA), YP051455; ECA3368 (recX), YP051457; ECA3420, YP051511; ECA3421-3426, YP051512-051517; ECA3427, YP051518; ECA3428 (hcp), YP051519; ECA3430, YP051520; ECA3432 (vasK), YP051522; ECA3433, YP051523; ECA3436 (vasG), YP051526; ECA3438, YP051528; ECA3440, YP051530; ECA3442 (vasA), YP051532; ECA3443, YP051533; ECA3444, YP051534; ECA3445, YP051535; ECA3427-3445, YP051518-051535; ECA3646 (celH), YP051234; ECA3672, YP051760; ECA3946, YP052033; ECA4067 (pelA), YP052154; ECA4068 (pelB), YP052155; ECA4069 (pelC), YP052156; ECA4070 (pelZ), YP052157; ECA4123 (rexZ), YP052210; ECA4142 (vgrG), YP052229; ECA4275, YP052362; ECA4276 (vgrG), YP052363; ECA4277, YP052364; ECA4305 (sftR), YP052392; ECA4483 (nac), YP052566. |
10.1371/journal.ppat.1004061 | The Pathogenic Mechanism of the Mycobacterium ulcerans Virulence Factor, Mycolactone, Depends on Blockade of Protein Translocation into the ER | Infection with Mycobacterium ulcerans is characterised by tissue necrosis and immunosuppression due to mycolactone, the necessary and sufficient virulence factor for Buruli ulcer disease pathology. Many of its effects are known to involve down-regulation of specific proteins implicated in important cellular processes, such as immune responses and cell adhesion. We have previously shown mycolactone completely blocks the production of LPS-dependent proinflammatory mediators post-transcriptionally. Using polysome profiling we now demonstrate conclusively that mycolactone does not prevent translation of TNF, IL-6 and Cox-2 mRNAs in macrophages. Instead, it inhibits the production of these, along with nearly all other (induced and constitutive) proteins that transit through the ER. This is due to a blockade of protein translocation and subsequent degradation of aberrantly located protein. Several lines of evidence support this transformative explanation of mycolactone function. First, cellular TNF and Cox-2 can be once more detected if the action of the 26S proteasome is inhibited concurrently. Second, restored protein is found in the cytosol, indicating an inability to translocate. Third, in vitro translation assays show mycolactone prevents the translocation of TNF and other proteins into the ER. This is specific as the insertion of tail-anchored proteins into the ER is unaffected showing that the ER remains structurally intact. Fourth, metabolic labelling reveals a near-complete loss of glycosylated and secreted proteins from treated cells, whereas cytosolic proteins are unaffected. Notably, the profound lack of glycosylated and secreted protein production is apparent in a range of different disease-relevant cell types. These studies provide a new mechanism underlying mycolactone's observed pathological activities both in vitro and in vivo. Mycolactone-dependent inhibition of protein translocation into the ER not only explains the deficit of innate cytokines, but also the loss of membrane receptors, adhesion molecules and T-cell cytokines that drive the aetiology of Buruli ulcer.
| Buruli ulcer is a progressive necrotic skin lesion caused by infection with the human pathogen Mycobacterium ulcerans. Mycolactone, a small compound produced by the mycobacterium, is the root cause of the disease pathology, but until now there has been no unifying mechanism explaining why. We have been using a model system to investigate the reason for the selective loss of protein that is a common feature of mycolactone exposure. Specifically, this involves identifying the point at which it stops immune cells making inflammatory mediators. In this work, we demonstrate that mycolactone inhibits production of such proteins by blocking the first step of protein export: translocation into a cellular compartment called the endoplasmic reticulum (ER). Proteins due for export are instead made in the cell cytosol where they are recognised as being in the wrong place and are rapidly degraded, causing a general cessation of the production of proteins that have to travel through the ER, including almost all secreted and surface proteins. This has a profound effect on basic cell functions such as growth, adhesion and survival. Therefore, we have identified the molecular basis underlying the key features of Buruli ulcer, and this will transform our understanding of disease progression.
| Mycolactone is a lipid-like polyketide macrolide virulence factor produced by Mycobacterium ulcerans, the infectious agent of Buruli ulcer (BU) [1], [2]. This progressive, necrotizing, cutaneous lesion is common in West Africa but also found in other regions, including Australia, Asia and South America. Mycolactone is a key factor in BU pathology: possession of a plasmid carrying enzymes involved in mycolactone synthesis is essential for virulence and injection of mycolactone alone can reproduce many characteristics of the infection, including ulceration, necrosis and suppression of immune responses [1], [3]. Mycolactone has been shown to have diverse effects on a range of cells and tissues but a unifying mechanism underlying its pleiotropic actions has remained elusive.
In vitro, exposure to pure mycolactone is cytotoxic for many cell lines, but the dose and exposure required is highly variable ([4] and references therein) and primary immune cells (including T-cells, monocytes and macrophages) are considerably more resistant [5], [6]. While early evidence from cell lines implicated G1/G0 growth arrest and apoptosis [7], recent work showed that a more likely mechanism driving cell death in vivo is anoikis due to direct binding of mycolactone to the Wiskott-Aldrich Syndrome Protein (WASP), leading to inappropriate activation of WASP and relocalisation of the actin nucleating complex Arp2/3 [8]. This disrupts the cytoskeleton, altering cell adhesion and migration. Detachment of monolayer cells is a common feature of the mycolactone response and precedes cell death by up to 48 hours.
One of the most striking characteristics of BU lesions is an almost complete absence of inflammation despite extensive tissue damage. In ulcerated lesions, where large amounts of mycolactone are produced by foci of extracellular bacilli, inflammatory cell infiltration is limited to the periphery [9]–[11]. Infection is accompanied by alterations in local and systemic immune responses in which mycolactone plays a central role [11]–[14], via direct and indirect effects on T-cells, dendritic cells, monocytes and macrophages [5], [15]–[17]. Mycolactone interferes with T-cell activation, down-regulating expression of the T-cell receptor and reducing IL-2 production in response to activating signals [15], [17], [18]. Lymphocyte homing is also impaired due to suppression of L-selectin and LFA-1 levels, leading to a dramatic depletion of T-cells in peripheral lymph nodes [6]. In monocyte-derived dendritic cells, mycolactone inhibits the production of costimulatory molecules (such as CD40 and CD86). In addition, secretion of various cytokines and chemokines is blocked and mycolactone treated dendritic cells show a reduced ability to activate T-cells [16].
The innate immunity provided by monocytes and macrophages is also suppressed by mycolactone. Tissue resident macrophages normally play a central role in mycobacterial infections. However, M. ulcerans differs from other pathogenic mycobacteria in that, except in very early infection, the vast majority of bacilli are not found within the host macrophage but are located extracellularly. Mycolactone inhibits key macrophage responses such as nitric oxide production and phagocytosis as well as phagosome maturation and acidification [2], [4], [19]. In addition, mycolactone prevents the induction of many proteins essential for driving inflammation, including TNF, other cytokines/chemokines (for example, IL-6, IL-8 and IP-10), and further inflammatory mediators (such as the prostaglandin synthetase Cox-2) [5], [10], [15].
There is good evidence that mycolactone diffuses through the lesion in advance of the proliferating bacilli and the necrotic centre (see for example [20]). Therefore, understanding exactly how this compound mediates its diverse immunosuppressive and cytotoxic effects on cells surrounding the developing lesion is crucial. As outlined above, many of these effects involve loss of expression of specific proteins, both induced and constitutive, such as inflammatory mediators. Consequently, the same molecular mechanism that prevents inflammatory protein production in the macrophage may also explain the inadequate protein production more generally. This makes it an excellent model system with which to examine the basic cell biology of mycolactone function, since the response is inducible by nature and it is therefore straightforward to separate new protein synthesis from baseline levels.
We have previously shown that inducible inflammatory mediator production is inhibited by a post-transcriptional mechanism, since mycolactone does not modulate the LPS-dependent activation of ERK, JNK, p38 MAPK or NFκB and induced levels of mRNA are maintained or even enhanced [5]. However there is no significant decrease in total protein synthesis, nor are phosphorylation patterns of Akt, p70S6K, eIF4E and eIF2α changed; a finding confirmed in another model system, Jurkat T-cells [17]. In the current manuscript we demonstrate conclusively that mycolactone does not selectively inhibit translation as predicted [2], [5], and instead blocks co-translational translocation into the ER. This leads to the rapid degradation of mislocalised proteins in the cytosol and hence loss of detectable expression. We show that the production of nearly all new glycosylated and secreted proteins ceases following mycolactone exposure, not only in macrophages but in fibroblasts, epithelial and endothelial cells. This mechanism therefore provides the necessary explanation for many of the pleiotropic effects of this unique molecule and accounts for much of the underlying disease pathology.
In order to establish the dose of synthetic mycolactone A/B required to completely inhibit the production of TNF in RAW264.7 cells, we carried out a dose response (Fig. 1A). It was determined that the effective dose was 125 ng/ml, and this also prevented LPS-dependent Cox-2 production without affecting cell viability (Fig. S1A). This dose is marginally higher than required for inhibition of TNF production by natural mycolactone A/B in primary human macrophages (Fig. S1B), probably reflecting the known variation in sensitivity between different cell types, preparations of mycolactone (natural vs. synthetic) and/or target activities (immunosuppressive vs. cytotoxic). We then performed polysome profiling of macrophages to investigate whether mycolactone selectively inhibits the translation of inflammatory mediators. This technique allows the association of TNF, IL-6 and Cox-2 transcripts with actively translating polysomes in various experimental conditions to be assessed. RAW264.7 cells were used because, in preliminary experiments, the low mRNA yields and high RNase content of primary human monocytes and macrophages precluded the use of these cells (data not shown). The post-transcriptional mechanism of mycolactone-dependent inhibition of cytokine production observed in primary cells is conserved in this cell line (Fig. S1C, performed as a control experiment for all profiles obtained).
Mycolactone exposure was found to consistently cause a change in the shape of the polysome profiles, associated with an increase in the size of the 60S peak and change in the profile in the area associated with heavy polysomes in both unstimulated (Fig. 1B; MYC) and stimulated (Fig. 1C; LPS+MYC) cells. However, these changes occurred gradually over several hours, while the inhibition of TNF production is manifest as little as 20 min after LPS addition (data not shown). This suggests it may be a secondary, rather than primary, effect. Mycolactone alone did not influence the quantity or location of TNF mRNA (not shown) and LPS stimulation in itself did not induce any gross changes to the polysome profiles (Fig. 1C). In each profile, poly-A tract binding protein (PABP) and β-actin are used as control transcripts that confirm the location of unformed ribosomes and polysomes, respectively (Fig. 1C).
While unstimulated RAW264.7 cells expressed very little TNF mRNA, as expected, LPS stimulation led to increased abundance of TNF, IL-6 and Cox-2 mRNAs and their location moved so that a higher proportion of the mRNAs were in the polysomal fractions (Fig. 1C, compare ‘control’ and ‘LPS’ - and quantified in Fig. 1D), due to the known translational derepression that occurs following stimulation [21]. Again, as expected, neither the location of β-actin (known to be mycolactone insensitive [5]) or PABP were affected by mycolactone (Fig. 1B and D). However, in stark contrast to expectations, mycolactone had no effect on the polysomal association of any of the three inflammatory transcripts; all remained in heavy-polysomal fractions (Fig. 1B, compare ‘LPS’ and LPS+MYC’). When quantitated, the distribution of the mRNAs was very similar in the presence and absence of mycolactone (Fig. 1D).
We confirmed this unexpected finding in a number of ways. First, the localisation of these transcripts was assessed at various times after LPS stimulation to investigate whether the findings were influenced by the kinetics of the LPS response or time of mycolactone exposure (>1 hr), but this was found not to be the case (data not shown). Second, we examined the effects of short term exposure to two translation-inhibiting drugs on polysome profiles (Fig. 2A and B). Puromycin (PURO) causes premature termination and ribosomal release from translating mRNAs, whereas homoharringtonine (HH) prevents translation initiation leading to ribosome run-off of translating mRNAs) [22], [23]. Neither drug influenced the production of TNF or its inhibition by mycolactone (Fig. S2), but both caused a change in the profiles obtained from LPS stimulated cells, with HH being the more efficient (Fig. 2A). As expected, there was a concomitant change in the location of β-actin mRNA to monosomes (HH) or lighter polysomes (PURO) (Fig. 2B, left panel LPS, compare the black with the blue or green lines respectively). Cox-2 and TNF mRNAs also both moved into lighter polysomal fractions, confirming that our experimental system was sensitive to inhibition of translation.
When the action of these drugs on mycolactone treated cells was assessed, it could be seen that, while mycolactone altered the profiles but not the location of β-actin, Cox-2 or TNF transcripts as before (Fig. 2A and 2B, black lines), the response to PURO and HH was the same in the absence and presence of mycolactone. For Cox-2, PURO caused a similar ∼2-fraction shift (Fig. 2B, green lines), whereas HH causes a similar ∼1-fraction shift (Fig. 2B, blue lines) in both untreated and mycolactone treated cells. It is interesting to note that the shift in the Cox-2 peaks following HH treatments were smaller than that seen for β-actin, suggesting that Cox-2 is being translated more slowly (compare the blue lines in Fig. 2B, LPS, Actin and Cox-2). Both drugs had a less marked effect on TNF but a movement of the peak of mRNA recovery to lower fractions could be seen that was not prevented by mycolactone. This shows that all of the tested mRNAs are undergoing active translation in both the presence and absence of mycolactone and are not stalled on the ribosomes.
Finally, in an independent approach, the cellular localisation of proinflammatory mRNAs in the presence of mycolactone was also examined. Since TNF and IL-6 are secreted proteins and Cox-2, is ER-resident and contains a signal peptide, their actively-translating, nascent polypeptide chains should be directly associated with the ER due to the interaction of the signal peptide with the signal recognition particle (SRP) and Sec61 complex [24]. Cells were selectively permeabilised with digitonin to separate the cytosolic and digitonin-resistant ER membrane fractions. Western blotting showed the presence GAPDH protein in the cytosol while the ER-resident protein glucosidase I (GCS1) was confined to the membrane fraction (Fig. 2C). As seen by others, GAPDH mRNA was fairly evenly distributed between cytosolic and membrane fractions [25], [26], but the mRNAs for TNF, Cox-2 and IL-6 were all predominantly in the membrane fraction, even in the presence of mycolactone, indicating sufficient synthesis had occurred to allow signal peptide recognition (Fig. 2D). This data also strongly argues against an inhibition of proinflammatory mRNA translation as the mechanism underlying the loss of protein production due to mycolactone.
As proinflammatory protein synthesis is maintained in the absence of detectable protein levels, mechanisms by which these proteins might be targeted for degradation by the cell were investigated. Degradation by the 26S proteasome seemed a likely candidate. However, examining this experimentally is complex for inflammatory mediators since their transcriptional activation requires proteasome-dependent degradation of IκBα [27]. Cells were therefore stimulated with LPS for 2 hrs prior to addition of the proteasome inhibitor (PSI) for an additional 2 hrs. This did not decrease LPS-dependent production of TNF in cell supernatants (Fig. 3A) indicating that this experimental design was satisfactory.
Treatment of RAW264.7 cells with mycolactone resulted not only in a profound decrease in LPS-dependent Cox-2 production, but the barely detectable immunoreactive protein also had a lower mol wt (Fig. 3B, compare lanes 2 and 3), equivalent to that seen in cells exposed to the N-glycosylation inhibitor tunicamycin (TUN; Fig. 3B lane 4). Remarkably, Cox-2 production was found to increase when 26S proteasome activity was blocked in mycolactone-treated cells (Fig. 3B, compare lanes 3 [open arrow] and 7 [closed arrow]). When this effect was quantified by densitometry (Fig. 3C), then considered as fold-change within each experiment, the extent of restoration of protein production was 2.31±0.35-fold (mean±SEM, P = 0.02, n = 3). Notably, this protein was also unglycosylated (Fig. 3B, Cox-2 lane 7 [closed arrow]).
An even more striking observation was made for TNF, detected as nascent pro-TNF (26 kDa) within cell lysates. In LPS-stimulated cells pro-TNF is rapidly exported and so is barely detectable in cell extracts (Fig. 3B lane 2). However, while the low-level ER stress induced by TUN does not reduce TNF in cell supernatants (Fig. 3A) it does slows transit through the ER and Golgi allowing detection of pro-TNF in lysates [28] (Fig. 3B, lane 4). In the absence of PSI, no pro-TNF could be observed following mycolactone treatment, but levels of pro-TNF increased dramatically upon inhibition of 26S proteasome activity (Fig. 3B, compare lanes 3 [open arrow] and 7 [closed arrow]). When this effect was quantified by densitometry and analysed as for Cox-2 (Fig. 3C) this equated to a mean fold-increase of 12.24±6.6-fold (mean±SEM, P = 0.0002, n = 3). This is the first time that the inhibition of cellular TNF and Cox-2 production has been overcome in the continuous presence of inhibitory concentrations of mycolactone, allowing detection of the previously undetectable protein. This transformative finding supports a mechanism for mycolactone action in which such proteins (including, but not restricted to, TNF and Cox-2) are being destroyed by proteosomal degradation in the cytosol.
The lack of glycosylation of Cox-2 and the failure of PSI treatment to cause TNF secretion (Fig. 3A), suggested the restored proteins could not gain access to the ER. Indeed, after digitonin permeabilisation of PSI-treated cells, unglycosylated Cox-2 and pro-TNF were predominantly found in the cytosolic fraction of mycolactone treated cells (Fig. 3D, lane 3 [closed arrow]), in contrast to the LPS-induced proteins that were glycosylated and membrane associated (Figure 3D lane 6). Unglycosylated Cox-2 was found in the membrane fraction after TUN treatment showing that lack of N-glycosylation alone was insufficient to explain its localisation in mycolactone-treated cells (Fig. 3D lane 8). Inhibition of the Sec61 translocon by small molecule inhibitors, rather than causing accumulation of proteins intended for export in the cytoplasm, tends to trigger rapid degradation of the mislocalised proteins [29], [30], similarly to mycolactone. We therefore assessed whether the degradative loss we observe could be due to a blockade of translocation, using assays in which different mRNAs (transcribed and capped in vitro) undergo in vitro translation (IVT) in the absence or presence of ER containing cellular membrane preparations (Fig. 4). When such membranes are absent, no modifications of the proteins are possible. However, when membranes are present, the nascent proteins produced can undergo co-translational translocation into the ER via the Sec61 translocon and can consequently be either glycosylated or processed to remove the signal peptide sequence. In agreement with our other findings, mycolactone had a minimal direct effect on the synthesis of TNF in both the absence (Fig. 4A; TNF and luciferase mRNAs) and presence (Fig. 4B, TNF compare lanes 1 and 4) of membranes provided by semi-permeabilised RAW264.7 cell extracts.
In order to test whether TNF could co-translationally translocate into the ER in the presence of mycolactone we used Proteinase K, which can only digest proteins that it can access (i.e. those outside of the added membranes). A proportion of newly synthesised TNF could be protected from Proteinase K digestion, resulting in a resistant band of slightly lower mol wt (due to loss of pro-TNF's cytoplasmic tail, Fig. 4B, lane 2). This protected fragment was lost on inclusion of detergent (Fig. 4B, lane 3), as expected, since the Proteinase K could now access proteins in the internal membrane compartment. Mycolactone efficiently prevented TNF from translocating into the protected membrane compartment since its addition led to a complete loss of the Proteinase K resistant band (Fig. 4B, compare lanes 2 and 5). This inhibition of translocation was dose-dependent (IC50≈15 nM, Fig. 4C).
A complete blockade of TNF translocation into the ER is necessary and sufficient to explain the loss of TNF production and rapid degradation we have observed. Co-translational translocation is a mechanism utilised by many proteins, and there are some well-established model precursor proteins that can be used to investigate whether mycolactone's inhibition is more generally applicable. The N-glycosylation of yeast prepro-α Factor (PPaF) is dependent on the addition of canine pancreatic microsomal membranes (CPMM, Fig. 4D) and can be reversed by the deglycosylating enzyme Endoglycosidase H (EndoH, Fig. 4E). Mycolactone reproducibly blocked PPaF from being glycosylated (Figs 4D and 4E), and this mimicked the activity of another known translocation inhibitor, Eeyarestatin 1 (ES1, Fig. 4E). Whilst 250 µM ES1 is typically employed to inhibit ER translocation in vitro [31], much lower levels of mycolactone (0.25–0.7 µM) achieved a comparable effect (Figs. 4D and 4E; PPaF). The precise concentration of mycolactone required for a complete inhibition of in vitro translocation may be influenced by the amount of ER derived membranes present in the assay (cf. Figs 4D and 4E). Likewise, the cleavage of the β-lactamase signal peptide can also be detected after the addition of CPMM, providing an alternative measure of ER translocation. In this case we observed a substantial reduction in signal sequence cleavage although the effect was not complete (Fig. 4D, LACTB).
Mycolactone is a lipid-like molecule that is reportedly present in the cytoplasm of treated cells [8], [32], but an inhibition of translocation suggests that it may be interacting with the ER membrane in some way. Cells that are allowed to recover for 24 hr after 1 hr of mycolactone exposure are still unable to produce TNF, suggesting that this activity is irreversible (Fig. 4F). To determine whether mycolactone mediates its functions simply by disrupting organelle membrane structures we studied whether mycolactone could prevent the insertion of tail-anchored membrane proteins into the ER. Such proteins, including the β subunit of the Sec61 complex (Sec61β) and cytochrome B5 (Cyt-B5), are inserted into the ER membrane in a post-translational, Sec61-independent, manner [33], and are subsequently glycosylated on artificial C-terminal reporters (Fig. 4E, compare lanes 1 and 2). While mycolactone prevented the in vitro glycosylation of PPaF, neither mycolactone nor ES1 affected the N-glycosylation of Sec61β or Cyt-B5 (Fig. 4E, lanes 3 and 6), ruling out an effect on either membrane integrity or the N-glycosylation machinery. In addition, transmission electron microscopy showed that mycolactone did not disrupt the ultrastructure of LPS-stimulated RAW264.7 (data not shown). On this basis we conclude that the loss of co-translational translocation into the ER resulting from treatment with mycolactone reflects a specific blockade of the membrane translocation machinery.
While inhibition of translocation across the ER is sufficient to explain the loss of inflammatory mediators by mycolactone, we also investigated whether other cellular mechanisms might also contribute. First, we examined whether it might activate ER-associated degradation (ERAD) since this pathway can recognise unfolded proteins, deglycosylate and then degrade them in a ubiquitin-dependent manner. Specifically we asked whether Kifunensine (KIF; a class I α-mannosidase inhibitor and a well-established suppressor of ERAD [34], [35]) could overcome mycolactone-dependent inhibition of protein production reasoning that, if this was the case, then KIF treatment should restore protein production. The biological activity of KIF in this system was confirmed by monitoring the expression of constitutive Cox-2 expression (Fig. 5A, compare lanes 1 and 4) in the absence of LPS or mycolactone, since this is known to be turned-over by ERAD [36]. KIF significantly increased Cox-2 expression under these conditions (Fig. 5B), so we can be sure that it does inhibit ERAD at this dose in RAW264.7 cells. However, KIF was unable to restore either production of Cox-2 (Fig. 5A, compare lanes 3 and 6) or TNF secretion (Fig. 5C) in mycolactone-treated cells. This argues against mycolactone being an activator of ERAD-dependent ER export.
In addition, as shown previously in primary human monocytes [5] and T cells [17], eIF2α was not significantly phosphorylated in mycolactone-treated RAW264.7 cells (Fig. 5D). Moreover, it did not cause the phosphorylation of PERK, induce expression of BiP (Fig. 5D) or cause the IRE-dependent splicing of XBP-1 (Fig. 5E), in contrast to the known inducer of ER stress, tunicamycin (TUN). Therefore ER stress, as defined by conventional markers, cannot explain mycolactone action.
Recently, inappropriate activation of WASP family proteins by mycolactone was shown to lead to changes in cell adhesion and migration, some of which are reversed by wiskostatin [8], an inhibitor of N-WASP GTPase activity. Since actin dynamics might conceivably contribute to the inhibition of TNF and Cox-2 production by mycolactone, as well as cell adhesion, the effect of wiskostatin in our system was investigated. In contrast to previous reports in human macrophages [37], wiskostatin itself had an inhibitory effect on TNFα secretion in RAW264.7 cells (Fig. 5F). When co-incubated with mycolactone it could not restore protein production. This included the loss of LPS-induced TNF (Fig. 5F) and Cox-2 (Fig. 5G) in RAW264.7 cells and the IL1β-induced production of IL-6 (Fig. 5H) and IL-8 (data not shown) in HeLa cells at our inhibitory dose of 125 ng/ml. Since the reported restorative effect of wiskostatin on HeLa cell adhesion was determined using a lower dose of natural mycolactone A/B [8] we examined wiskostatin's effect at a range of doses of both mycolactone and wiskostatin, but could find no evidence to support an influence of WASP inhibition over cytokine production (Fig. 5H).
Translocation of proteins into the ER is a widespread cellular phenomenon but is nevertheless restricted to a defined subset of proteins, most of which carry a canonical signal peptide. Our findings suggest that in mycolactone-exposed cells, the production of most proteins within this subset may be blocked. We tested this by examining protein synthesis in different cellular compartments (Fig. 6A and B). The translation elongation inhibitor cycloheximide (CHX) led to an almost complete block in protein synthesis in both cytosolic and membrane fractions of RAW264.7 cells, as expected. Yet, while mycolactone caused little change in cytosolic protein synthesis, it did cause a selective ∼30% decrease in membrane-associated proteins (Fig. 6A and B, P<0.01). Since the membrane fraction would include proteins from various intracellular organelles (nucleus, mitochondria etc) as well as exported proteins, Concanavalin A (ConA) agarose was used to isolate glycosylated proteins in order to better represent the subset of proteins that must translocate into the ER. As predicted, a large decrease in the recovery of such constitutive and induced proteins from mycolactone-treated cells was found (Fig. 6C). Notably, the degree of loss varied and a minority of proteins showed little or no change. When supernatants were examined, the abundance of almost all constitutive and LPS-induced proteins was reduced, with only a single ∼32 kDa protein unaffected (Fig. 6C). This was dose-dependent (Fig. 6D) and production of most mycolactone-sensitive proteins showed similar IC50 values (57.8±7.9 nM, mean±SEM, n = 3). Indeed, when we investigated the production of 18 cytokines and chemokines produced by RAW264.7 cells (Fig. S3A), we found that 17 were almost completely blocked by mycolactone after LPS stimulation (Fig. S3B and C). Those not reported as targets of mycolactone previously include TIMP-I, soluble ICAM-1 and IL-RA. A single chemokine, MIP-1α seemed to be completely insensitive to mycolactone inhibition in this antibody array, however validation by ELISA showed that these antibodies may be saturated by the high levels of constitutive protein, since this more quantitative analysis showed a profound inhibition by mycolactone (Fig. S6D). A similar phenomenon is most likely also responsible for the apparent reduced sensitivity of constitutively expressed JE (CCL2, commonly known as MCP-1 in humans) and MIP-1β in the absence of LPS (Fig. S3B and C), especially since others have previously shown that both are mycolactone targets in other systems [16], [18].
Since many different types of cells are exposed to mycolactone in BU lesions and translocation by the Sec61 complex is highly conserved [38], we examined whether the production of secreted and glycosylated proteins would be similarly prevented in non-immune cells. Interestingly this was found to be the case in human dermal microvascular endothelial cells (HDMVEC), murine L929 fibroblasts and HeLa cells (Fig. 6E). In all cases, there was a profound inhibition of protein production affecting the majority of proteins in these compartments. Therefore, a block in translocation of proteins into the ER represents a widely applicable mechanism underlying mycolactone's pathogenic effects.
In this manuscript we have identified an important new activity for the M. ulcerans virulence factor, mycolactone. By investigating the inhibition of cytokine production as a model system to examine its basic cell biology, we have shown that mycolactone effectively blockades the translocation of nascent proteins across the ER membrane in a mechanism that seems to involve the Sec61 translocon. This finding is remarkable because previous data suggested that mycolactone was probably inhibiting the translation of inflammatory mediators such as TNF, IL-6 and Cox-2 [2], [5]. However, our detailed investigation refutes this. We showed conclusively that the transcripts were actively translating in vitro and in a RAW264.7 cell model in both the absence and presence of mycolactone. The polysomal location of each of these transcripts in treated cells is remarkably similar to that seen in untreated cells and their relocation in response to inhibitors of translation is also unaffected. Likewise the demonstration of an association between these mRNAs and the membrane fraction in mycolactone-treated cells provides independent evidence of sufficient translation having occurred to allow signal peptide-directed targeting to the ER.
We provide multiple lines of evidence in support of mycolactone's ability to prevent translocation across the ER and induce subsequent degradation in the cytosol. This is supported first by the finding that translocation of TNF into a protease resistant membrane compartment could no longer occur in the presence of mycolactone in vitro. Furthermore, production of both Cox-2 and cellular pro-TNF was restored when the protease activity of the proteasome was blocked by PSI after transcriptional activation and translational derepression had occurred. The restored Cox-2 was unglycosylated; however this is most likely a consequence of never having entered the ER rather than a cause of degradation. The effects of TUN and mycolactone on TNF secretion and Cox-2 production are quite distinct and the restored Cox-2 and TNF were both present in the cytosol rather than the membrane. Therefore, while blocking the proteasome prevented the degradation of the proteins, it could not overcome mycolactone's translocation blockade and the restored proteins would not be able to carry out their normal cellular function. Proteasome inhibitors such as Bortezomib have attracted some attention as cancer therapeutics, but for the reasons laid out above such drugs should not be considered for BU treatment.
There are a number of mechanisms by which proteins can enter the ER and then go on to be secreted, retained in the ER or inserted into the membrane. These include co- and post-translational Sec61-dependent translocation as well as Get/TRC40-dependent insertion of tail anchored proteins, reviewed in [39]. Among cytokines, TNF is slightly unusual in that it is made as pro-TNF, co-translationally inserted into the ER membrane as a Type II membrane protein then cleaved at the cell surface (by TNF-converting enzyme; TACE). Most others, including IL-6, undergo conventional trafficking via post-translational translocation into the ER, followed by signal sequence cleavage. IL-1β and IL-18 provide notable exceptions to this rule, and are secreted by an unconventional mechanism following cleavage of the cytosolic pro-isoforms in a caspase- and inflammasome-dependent manner [40]. The precise cellular mechanism that results in consequent export of mature cytokines is still somewhat controversial. Neither IL-1β nor IL-18 can be made by RAW264.7 cells (see Fig. S3C) as they lack the required ASC (Apoptotic speck protein containing a caspase recruitment domain [41]), and so they were not studied further here. However, it is interesting to note that in our previous work with primary human monocytes, IL-1β production was only partially blocked by mycolactone (between 30–70% depending on the TLR ligand used to activate the cells [5]). The ability of mycolactone to inhibit unconventional secretion is now the subject of further investigation.
To our knowledge, mycolactone is the first and only virulence factor shown to inhibit translocation into the ER. A very few, non-pathogenic, compounds have been described that block Sec61-mediated transport: the substrate selective inhibitors cotransin (derived from the natural product HUN-7293) and its close relative CAM741, as well as ES1 and Apratoxin A. In each case a block in the cotranslational translocation of exported peptides leads to their rapid degradation in the cytosol [29]–[31], [42]. ES1 is arguably the best characterised of the translocation inhibitors but it must be used at a much higher dose (8 µM for treatment of cells, 250 µM in IVT) than is required for mycolactone-dependent inhibition. In a direct comparison, both were found to be similarly inactive against the insertion of tail-anchored proteins (Fig. 4E). In addition to its ability to inhibit entry into the ER, ES1 also inhibits ERAD driven exit and triggers an unfolded protein response ([43], [44] and data not shown). We investigated both these other pathways here. The inability of KIF to enhance expression of TNF or Cox-2 in the presence of mycolactone makes increased export of misfolded proteins via the ERAD pathway an unlikely mechanism contributing to the loss of inflammatory mediators (Fig. 5A and B). Moreover, we found no evidence of induction of the unfolded protein response in mycolactone-treated cells. Therefore mycolactone's cellular effects are similar to, but discrete from, ES1.
In contrast to mycolactone and ES1's ability to prevent the production of nearly all glycosylated and secreted proteins (Fig. 6C–E and [31]), cotransin is both non-toxic and highly selective in its action, affecting only a small subset of substrates [30]. Apratoxin A causes a wide inhibition of protein secretion but, also differs from mycolactone because its effects are completely reversible [42]. To date in the literature, there have been contradictory reports on the reversibility of mycolactone action. While L929 fibroblasts are reported to regrow after the removal of mycolactone [1], dendritic cells cannot regain the ability to respond to maturation stimuli after a prolonged exposure (24 hrs) [16]. Here we show that, for cytokine production, the effects remain irreversible after 24 hours, even after a brief exposure (Fig. 4F).
The recently reported ability of mycolactone to enhance actin polymerisation and inappropriately activate WASP [8] does not seem to be a major contributor to the inhibition of production of inflammatory mediators. Mutations in human WASP are associated with a range of immune dysfunction disorders and WASP has also been implicated in Golgi to ER transport [45], [46]. However, doses of the WASP inhibitor wiskostatin (equivalent or higher than those shown to partially restore adhesion in HeLa cells [8]) could not restore secreted protein production at a range of doses of mycolactone (Fig. 5F–H). Therefore, the decisive step in mycolactone-dependent inhibition of protein production seems to depend on the translocation blockade rather than WASP activation. Determining the precise mechanism by which mycolactone blocks translocation into the ER, and the molecular consequences of this action are the subject of ongoing investigation. It will be interesting to determine the molecular explanation of how some signal-peptide containing proteins including β-lactamase (Fig. 4D) and CCR7 [17] escape this translocation block. Indeed selectivity in mycolactone action has been fairly widely reported and the variable sensitivity to mycolactone inhibition observed in our in vitro translocation assays is consistent with this [5], [6], [16]–[18].
The broad spectrum inhibition of protein translocation induced by mycolactone in various cell types has important implications across diverse pathologies of Buruli ulcer. The suppression of innate and adaptive immune responses mediated by secreted cytokines and chemokines is one important aspect of this. In this work, we have identified nine novel immune proteins that are sensitive to mycolactone (BCA, MIP-2, G-CSF, GM-CSF, IL-27, C5/C5a, sICAM-1, IL-1RA and TIMP-1). In addition, the general block in production of glycosylated proteins would also affect many other proteins destined for the cell surface that mediate essential cellular functions. This would include those involved in aspects of adaptive immunity driven by maturation markers and costimulatory molecules on dendritic cells [16] where an inability to produce these proteins would lead to a failure to interact with lymphocytes, causing the T cell anergy seen in BU [2], [47], [48]. Likewise, the loss in T cell homing has been attributed to loss of receptor expression and this powerful mechanism of suppression would substantially add to the effect of let-7b [6]. These authors did investigate the role of proteasomal inhibition with MG132, but the experimental technique used (flow cytometry of fixed, permeabilised cells) would probably be insensitive to the restoration of protein that would now be located in the cytosol (see Fig. 3D) and therefore unlikely to be in its native conformation. In addition, the translocation block may contribute, along with WASP activation [8], to the rounding up and loss of adhesion seen in fibroblasts due to gradual loss of cell adhesion molecules as they turn over. Since anoikis is described as driving cell death, this could be an important contributory factor in the necrosis seen in ulcers. Moreover, a steady depletion of endogenous proteins in this way would lead to a gradual winding down of cellular function, rather than rapid cell death, which ties in well with the slow progression of the disease in humans, as well as in vitro and in vivo data [4]. For instance, the delay between injection of even 100 µg of mycolactone into guinea pig and ulcer formation is 5 days [1], although some early signs of apoptosis are evident from 2 days [7].
BU is the third most common human mycobacterial infection in the world after tuberculosis and leprosy. Although not fatal, patients can suffer lifelong disfigurement and disability unless the infection is recognised and treated at an early stage. Treatment with antibiotics is often accompanied by increased swelling and ulceration, or even the appearance of new lesions [49]. Such paradoxical responses are thought to arise from a gradual increase in proinflammatory activity as mycolactone levels decline. Several studies in both human patients and animal models have shown that a progressive and structured immune response follows initiation of therapy with the standard rifampicin/streptomycin combination [20], [50], [51]. As bactericidal activity does not always correlate with healing rate, it has been suggested that the drugs, rifampicin in particular, may inhibit the ability of M. ulcerans to synthesise mycolactone and that this is the more important factor in recovery [51], [52]. It seems unlikely that this phenomenon is due to a direct effect of the antibiotics on the immune system. Although rifampicin is reported to be immunomodulatory, responses vary and while some proinflammatory responses are enhanced by exposure to the drug, others are blocked [53], [54]. The molecular mechanisms of these effects seem independent from Sec61-dependent translocation and are mostly upstream of this process. In Jurkat T cells for example, rifampicin inhibits TNF production by blocking NFκB activation [54]. However, the possibility that antibiotics may promote restoration of immune function by altering the interaction between mycolactone and its cellular target cannot yet be ruled out and is worth investigating. A greater understanding of these interactions would give insights valuable to improving diagnosis and treatment of this debilitating disease. Furthermore, unravelling the mechanisms of mycolactone activity should also yield novel insights into the vital, basic cellular process of ER translocation.
A full list of reagents, primers and antibodies can be found in Information S1. We used synthetic mycolactone A/B (kind gift of Prof. Yoshito Kishi, Harvard University) throughout these investigations [55]. All reagents used in tissue culture were routinely tested for endotoxin contamination by LAL assay (Lonza) and were <0.1 U/ml LPS.
The RAW264.7 murine macrophage cell line (ATCC) was routinely cultured at 37°C and 5% CO2 in high glucose DMEM medium (PAA) supplemented with 10% FBS (Life Technologies). Culture conditions for additional cell types can be found in Information S1. Cells were pre-incubated with mycolactone or other inhibitors for 1 hr before stimulation with 100 ng/ml TLR-grade LPS (Enzo Life Sciences) then incubated for 1–4 hr before harvesting of cells and culture supernatants. Mycolactone was routinely used at a final concentration of 125 ng/ml; DMSO diluted to the same extent (0.0125%) was the control. Other inhibitors used were PSI (5 µM, Calbiochem), CHX (10 µg/ml), ActD (2 µg/ml), KIF (50 µM), TUN (5 µg/ml), and wiskostatin (1 µM). For proteasome dependent degradation experiments, cells were pre-incubated with mycolactone or TUN and stimulated with LPS for 2 hr, prior to the addition of PSI for a further 2 hr. To investigate the reversibility of mycolactone action, cells were exposed to mycolactone for 1 hr, then mycolactone was removed, the cells were washed with PBS then incubated in complete DMEM for the indicated time, after which cells were stimulated with LPS for 4 hrs before harvesting the supernatants. HeLa cells were stimulated with 10 ng/ml recombinant IL-1β (Peprotech).
Polysome profiling by sucrose density gradient ultracentrifugation was carried out according to a previously published method [56], [57]. A full protocol can be found in Information S1. Briefly, treated RAW264.7 cells were harvested 10 min after ribosomes were stalled with CHX (10 µg/ml). In some cases, 100 µg/ml PURO or 5 µM HH were added 3 min prior to addition of CHX. Cell lysates were separated over a 10–50% sucrose gradient. RNA was extracted from 1 ml fractions and analysed by Northern blotting using 32P-labelled (full coding region) cDNA probes.
Membrane bound and cytosolic cellular fractions were separated by digitonin permeabilisation using a previously published method [26] with minor alterations (full protocol in Information S1). Briefly, proteins from treated RAW264.7 cells were extracted by sequential use of permeabilisation (0.03% digitonin) and solubilisation (1% NP-40, 0.5% sodium deoxycholate) buffers.
TNFα cDNA was prepared from LPS-stimulated primary human monocytes by RT-PCR and capped mRNA synthesised using Message Machine (Applied Biosystems). Control mRNAs (luciferase, α Factor and β lactamase) were all from Promega. In vitro translation (IVT) reactions were carried out using nuclease-free rabbit reticulocyte lysates (Promega) with 0.5–1.0 µg mRNA. Mycolactone was diluted in 5% (w/v) BSA in nuclease-free water before addition, controls contained BSA alone. Where used, canine microsomal membrane (CPMM) preparations (Promega, or prepared in house [58]; a kind gift of Prof Barnhard Dobberstein [University of Heidelberg]) or semi-permeabilised RAW264.7 cell extracts freshly prepared as described in [59] were added to a final concentration of 10%. Samples were incubated for 30 min at 30°C. For protease protection assays, samples were diluted in 1∶5 in 20 mM TrisCl pH 8.0, 10 mM CaCl2 and split into 3 aliquots: to the control sample buffer alone was added to a final volume of 50 µl, the second aliquot contained 20 µg/ml Proteinase K and the third 20 µg/ml Proteinase K and 0.1%Triton-x-100. Samples were incubated for 1 hr at 4°C, then reactions were stopped by addition of 5 mM PMSF and boiling sample load buffer. For glycosylation assays of tail-anchored proteins and PPaF, ES1 was pre-incubated with in-house CPMM for one hour before the addition to other components. Mycolactone-containing CPMM were used immediately. Membranes were recovered as described [31] by centrifugation through 750 mM sucrose, 500 mM KOAc, 5 mM Mg(OAc)2, 50 mM HEPES-KOH (pH 7.9) at 100,000×g for 10 mins. The membrane pellet was resuspended in 100 mM sucrose, 100 mM KOAc, 5 mM Mg(OAc)2, 50 mM HEPES-KOH (pH 7.9), 1 mM DTT and treated with 250 µg/ml RNaseA at 37°C for 10 mins to remove any residual peptidyl-tRNA species. Samples were separated by SDS-PAGE. Luciferase, α Factor, β lactamase, Sec61β and Cyt-B5 were detected by labelling with 35S methionine; TNFα was detected by SDS-PAGE (15% acrylamide) and Western blotting.
Total RNA was extracted from cell lysates or digitonin permeabilised fractions using the RNeasy kit (Qiagen) and quantified by Nanodrop. One-step qRT-PCR gene expression assays (Life Technologies) were carried out on either an Applied Biosystems 7900 (Life Technologies) or on a Stratagene Mx3005P (Stratagene). For relative gene expression the ΔΔCt method was used. For absolute quantitation, full-length murine cDNAs were prepared and used to form standard curves. XBP-1 splicing was investigated by RT-PCR of total RNA followed by Pst1 digestion as described [60].
Secreted TNF was detected in culture supernatants by ELISA. For Western blots, cells were lysed directly in gel sample buffer then sonicated or permeabilised with digitonin. Proteins were separated by SDS-PAGE (12.5% acrylamide) followed by conventional blotting. Where quantitation was performed, pixel density was assessed using ImageJ analysis of non-saturated images and data were normalized to an appropriate loading control (GAPDH)
Metabolic labelling was performed as previously described [5]. Met/Cys-starved cells were stimulated as described above for 2 hrs before the addition of 0.37MBq Tran35S-Label for a further 2 hr at 37°C. They were then either lysed in 20 mM TrisCl, pH 7.4, 0.5M NaCl, 1% Triton-X-100. 1× protease inhibitor cocktail) or separated into cytosolic and membrane fractions as described above. Samples were separated by SDS-PAGE, exposed to a phosphorimager screen and analysed using a Personal FX imager (BioRad).
α factor (PPaF), β-actin: P60710, BiP: P20029, Cox-2: Q05769, Cytochrome b5: P00167, EIF2α: Q6ZWX6, GAPDH: P16858, Glucosidase 1: Q80UM7, IL6: P08505, β lactamase: Q9L5C7, Luciferase: P08659,: P01149, PABP1: P29341, PERK: Q9Z2B5, TNF (murine): P06804, TNF (human): P01375, Sec61β: P60468, XBP1: Q6ZWX6.
|
10.1371/journal.pcbi.1000327 | T-Cell Epitope Prediction: Rescaling Can Mask Biological Variation between MHC Molecules | Theoretical methods for predicting CD8+ T-cell epitopes are an important tool in vaccine design and for enhancing our understanding of the cellular immune system. The most popular methods currently available produce binding affinity predictions across a range of MHC molecules. In comparing results between these MHC molecules, it is common practice to apply a normalization procedure known as rescaling, to correct for possible discrepancies between the allelic predictors. Using two of the most popular prediction software packages, NetCTL and NetMHC, we tested the hypothesis that rescaling removes genuine biological variation from the predicted affinities when comparing predictions across a number of MHC molecules. We found that removing the condition of rescaling improved the prediction software's performance both qualitatively, in terms of ranking epitopes, and quantitatively, in the accuracy of their binding affinity predictions. We suggest that there is biologically significant variation among class 1 MHC molecules and find that retention of this variation leads to significantly more accurate epitope prediction.
| The use of prediction software has become an important tool in increasing our knowledge of infectious disease. It allows us to predict the interaction of molecules involved in an immune response, thereby significantly shortening the lengthy process of experimental elucidation. A high proportion of this software has focused on the response of the immune system against pathogenic viruses. This approach has produced positive results towards vaccine design, results that would be delayed or unobtainable using a traditional experimental approach. The current challenge in immunological prediction software is to predict interacting molecules to a high degree of accuracy. To this end, we have analysed the best software currently available at predicting the interaction between a viral peptide and the MHC class I molecule, an interaction that is vital in the body's defence against viral infection. We have improved the accuracy of this software by challenging the assumption that different MHC class I molecules will bind to the same number of viral peptides. Our method shows a significant improvement in correctly predicting which viral peptides bind to MHC class I molecules.
| Cytotoxic T lymphocytes (CTLs) discriminate between healthy and pathogen-infected cells by recognizing and responding to a molecular complex on the surface of the infected cell. This complex consists of a specific major histocompatibility complex (MHC) molecule and a peptide derived from the proteins contained in the cell. If the cell contains a pathogen, peptides from the pathogen proteome will be presented and, with the right MHC – peptide complex, a CTL response will be elicited.
Of the large number of peptides that can be derived from a pathogen only a small minority elicits a CTL response. This number has been estimated to be between 1 in 2,000 and 1 in 5,600 [1],[2]. This limitation in the number of peptides that are immunogenic is conferred by three main constraints: the requirement for peptide cleavage and transport, the requirement for MHC-peptide binding and the requirement for CTL recognition. By far the most stringent of these is the requirement for MHC-peptide binding, because only 1 in 40–200 peptides binds a specific MHC molecule with sufficient affinity to elicit an immune response [1],[2]. Further selection is largely due to the limitations of peptide processing and transport. In these processes, individual peptides are produced from the precursor polypeptides by proteasomal cleavage of the polypeptide, which can be followed by N-terminal trimming by other peptidases. This is followed by the transport of the peptides from the cytosol to the endoplasmic reticulum, mediated by the TAP complex. Further N-terminal trimming may occur before the peptide binds to the MHC molecule. The requirements of processing and transport eliminate approximately 80% of potential epitopes [1]. Finally, T cell specificity, i.e. the requirement for T cell receptor binding of the MHC-peptide complex, further halves the number of presented peptides that elicit a response. The probability of each of these steps is determined by the polypeptide sequence, amongst other factors [3].
Once CTLs recognize the MHC-peptide complex, they are capable of destroying the infected cell by the release of lytic granules containing cytotoxic effector proteins. This results in the destruction of the target cell by apoptosis. An effective CTL response has been shown to confer protection against viral infection, such as HIV [4] and HTLV-I [5]. Hence, the identification of T cell epitopes is of vital importance in the design of vaccines and understanding of the immune system [6],[7],[8]. However, given the scarcity of epitopes, experimentally screening all possible peptides for each MHC allele (e.g. by IFNγ ELISpot) is time consuming, expensive and inefficient. One way to improve the efficiency of the identification process is to first use theoretical algorithms to predict which peptides are more likely to be epitopes and then experimentally screen this much smaller, selected list of peptides. This method is widely used [9]–[12] and has been applied in a number of studies to identify potential vaccine targets [13],[14]. The use of theoretical methods to “pre-screen” peptides is of particular importance in the case of emerging infections such as avian influenza [15] where rapid vaccine development would be vital. This approach underpins a large bio-preparedness initiative coordinated by the Large-Scale Antibody and T Cell Epitope Discovery Program [7], which intends to foster development of immune-based therapeutics for emerging and reemerging pathogens including potential bioterrorism agents. The accuracy of these methods has also been demonstrated by the prediction of the vast majority of CTL epitopes from the vaccinia virus [16].
More generally, epitope prediction algorithms are being increasingly used to understand the CTL response. For example, in the case of HIV-1 infection, algorithms have been used to confirm which MHC-associated epitope mutations are likely to confer escape from a CTL response [17] and to understand why some MHC class I alleles are associated with slow rates of disease progression [18],[19].
A range of computational algorithms have been developed to predict CTL epitopes in pathogen protein sequences. Since the most selective requirement for a peptide to be immunogenic is the ability of the peptide to bind to the MHC molecule, most prediction methods focus on this stage of the pathway. As a general rule, information gained from experimental binding assays is used to train the algorithm until it is efficient at predicting novel MHC–peptide complexes. The algorithms that are used vary in complexity and accuracy. Some can be trained to recognize peptide motifs that are required for binding to a particular MHC molecule [20], others use a weight-matrix method to identify amino acids that occur at a higher-than-expected frequency at specific epitope positions [21],[22],[23]. However, the most accurate methods available use logistic regression [24] and, more generally, artificial neural networks [3].
Artificial neural networks (ANNs) take into account, in addition to the identity of each amino acid residue, the interactions between adjacent amino acids in a potential epitope. In summary, an ANN for a particular MHC molecule is trained to recognize associated inputs (a peptide sequence) and outputs (the binding affinity for that sequence with the MHC molecule) [25]. Once an ANN is trained for a particular molecule, it can predict the binding affinity of novel peptide sequences.
NetCTL [3] and NetMHC [25],[26],[21] are two of the most accurate prediction methods currently available [27]. NetMHC uses ANNs for a number of alleles to predict MHC molecule-peptide binding affinities. NetCTL, as well as using ANNs to predict MHC – peptide binding, also utilizes information about the proteasomal cleavage of the input peptide sequence, and its ability to bind to TAP. NetCTL or NetMHC will predict a score (either integrated or simply a binding affinity, respectively) for every overlapping nonamer peptide sequence in an input sequence to each MHC molecule for which the method has an ANN. Henceforth, we refer to the trained prediction algorithm for each MHC class I allele as an “allelic predictor”.
In order to make the prediction values comparable between each MHC molecule, it is recommended that the MHC-peptide binding affinity scores are rescaled [28]; this is explicitly implemented in NetCTL. The method of rescaling involves obtaining the predicted binding affinities of 500,000 random natural peptides for each MHC allelic predictor. From these affinities, a rescale value is calculated, defined as the binding affinity that is the threshold for the top 1% of total binding affinities. The rescaled affinity is then defined as the predicted affinity score divided by this rescale value [3]. Hence, from this calculation, all alleles are predicted to bind the same number of high-affinity peptides. One pragmatic reason for rescaling is to correct for any discrepancies between the allelic predictors that resulted from inconsistent training data (e.g. data that came from different sources), by assuming that all alleles should bind the same number of epitopes (C. Keşmir, pers. comm.). Additionally, there are biological arguments for believing that different alleles should bind similar numbers of epitopes. It has been postulated that the opposing constraints of effective pathogen recognition but tolerance of self would result in a very narrow range of optimal promiscuity for viable MHC class I molecules. A narrow range of promiscuity would also be predicted as a direct outcome of effective tapasin-dependent peptide optimization in the endoplasmic reticulum [29],[30],[31].
However, we will present evidence in this paper that in correcting for differences between the allelic predictors, information is being lost that reflects true biological variation between MHC molecules and, by extension, differences in their ability to bind to peptide sequences. We show that, for both qualitative and quantitative measures of binding, rescaling impairs rather than improves allelic predictor performance. This is of importance for vaccine design and to understand the nature of the CTL response. In particular, crucial between-allele variations in binding affinity and preference which may contribute to differences in the outcome of infection are likely to be obscured by rescaling.
In order to test the effect of rescaling on epitope prediction accuracy, we used two web-based prediction methods, NetCTL v1.2 [3] and NetMHC v3.0 [25],[26],[21]. NetCTL is an integrated method that uses information pertaining to TAP and protein cleavage in its predictions, together with MHC binding. The output is combined by rescaling the MHC binding result and adding this to the weighted scores for TAP and protein cleavage. NetCTL has allelic predictors for 12 different class I alleles that are chosen to be representative of each of 12 supertypes; hence it has 12 different rescaling factors.
NetMHC v3.0 simply predicts MHC-peptide binding, using ANNs to predict binding affinities for 43 MHC molecules. In order to test the effect of rescaling, it was necessary to produce rescale values for each of the 43 allelic predictors. This was performed as in NetCTL; 500,000 unique random nonamers were obtained from the proteome of Mycobacterium tuberculosis, their binding affinity was predicted and the rescale value (top percentile) was found for each allelic predictor. We also performed this calculation with 500,000 random natural peptides to test for the possibility of error from bias in amino acid usage in Mycobacterium tuberculosis. There was no significant difference in the rescale values obtained using these two different sources (supplementary material, figure S4).
In summary, we tested two sets of rescaling values: those obtained from NetCTL v1.2 and those that we calculated using NetMHC v3.0.
Epitope datasets were constructed from sources detailed below. In each case, the prediction methods were tested by their ability to detect these epitopes amongst the full set of overlapping nonamers derived from the proteins that contained the epitopes. The full set of nonamers will contain a small number of known epitopes and the remainder will be ‘non-epitopes’. Of course, this set of non-epitopes could include epitopes that have not been experimentally verified. However, the majority (see Introduction) would be non-binders with the corresponding MHC molecule. Added to this, the labelling of epitopes as ‘non-epitopes’ impact on both rescaled and non-rescaled calculations equally. Previous research has also shown that this property of the ‘non-epitope’ set did not produce significantly different results [24]. Each respective set of experimentally defined epitopes was denoted the positive dataset and the set of non-binding (or unknown) peptides was denoted the negative dataset.
The SYF1 dataset is a supertype dataset derived from SYFPEITHI [20] and is identical to that used in the original paper for NetCTL [3]. Each epitope in SYF1 was experimentally verified to bind to one of 10 MHC class I supertypes [32]. The resulting dataset consisted of 148 epitope-supertype pairs. The corresponding negative dataset was obtained by concatenating the SwissProt entry proteins from which each of the epitopes was derived. The length of the concatenated protein sequence was 78,259 amino acids. The ROC curve (see below for explanation) was generated using a negative set of ((78,259*10)−148) = 782,442 nonamers and a positive set of 148 nonamers. The positive set of SYF1 is available in the supplementary material (dataset S2).
Experimentally defined epitopes in HIV-1 were extracted from the HIV Molecular Immunology Database [33]. In total, 1,618 CTL epitopes were found that were bound by human MHC molecules. However, this set was highly redundant; the epitope lengths were variable and a large number of epitopes differed only by mutations within the sequence. Also, resolution of their MHC typing varied from 2 to 4 digits. To correct for this variability, a number of changes were made to the MHC allele-epitope list. Firstly, all MHC alleles were defined to two digits. Secondly, variant epitopes binding the same allele were discarded. Finally, as the prediction software only produced binding predictions for nonamer epitopes, all epitopes that were not 9 amino acids long were removed from the list.
In summary, it was possible to test 41 of the 43 allelic predictors for MHC molecules in NetMHC v3.0. The positive set consisted of 661 epitopes, defined in terms of start and end positions relative to the HIV reference strain HXB2 (supplementary dataset S1) and a matching MHC type to 2 digits. The input protein sequence to NetMHC contained 3,000 overlapping nonamers that covered the proteome from which the whole positive set of epitopes was derived. The total ‘negative set’ for the ROC analysis was (3,000 * 41)−661) = 122,339 nonamers, and a positive set of 661 nonamers. The positive set of Lanl661 is available in the supplementary material (dataset S3).
The Lanl661 dataset was modified for testing with NetCTL. From these 661 epitopes, a total of 179 bound to the 12 alleles for which NetCTL has allelic predictors. The input sequence to NetCTL contained 3,000 overlapping nonamers. For this experiment, the negative set consisted of ((3,000 * 12)−179) 35,821 nonamers, and a positive set of 179 nonamers. The positive set of Lanl179 is available in the supplementary material (dataset S4).
ROC curves give a visual measure of the accuracy of a prediction method. The threshold at which the prediction method identifies a peptide as being an epitope varies along the length of the curve. Each point on the curve gives the fraction of true positive epitopes found as a function of the number of false positive ‘epitopes’ at that threshold. Hence, setting a strict threshold for epitope detection will result in high specificity (correct predictions) but low sensitivity (missing a high proportion of true binders). The area under the ROC curve gives the AUC (Area under Curve) measurement. In order to test for significant difference between ROC curves, we conducted the bootstrapping analysis detailed in [34]. Briefly, using bootstrapping with replacement, 100 replicates were formed from each dataset and the resulting non-rescaled and rescaled whole AUC values were compared using a paired t test.
Using the 2 epitope datasets, HIV216 and SYFPEITHI863, and the same methods from [35], we repeated 3 of the measurements described in that paper for the rescaled and non-rescaled results of NetCTL v1.2. For the Rank measure, we analysed the proteins from which each epitope was derived. For each protein, we calculated the rank of the epitope amongst all overlapping 9-mers using rescaling and non-rescaling scoring methods for all alleles. We then analysed these ranks to see which method ranked the epitopes higher. For the second method, we measured the specificity of both rescaling and non-rescaling at predefined sensitivities. Finally, we measured the sensitivity among the top 5% top-scoring peptides, again for the rescaled and non-rescaled binding affinities.
The training data for NetMHC v3.0 is available at http://mhcbindingpredictions.immuneepitope.org/. An independent set of experimental epitope-allele binding affinities was obtained from the Immune Epitope Database and Analysis Resource (IEDB) by selecting all experimental data that did not originate from the laboratories of Sette et al. or Buus et al. (the training data originated from these two sources).
ROC curves were used to analyse the effects of rescaling on epitope prediction. Both NetCTL v1.2 and NetMHC v3.0 were tested and 3 datasets were used (figure 1 and table 1). In each case, rescaling resulted in a significant loss of performance (bootstrap test: p<0.001).
One possible explanation for why rescaling has a detrimental impact on prediction is that there may be a positive correlation between rescale factor and allelic predictor accuracy. To check this hypothesis we calculated the AUCs for each NetMHC v3.0 predictor using the Lanl661 dataset and plotted this against the corresponding rescale factor, the results of which are shown in figure 2. This shows no evidence of a correlation between rescaling values and the AUC values (R2 = 0.0068, p = 0.606).
Consequently, it is unlikely that a correlation between rescale values and AUC values explains our findings. However, certain alleles like B0801 do have both a low rescale value and a low AUC. To double check that these poor accuracy predictors were not causing the inaccuracies in rescaled predictions we repeated our ROC curve analysis for Lanl661 without the low accuracy predictors (those with an AUC value below 0.9; namely A6801, A6802, B3501, B0702, B0801, B0802 and B4501). In the remaining, reduced subset of predictors there was even less evidence for a correlation between AUC and rescale factor (R2 = 0.0007, p = 0.887). For this subset of predictors the accuracy was still significantly better if rescaling was not applied (figure S1; bootstrap test: p<0.001) and comparable to the ROC curve analysis using the full set of alleles (figure 1C).
Therefore, we believe there is no evidence to support the hypothesis that the reason rescaling is detrimental is because there is a correlation between rescale factors and AUC.
We used 3 other metrics [35] to compare predictive performance with and without rescaling.
Using 2 sets of experimentally-derived epitope-allele binding affinities, we also showed that the correlation between predicted and experimental affinities was weaker with rescaling than without (supplementary figure S3).
Rescaling is, in theory, a sound approach to improving epitope prediction and in particular comparability of predictions obtained using different allelic predictors. However, using a number of different measures of accuracy, in the context of two commonly used prediction methods, we have demonstrated that rescaling actually impairs rather than improves predictive performance and comparability. We suggest that rescaling predicted affinities results in a loss of information that outweighs any advantage gained in correcting for differences in training data.
The first approach used ROC curve analysis and showed clear differences between rescaling and non-rescaling. The ROC curve gives a graphical representation of how well the prediction method ranks true epitopes among a set of non-binding peptides. Or to use an analogy, how efficient it is at finding the epitopic needle in a haystack of random peptides. From figure 1, it is clear that rescaling across all allelic predictors results in a performance loss in terms of how well the method ranks its peptides by binding affinity; that is, rescaling impairs intra-allelic comparisons. This loss could be demonstrated using epitope data from a number of sources (SYFPEITHI, the HIV Molecular Immunology Database) and with two different methods of prediction (the combined approach of NetCTL v1.2 and NetMHC v3.0). This effect of rescaling would be detrimental to any studies screening across a number of alleles for possible epitopes (such as [15]). The effect of this performance difference can be gauged from figure 1 (A). In order to identify correctly 85% of the epitopes the percentage of false positives detected was 9% and 15%, for non-rescaled and rescaled methods respectively. To put this result into context, the viral protein NS1 from the H5N1 strain of Avian Influenza A consists of 221 overlapping nonamers. To screen this protein for potential epitopes, 33 epitopes would need to be experimentally checked for each MHC molecule of interest if rescaled predictions were used, as opposed to 20 for the non-rescaled predictions (providing 85% epitope coverage was sufficient).
Added to the significant results from the ROC curve analysis, the supplementary analysis demonstrated the positive effect of removing rescaling in terms of the correlation with experimental data (supplementary figure S3) and also in terms of per-protein and sensitivity analysis (supplementary figure S2 and tables S1 and S2). Taken together, these results strongly demonstrate the improvement in accuracy of removing the condition of rescaling when comparing predictions between alleles.
There has been little research on the variation in ‘stickiness’ among MHC molecules, i.e. whether some MHC class I molecules are capable of binding to a greater number of epitopes than others. The binding motifs for MHC-peptide binding vary across the range of alleles, but the assumption made for rescaling is that each molecule would bind to the same number of peptides out of a large random selection. Estimates based upon mass spectrometry suggest that over 2,000 peptides are associated with HLA-A2.1 and −B7 and it is speculated that the actual total could be over 10,000 per MHC molecule [36]. However, it is not known how this number varies between molecules. It has been postulated that the twin constraints of effective pathogen recognition but tolerance of self would result in a very narrow range of promiscuity for viable MHC class I molecules [29]. Contrary to this, recent research has shown that this range may be wider than initially envisaged [37] and our results suggest that there is considerable inter-allelic variation in promiscuity.
This data may also be informative regarding optimization of peptide cargo in the endoplasmic reticulum (ER). We would argue that peptide optimization is the biological interpretation of rescaling: alleles have similar numbers of epitopes because peptides with a lower binding affinity are replaced in the ER. We know that optimisation cannot be complete because otherwise every allele would just present one epitope: the one with highest affinity. However, it seems likely that there is a degree of optimization [30],[31]. The observation that rescaling gives worse predictions may put a bound on how much optimisation is occurring. Allied to this, it has been observed that the release of an MHC class I molecule from the peptide-loading complex with a suboptimal peptide takes precedence over the prolonged detention of the MHC class I molecule in the complex until an optimal peptide comes along [30]. Hence, peptide optimization acts to reduce inter-allelic variation and promiscuity results from inter-allelic variation in allele-peptide affinity. However, this peptide optimization is limited by time and is not complete and hence, we note this variation in promiscuity across different alleles.
In summary, we suggest that much of the observed variation between allelic predictors reflects genuine biological information which should not be discarded as experimental noise and that rescaling is based on an unjustified assumption: that all alleles bind the same number of peptides. Removing this assumption, we have demonstrated a significantly improved predictive performance.
These conclusions are important both for studies that use prediction methods to understand the CTL response and for T cell epitope discovery programs where avoiding rescaling could save a large amount of experimental effort, ultimately leading to improved vaccine implementation.
|
10.1371/journal.pgen.1005521 | Cognitive Function Related to the Sirh11/Zcchc16 Gene Acquired from an LTR Retrotransposon in Eutherians | Gene targeting of mouse Sushi-ichi-related retrotransposon homologue 11/Zinc finger CCHC domain-containing 16 (Sirh11/Zcchc16) causes abnormal behaviors related to cognition, including attention, impulsivity and working memory. Sirh11/Zcchc16 encodes a CCHC type of zinc-finger protein that exhibits high homology to an LTR retrotransposon Gag protein. Upon microdialysis analysis of the prefrontal cortex region, the recovery rate of noradrenaline (NA) was reduced compared with dopamine (DA) after perfusion of high potassium-containing artificial cerebrospinal fluid in knockout (KO) mice. These data indicate that Sirh11/Zcchc16 is involved in cognitive function in the brain, possibly via the noradrenergic system, in the contemporary mouse developmental systems. Interestingly, it is highly conserved in three out of the four major groups of the eutherians, euarchontoglires, laurasiatheria and afrotheria, but is heavily mutated in xenarthran species such as the sloth and armadillo, suggesting that it has contributed to brain evolution in the three major eutherian lineages, including humans and mice. Sirh11/Zcchc16 is the first SIRH gene to be involved in brain function, instead of just the placenta, as seen in the case of Peg10, Peg11/Rtl1 and Sirh7/Ldoc1.
| Retrotransposon-derived DNA sequences occupy approximately 40% of the mammalian genome, compared with only 1.5% of protein coding genes. They have been commonly considered “junk DNA” and even potentially harmful for host organisms. However, a series of knockout (KO) mouse analyses demonstrated that at least some of the LTR retrotransposon- and retrovirus-derived sequences play essential roles in the current mammalian developmental system as endogenous genes, such as Peg10, Peg11/Rtl1, Sirh7/Ldoc1, SYNCYTINs and FEMATRIN-1, which are active in multiple aspects of placental function. Here we demonstrate that another LTR retrotransposon-derived gene, Sirh11/Zcchc16, plays an important role in cognitive function in the brain. Sirh11/Zcchc16 KO mice exhibit abnormal behaviors related to cognition, including attention, impulsivity and working memory, possibly due to the locus coeruleus-noradrenaline (LC-NA) system, suggesting that human SIRH11/ZCCHC16 may be involved in X-linked intellectual disability and/or attention-deficit/hyperactivity disorder. Comparative genome analysis demonstrates that SIRH11/ZCCHC16 was acquired in a common eutherian ancestor, suggesting that it contributed to eutherian brain evolution because it confers a critically important advantage in the competition that occurs in daily life. This study provides further insight into the impact of LTR retrotransposon-derived genes on mammalian evolution.
| Mammals, including human beings, have evolved a unique viviparous reproductive system using a placenta and a highly developed central nervous system. How did these unique characteristics emerge in mammalian evolution? Retrotransposons occupy approximately 40% of the mammalian genome. They recently have attracted attention as one of the driving forces of genomic evolution, providing novel endogenous genes [1–11] as well as rewiring the genetic network in the form of novel cis-elements, such as promoters, enhancers, insulators and transcription factor binding sites [12–16].
In a series of KO mouse experiments we have demonstrated that at least three LTR retrotransposon-derived genes are essential for mammalian development and reproduction via multiple placental functions; Peg10 is involved in the formation [8] and Peg11/Rtl1 in the maintenance of the placenta [9], while Sirh7/Ldoc1 is involved in endocrine regulation via the differentiation/maturation of a variety of placental cells [11], suggesting that they all have profoundly contributed to the evolution of viviparity during mammalian evolution [8–11, 17, 18].
Two major families of genes derived from Ty3/Gypsy LTR retrotransposons have been identified: one is the SIRH family, comprising the 11 genes mentioned above, while the other is the paraneoplastic MA antigen (PNMA) family derived from a gypsy_12DR-related retrotransposon comprised of at least 19 and 15 genes in humans and mice, respectively [7, 19–22]. It should be noted that Peg10 is the only gene commonly conserved in both the eutherians and marsupials [23], while all the others exist as eutherian- or marsupial-specific genes [11, 22, 24, 25]. Among the SIRH family, Sirh11/Zcchc16 (also called Mart4) is unique because it does not exhibit any placental expression during development, but rather, is expressed in the brain, testis, ovary and kidney.
In this work, we set out to examine whether Sirh11/Zcchc16 plays a role in organs other than the placenta, then generated and analyzed Sirh11/Zcchc16 KO mice. Interestingly, the Sirh11/Zcchc16 KO mice exhibited a variety of behavioral abnormalities related to cognition, indicating Sirh11/Zcchc16 is involved in brain function. We also found abnormal regulation of the NA level in the prefrontal cortex of KO mice. As the noradrenergic system in the LC in the brainstem sends projections to virtually all brain structures, including the prefrontal cortex of the cerebrum, and has been proposed to be involved in cognitive function, such as impulsivity, attention, working memory and their associated behaviors in mammals [26–29], we investigated the potential role of the Sirh11/Zcchc16 protein in the noradrenergic system, suggesting the relationship to human mental disorders and the impact on brain evolution in eutherian mammals.
Mouse Sirh11/Zcchc16 encodes a Gag-like protein comprising 304 amino acids with a typical CCHC RNA-binding motif at the C-terminus (Fig 1A). It exhibits 37.5% homology with the entire sushi-ichi retrotransposon Gag, consisting of 371 amino acids, except for the N-terminus. Sirh11/Zcchc16 is located on the X chromosome between Trpc5 and Lhfpl1. Its location is conserved in all of the eutherian lineages, euarchontoglires, laurasiatheria, afrotheria and xenarthra (Fig 1B and 1C). However, it became a pseudogene by frameshift and nonsense mutation in xenarthran species, such as the armadillo and sloth (S1 Fig). In the case of the armadillo (Dasypus novemcinctus), contig including pseudoSIRH11/ZCCHC16 is short and thus does not reach LHFPL1 or TRPC5. However, the presence of several evolutionarily conserved sequences (ECSs) in its surrounding 20 kb sequence confirms that it is orthologous to SIRH11/ZCCHC16 (Fig 1B, lower column). The absence of SIRH11/ZCCHC16 from marsupials, monotremes and birds was also confirmed, because there is no orthologous gene between TRPC5 and LHFPL1 in the opossum and Tasmanian devil, or between TRPC5 and AMOT in the chicken and platypus, respectively (Fig 1B and 1C). This indicates that the insertion of SIRH11/ZCCHC16 occurred in a common eutherian ancestor after the spilt of the eutherians and marsupials 160 million years ago (Ma), before the diversification of the three major eutherian lineages, boreoeutheria (including euarchontoglires and laurasiatheria), afrotheria and xenarthra, 120 Ma [30].
Importantly, the dN/dS ratio in the pairwise comparisons of SIRH11/ZCCHC16 orthologs between mouse and seven eutherian species other than xenarthran species is approximately 0.35~0.45 (< 1) (Table 1)[31], suggesting that SIRH11/ZCCHC16 has been subjected to purifying selection after its domestication (exaptation) in the common eutherian ancestor. Furthermore, to examine whether the functional constraint is relaxed in the xenarthran lineage, we also performed an analysis using the Phylogenetic Analysis by Maximum Likelihood (PAML) program for comparing two models [31]. One is that all the twelve species including the four xenarthran species have the same ω value (dN/dS ratio) (model 1). The other is based on the assumption that the ω value in the xenarthran branch, including the two armadillo and two sloth species, is different from that of all the other eight eutherian species (model 2). The analysis demonstrated that model 2 is statistically significant (p = 8.80E-04): ω2 = 1.11 for the xenarthran branch, with ω1 = 0.766 for all of the others (Fig 1D), that is, relaxed or neutral evolution is ongoing in xenarthra. All of these results indicate that SIRH11/ZCCHC16 is a protein-coding gene in the eutherians except in xenarthra.
Sirh11/Zcchc16 is basically comprised of 7 exons, with its ORF in the last exon (Fig 2A). According to the NCBI database, there are at least three variants with different first exon sequences, presumably dependent on the tissues and organs where it resides. A low level of Sirh11/Zcchc16 expression was observed in the brain, liver and heart on embryonic day 14.5 (d14.5), with a moderate level of expression in the brain, kidney, testis and ovary in adults (8 weeks (8 w)) (Fig 2B). We generated Sirh11/Zcchc16 KO mice using TT2 ES cells by means of a complete deletion of its protein coding sequence (S2 Fig). After removing the neomycin cassette, Sirh11/Zcchc16 KO mice were backcrossed to B6 more than 10 generations. We confirmed the lack of the ORF region by RT-PCR using a primer set (F5R5). However, it should be noted that the RT-PCR experiment using primer sets amplifying its 3’-UTR region (such as F9R9) exhibited an approximately 1.5 fold increment in these organs in the KO mice (although not significant in the whole brain or kidney), suggesting the existence of a feedback mechanism regulating Sirh11/Zcchc16 at the protein level (Fig 2A and 2C).
Sirh11/Zcchc16 KO mice did not exhibit lethality or growth retardation in the pre- and postnatal periods in either female (homo) or male (null) KO mice (Tables 2 and 3). Despite the relatively higher expression of Sirh11/Zcchc16 in the testis and oocyte, both the male and female KO mice were fertile, even in the case of mating between female homo KOs and null male mice, indicating that Sirh11/Zcchc16 has no apparent role in sperm and/or egg production (Table 2). No abnormalities were detected in the urine of the Sirh11/Zcchc16 male KO mice (10 w), such as pH or the amount of glucose, total protein, urobilinogen, ketone body, bilirubin and occult blood, suggesting that kidney function is also normal in these KO mice (S3 Fig).
Although no evident structural abnormalities were found in the KO mouse brain (S4 Fig), it was noticed that abnormal behaviors were exhibited. For example, they displayed agitated movement in their cages when staff personnel entered the breeding room, and sometimes jumped out when their cages were exchanged. Therefore, comprehensive behavior tests were carried out using 8–10 w males. Sirh11/Zcchc16 KO mice exhibited no abnormality in the open field test (9 and 10 w), but in the Light/Dark transition test (8 w) the latency before entering into the light chamber was significantly decreased (unpaired two sample t-test, t(12) = 2.52, p = 0.0269), while the number of transitions was significantly increased (unpaired two sample t-test, t(12) = -2.58, p = 0.0242) compared to the wild type (WT), suggesting a reduced attention and/or enhanced impulsivity (Fig 3A). They also exhibited significantly higher activity during the dark period of the home-cage activity test (10 w), especially just after the “light to dark (Zeitgeiber time (ZT) 12 and 13)” (effect of genotype, F(1, 60) = 7.86, p = 0.00681 and effect of genotype, F(1, 60) = 5.35, p = 0.0241, respectively) as well as just before the “dark to light (ZT23)” (F(1, 60) = 9.77, p = 0.00274) transition periods. There was also lower activity at ZT19 (F(1, 60) = 7.56), p = 0.00788), while there was no significant difference in the light periods (ZT0-ZT11), suggesting hyperactivity, especially when the light conditions are changed or about to be changed (Fig 3B).
Next, the Y-maze test was conducted to assess spatial working memory by recording spontaneously alternating search behavior during a 5-min session in a Y-maze (16–18 w). It is considered that alternating search behavior reflects a primitive working memory capacity because it is based on the tendency of normal animals to enter the arm of the Y-maze which was least recently explored [32]. The Y-maze test also allows the simultaneous assessment of hyperactivity independently of spatial working memory. Hyperactivity may interfere with learning and memory, therefore, its assessment is crucial for the interpretation of memory test results [33]. Importantly, KO mice exhibited a lower level of alternation (Mann-Whitney U test, p = 0.032), although the total number of arm entry events was the same (Mann-Whitney U test, p = 0.358) as the WT controls, suggesting that they have a poor working memory (Fig 3C). Some of the KO mice jumped out of the Y-maze stage before the test started, while none of the control mice exhibited such behavior. Although we excluded this data from the results, this also appears to indicate that these KO mice tend to exhibit extreme behavior when transferred to a new environment. From these results, we reasoned that Sirh11/Zcchc16 KO mice have some abnormality in cognition, possibly related to monoamine function in the brain, that impacts impulsivity, attention and/or memory.
We used null KO male mice for behavioral tests. Although Sirh11/Zcchc16 is also expressed in the kidney, testis and ovary, there was no evidently unusual phenotype in these organs and also no viability or growth effects observed, as mentioned. Therefore, we think that the abnormal behaviors observed in the KO mice mainly reflect an impairment of brain function. In addition, we used mice generated by in vitro fertilization (IVF) between KO males and hetero KO females for these behavioral tests, as described in Materials and Methods. The fertilized eggs were transferred to the pseudopregnant ICR females, after which the pups were born naturally and taken care of by the ICR mothers. Thus, harmful effects on the pups from the potentially abnormal KO mothers in terms of their nursing behavior were completely excluded. Therefore, we believe that these results reflect a difference in genotype.
We carried out microdialysis analysis in the prefrontal cortex of the cerebrum to directly examine the monoamine levels in the KO brain [34–35] because it is well documented that prefrontal cortical NA as well as dopamine (DA) plays an important role in spatial working memory [36]. The levels of various monoamines were measured, mainly at 11~13 w of age, including DA, NA, adrenaline (AD), 3-methoxytyramine (3-MT), 5-hydroxyindole acetic acid (5-HIAA), serotonin (5-HT), 3, 4-dihydroxyphenylacetic acid (DOPAC), homovanillic acid (HVA), 3-methyl-4-hydroxy-phenylglycol (MHPG) and normetanephrine (NM) using isoproterenol (ISO) as the control. During overnight experiments for long-term recording, we found that these levels varied suddenly and unexpectedly, presumably because of abrupt changes in sound and vibration in the experiment room. The levels were also greatly dependent on the initial condition of the mice. As it was difficult to obtain reproducible data in the normal experimental setting, we adopted the perfusion method to overcome this difficulty. Perfusion by high-K solution intensively stimulates the release of neurotransmitters and hence enables an evaluation of the sum of neurotransmitter content at the synapse and in the neuron. High K solution was applied twice, with a 120-min interval. The release of neurotransmitter in response to the second perfusion of the high K solution is dependent upon the catabolic activity during a 120 min period. It was demonstrated that the recovery of the NA level was significantly delayed compared with the DA level (Mann-Whitney U test, p = 0.0159), while the levels of two other DA metabolites, DOPAC and 3-MT (Mann-Whitney U test, p = 0.532 and 0.310, respectively), were unchanged in the KO mice (Fig 4A).
DA is catabolized in three different pathways to NA, DOPAC and
3-MT by the enzymes dopamine beta monooxigenase (Dbh), monoamine
oxidase (MAO) and catechol-O-methyltransferase (COMT), respectively
[37]. Dbh exhibits a dominant expression pattern in the brainstem, approximately 32 fold greater than the cerebrum (Fig 4B) although its brainstem level was not affected in the Sirh11/Zcchc16 KO. Sirh11/Zcchc16 also exhibited a relatively higher expression in the mesencephalon, diencephalon and brainstem compared with other parts of the brain (Fig 4C). Interestingly, both the Dbh and Sirh11/Zcchc16 expression levels fluctuated in the brainstem, but their expression patterns exhibited a negative correlation (i.e. the test for non-correlation, Pearson correlation coefficient (r) = −0.591, p = 0.033, Fig 4D). Therefore, it is possible that they are reciprocally regulated by the same environmental signals. Although the precise biological role of Sirh11/Zcchc16 as well as the relationship to Dbh in the brain are presently unknown, all of these data suggest that behavioral defects of the Sirh11/Zcchc16 KO mice are somehow related to a dysregulation of the noradrenergic system in the brain.
We used two cohorts of mice, one for the comprehensive behavioral tests, including biochemical analyses, and the other for expression analysis in a range from embryos to adult tissues, as well as microdialysis and the Y-maze test. We believe that the results of the monoamine analysis (microdialysis) accurately reflect the Y-maze test. They also are consistent with the Light/Dark transition and home-cage activity tests in the first cohort of mice.
It has been proposed that the LC-NA system plays an important role in cognitive function, presumably by regulating the balance between focused versus flexible responding, or selective versus scanning attention [25–28]. In the phasic mode, LC cells exhibit selective phasic activation for target stimuli, but only a moderate level of tonic discharge, leading to excellent performance on the specific task with few errors and focused, selective attention. In the tonic mode, the LC cells fail to respond phasically to any task stimuli, but rather, exhibit higher levels of ongoing tonic activity, leading to poor performance with many errors and a form of scanning, labile attention. In their overarching theory, Bouret and Sara proposed that phasic activation of the NA neurons of the LC takes place in time with the cognitive shifts that facilitate dynamic reorganization of target neural networks, permitting rapid behavioral adaptation to the demands of changing environmental imperatives [27].
In this work, Sirh11/Zcchc16 KO mice exhibited certain clearly evident behavioral abnormalities, such as increased activity during the light/dark transition test, higher daily activities in the period just after dark as well as just before light and a lower score in alternating search behavior in the Y-maze test, indicating that Sirh11/Zcchc16 plays a role in cognition. Together with certain other sudden and unexpected movement activity that was observed, it is likely that Sirh11/Zcchc16 KO mice tend to exhibit behaviors related to the classic tonic LC mode. In support of this idea, a significantly lower recovery rate of NA compared to DA in the prefrontal cortex was observed in a microdialysis analysis of the KO mice after perfusion with high potassium solution.
Importantly, activation of the LC-NA system is also associated with an increased accuracy of the response to task-relevant stimuli [26]. Using microdialysis analysis, Rossetti and Carboni demonstrated that both prefrontal cortical DA and NA are involved in the modulation of working memory [36]. From an application of the T-maze test using rats that analyzes such function along with memory and spatial learning via an application of various stimuli, they demonstrated that the prefrontal cortical DA and NA dialysate levels are both phasically increased when rats perform correctly in a delayed alternation task in a T-maze. Together with the findings from other experiments, they ultimately concluded that DA is primarily associated with reward expectancy, whereas NA is involved in the active maintenance of goal-related information as well as the rules for realizing the goal [36]. This is in good accord with the results of our Y-maze test, indicating that Sirh11/Zcchc16 KO mice with a lower NA recovery rate in the neurons of the prefrontal cortex have impaired spatial working memory.
It has also been proposed that the phasic as opposed to tonic LC activity participates in certain critically important normal behavioral functions as well as severe mental problems, including attention-deficit/hyperactivity disorder (ADHD) and a variety of emotional and affective disorders [25–28]. Therefore, the behavioral abnormalities observed suggest a possible role for SIRH11/ZCCHC16 in mental disorders.
Although a relatively higher expression level was observed in the brainstem, diencephalon and mesencephalon in the RT-PCR experiment, we have no direct evidence at present as to precisely where Sirh11/Zcchc16 is expressed in the brain or whether the putative Sirh11/Zcchc16 protein in the NA neurons from the LC. This is despite numerous attempts using in situ, Western blotting and immunostaining. However, the conservation of the amino acid sequence of the Sirh11/Zcchc16 in both boreoeutheria and afrotheria clearly demonstrates that it is subjected to purifying selection (dN/dS <1), providing indirect but nonetheless supportive evidence that Sirh11/Zcchc16 is a protein-coding gene (Table 1). This conclusion is consistent with the finding that the xenarthran pseudoSirh11/Zcchc16 bearing many mutations in its coding frame has been subjected to relaxed or neutral evolution (dN/dS ~1) (Fig 1D). The Sirh11/Zcchc16 protein possesses a conserved CCHC zinc finger domain with potential RNA-binding capability. Therefore, it is possible that it functions as a part of messenger ribonucleoprotein particles (mRNPs) in neurons because mRNAs synthesized in the soma are transported to neurites and/or synapses as mRNPs by binding to RNA-binding proteins and are translated there [38–40]. However, since the possibility that Sirh11/Zcchc16 functions as a non-coding RNA has not been completely excluded, these issues need to be further addressed in more detail in the future.
Human SIRH11/ZCCHC16 is located on Xq23, where several X-linked intellectual disability (XLID) genes have been mapped, such as PRPS1 (–5 Mb from SIRH11/ZCCHC16), ACSL4 (–2.5 Mb), PAK3 (–1 Mb), DCX (–1 Mb), AGTR2 (+4 Mb), LAMP2 (+8 Mb) and GRIA3 (+11 Mb). As some of the genes responsible for XLID remain to be identified in this chromosomal region, SIRH11/ZCCHC16 may also be a good candidate for XLID [41, 42]. Interestingly, Cho et al. recently reported that some XLID patients have mutations in the SIZN1/ZCCHC12 (PNMA10) gene that locates approximately 6 Mb downstream of SIRH11/ZCCHC16 and indicated this gene to be a good candidate for XLID [43]. PNMA10/ZCCHC12 is another eutherian-specific Ty3/Gypsy LTR retrotransposon-derived gene that is known to be involved in transcriptional regulation [44]. Mouse Pnma10/Zcchc12 is expressed in the embryonic ventral forebrain in a cholinergic-neuron-specific manner and binds to SMAD family proteins. It also acts as a transcriptional co-activator for bone morphogenic protein (BMP) signaling [44]. Both Sirh11/Zcchc16 and Pnma10/Zcchc12 encode Gag-like proteins with the CCHC zinc-finger motif, so it will be of interest to determine whether these two retrotransposon-derived proteins have different activities which function in a noradrenergic- and a cholinergic-neuron-specific manner, respectively, in the extant eutherian mammals.
Comparative genome analysis clearly demonstrated that SIRH11/ZCCHC16 is highly conserved in the three major groups of eutherian mammals, but not in xenarthra, strongly implying that the presence of Sirh11/Zcchc16 is beneficial in these three eutherian groups, including humans and mice. Our results suggest that SIRH11/ZCCHC16 contributed to the evolution of the brain by modulating the NA neuronal network in a complex manner. The noradrenergic system is also conserved in other vertebrates, such as fish, amphibians, reptiles and birds, and the general system characteristics are strikingly preserved across a wide range of phylogenetic groups [27]. Therefore, the biological function of the Sirh11/Zcchc16 protein is of great interest in terms of elucidating the evolution of the neuromodulatory system of the brain in the eutherian mammals. What was the function that was replaced in the noradrenergic system by domestication of SIRH11/ZCCHC16 in these three eutherian lineages, thus permitting rapid behavioral adaptation to changing environmental imperatives? Do the xenarthran species, i.e. the armadillos and sloths, have diminished and/or distinct cognitive activity related to SIRH11/ZCCHC16? Although a variety of issues of this type require future investigation, this is the first demonstration that one of the SIRH genes plays a role in cognitive function in the brain, presumably via the noradrenergic system.
We previously demonstrated that the three SIRH genes, such as Peg10/Sirh1, Peg11/Rtl1/Sirh2 and Sirh7/Ldoc1, play essential roles in eutherian development and reproduction, and proposed that these three SIRH genes have profoundly contributed to the evolution of viviparity in mammalian evolution as newly acquired genes [8–11]. In addition, the present investigation of Sirh11/Zcchc16 provides further insight into the impact of LTR retrotransposon-derived genes on the neuromodulatory system in the brain as key step in the evolution of the eutherian mammals.
The present animal experiments were performed in strict accordance with the guidelines of Tokai University and Tokyo Medical and Dental University (TMDU), and were approved by the Animal Investigation Committees of Tokai University and TMDU. The open-field test, Light/dark transition test, home-cage activity test and urinalysis were performed in accordance with the guidelines issued by the RIKEN Bioscience Technology Center in their “Outline for Conducting Animal Experiments” (August 1999, revised October 2001).
Sushi-ichi gag (AAC33525.1) and mouse Sirh11/Zcchc16 (NP_001028967.2) protein sequences were obtained from NCBI. Amino acid identity and similarity were calculated using the EMBOSS Water program (http://www.ebi.ac.uk/Tools/psa/emboss_water/) in the default mode. The orthologues of SIRH11/ZCCHC16 were identified by search of NCBI Gene (http://www.ncbi.nlm.nih.gov/gene/) using ZCCHC16 as the query term. Genomic homology analysis was performed using the mVISTA LAGAN program (http://genome.lbl.gov/vista/mvista/submit.shtml). We obtained TRPC5-AMOT genomic sequences from the NCBI database. The sequences used for analysis were the following: Chicken (Gallus gallus): gi|358485508:c13102445-12913720; Platypus (Ornithorhynchus anatinus): gi|149729612:c11732082-11450308; Opossum (Monodelphis domestica): gi|126362945:c69157606-68740637; Mouse (Mus musculus): gi|372099090:144381671–145505458; Human (Homo sapiens): gi|568815575:111774314–112840908; Dog (Canis lupus familiaris): gi|357579592:84841858–85807530; African savanna elephant (Loxodonta Africana) gi|343530165:c19381570-18463351; Armadillo (Dasypus novemcinctus) gi|476561443; Sloth (Choloepus hoffmanni) gi|692243298|gb|KN194663.1|. PseudoSIRH11/ZCCHC16 protein sequences were aligned to Florida manatee SIRH11/ZCCHC16 using Clustal Omega (http://www.ebi.ac.uk/Tools/msa/clustalo/).
Genomic DNA was isolated from frozen muscle using the DNeasy Blood & Tissue Kit (QIAGEN). For PCR, the primers were designed at the 5'- and 3'-UTR of pseudoSIRH11/ZCCHC16 using the consensus sequence between Dasypus novemcinctus and Choloepus hoffmanni. The PCR reaction was performed using ExTaqHS (TaKaRa) with the following conditions: 30 cycles of 98°C, 10 sec; 60°C, 30 sec; 72°C, 1 min. The following PCR primers were used: pseudoSIRH11-F1: 5'-CTTACTGCCTGCCCATTGGT-3' and pseudoSIRH11-R1: 5'-GGATTTTAAAAGTTGGTGCAGG-3'. PCR products were direct-sequenced using the above primers after Exo-SAP-IT (USB) treatment. DNA Data Bank of Japan (DDBJ) accession numbers: LOC064756 for Tolypeutes matacus SIRH11/ZCCHC16 and LOC064757 for Choloepus didactylus SIRH11/ZCCHC16.
A phylogenic tree was constructed with ClustalW2 (Neighbor-joining method) (http://www.ebi.ac.uk/Tools/msa/clustalw2/) using protein coding and pseudo SIRH11/ZCCHC16 sequences obtained from twelve species. The codon alignment of cDNA was created with the PAL2NAL program (http://www.bork.embl.de/pal2nal/) [45]. The non-synonymous/synonymous substitution rate ratio (ω = dN/dS) was estimated by using CodeML in PAML [31].
To generate Sirh11/Zcchc16 MT mice (Accession No. CDB0557K: http://www.clst.riken.jp/arg/mutant%20mice%20list.html), we obtained three genomic fragments, the 5’-arm (8 kb: 145111478–145119486; NC_000086), middle arm (1.1 kb: 145119487–145120587; NC_000086) and 3’-arm (2.1 kb: 145120588–145122703; NC_000086) by recombination from the RP23-319K12 BAC clone (BACPAC Resources), and then cloned them into a targeting vector. The targeting vector was introduced into TT2 ES cells (C57BL/6 × CBA genetic background) by electroporation [46]. To generate chimeric mice, ES cells in which homologous recombination had occurred were injected into 8-cell stage embryos. Germ line transmission of the Sirh11/Zcchc16 MT allele was confirmed by Southern blot and PCR using the genome prepared from pups in which male Sirh11/Zcchc16 chimeric mice had been crossed with female C57BL/6J. To remove the flox region, we injected a pCAG/NCre plasmid [47] into the fertilized eggs generated by in vitro fertilization (IVF) from Sirh11/Zcchc16 MT hetero eggs and C57BL/6J sperm, thus establishing Sirh11/Zcchc16 neo mice. To obtain Sirh11/Zcchc16 KO mice, we injected a pCAGGS-FLPe plasmid (Gene bridge) into the fertilized eggs generated by IVF from Sirh11/Zcchc16 neo hetero eggs and C57BL/6J sperm. Exclusion of the neo cassette was confirmed by genomic PCR of the pups’ DNA. Southern blot analysis was performed using a standard protocol. Five micro grams of genomic DNA from the tail were digested by restriction enzymes NheI (for the 5’ probe) and NcoI (for the 3’ probe), respectively. Hybond-N+ (GE Healthcare) membranes blotted with digested DNA were hybridized in Church buffer with radio isotope-labelled probes. The 5’ and 3’ probes were generated by genomic PCR using the following sequences: 5’ probe: 145109101–145110193; NC_000086; 3’ probe: 145124154–145124725; NC_000086. The Sirh11/Zcchc16 KO allele was detected by genomic PCR. Genome DNA was prepared from the tail or ear tip using a DNeasy Blood & Tissue Kit (QIAGEN). PCR was performed using ExTaqHS polymerase (TaKaRa) with the following primers: Sirh11-F1: 5’-ATGTATCCTAAGGTGATCCG-3’ and Sirh11-R2: 5’-ATGTGATGCCACAGCAACTC-3’. Sirh11/Zcchc16 KO mice were backcrossed to C57BL6/J for more than 10 generations.
Total RNA was prepared from frozen tissues using ISOGEN (NIPPON GENE) and ISOGEN-LS (NIPPON GENE). The cDNA was made from total RNA (1 μg) using Revertra Ace qPCR RT Master Mix (TOYOBO). Quantitative RT-PCR analysis was performed using Fast SYBR Green Master Mix (Life technologies) and a StepOnePlus System (Life technologies) by means of an absolute quantification method. Data was normalized by Actb expression. Student’s t-test was used for statistical analysis. The following primer sequences were used: Sirh11-F9: 5’-TGGTGCTGGTGTATTTCCCC-3’ and Sirh11-R9: 5’-TGGCACAGTGGTTAGTGAGGC-3’; Sirh11-F5: 5’-AAGAGGAGGATAGGAAATCACTTTG-3’ and Sirh11-R5: 5’-GTTGTTAGGACAAGGTTGAGG-3’; Dbh-F1: 5’-ACTGAACGGAGAAGCCCTGGAC-3’ and Dbh-R1: 5’-CACCAGAGGACCAACAGGGTCG-3’; Actb-F: 5’-AAGTGTGACGTTGACATCCG-3’ and Actb-R: 5’-GATCCACATCTGCTGGAAGG-3’.
To reproduce the hetero and wild-type progeny for the behavior-screen, in vitro fertilization (IVF) was performed. Wild type male mice were used as the source for sperm, while hetero female mice were used as the source for the oocytes used for IVF. ICR female mice (CLEA Japan, Tokyo, Japan) were used as pseudopregnant recipients for embryo transfer. Sperm were collected from the caudae epididymides of adult male mice (20 w) and allowed to diffuse in human tubal fluid (HTF) medium. After preincubation for approximately 1 hour to allow for capacitation, the sperm were used for insemination. Meanwhile, immature hetero female mice (4 w) were superovulated using intraperitoneal injections of PMSG and HCG (Serotropin and Gonatropin; ASKA Pharmaceutical Co., Tokyo, Japan) with an interval of 48 hours between injections. Approximately 15–17 hours after the HCG injection, the oocyte-cumulus complexes were collected from the oviducts of the superovulated female mice. The complexes from several female mice were then placed in the HTF fertilization medium. Insemination was performed by adding the sperm suspension to the fertilization medium containing complexes and culturing at 37°C with 5% CO2 in air. Twenty-four hours after insemination, 2-cell embryos were transferred into the oviducts of pseudopregnant ICR females mated to vasectomized ICR males. All pups were delivered naturally after embryo transfer. After weaning at four weeks age, they were employed as breeding individuals in a single-breeding cage.
Each mouse was placed in the corner of an open-field apparatus (400 mm wide x 400 mm long x 300 mm high; O’Hara & Co., Ltd., Tokyo, Japan) made of white polyvinyl chloride. The distance traveled by each animal in the open field was recorded for 20 min with a video-imaging system (Image OF9; O’Hara & Co., Ltd., Tokyo, Japan). The mice were tested on two separate occasions at 9 and 10 w. The unpaired two sample T-test was used for statistical analysis.
A commercially available light/dark chamber (O'Hara & Co., Ltd.) was used for the light/dark transition test. The apparatus consists of a light chamber (200 mm long × 200 mm wide × 250 mm high) made of white vinyl chloride plates and a dark chamber with the same dimensions made of black vinyl chloride plates. The apparatus has an opening (50 mm wide × 30 mm high) in the middle of the wall that joins the two chambers. The opening is controlled by a guillotine door. The latency for entering into the lighted chamber and number of transitions between the light and dark chambers were measured. The mice were tested at 8 weeks of age. The unpaired two sample T-test was used for statistical analysis.
Each mouse was placed alone in a testing cage (227 mm wide x 329 mm long x 133 mm high) under a 12-h light–dark cycle (light on at 08:00 h) and had free access to both food and water. After 1 day of acclimation, spontaneous activity in the cage was measured for 5 continuous days (starting at 08:00) with an infrared sensor (activity sensor, O’Hara & Co., Ltd.). The mice were tested at 10 w. The two-way analysis of variance (ANOVA) (effects of genotype and date) was used for statistical analysis.
The Y-maze test was performed using male mice at 16–18 weeks of age (WT: N = 6, KO: N = 8). The apparatus was a black, plastic maze with three arms (400 mm long × 30 mm wide × 150 mm high, 120 degrees). Mice were placed at the center of the apparatus and allowed to move freely through the maze for 2 min. The sequence and total number of arm entries were recorded by video camera for 5 min. When all 4 limbs of the mouse were within a pathway, it was considered an entry. An “alternation” was counted when a mouse successively entered 3 different arms. Spontaneous alternating search behavior was calculated by the following equation: alternation behavior (%) = [the number of alternations/(total number of arm entries – 2)] × 100. The results are given as the mean and standard error of the mean (S.E.M.). Statistical analysis was conducted using computer software (Prism, version 6.0c, GraphPad Software, San Diego, CA, USA) for a comparison across the experimental conditions. Statistical evaluations for the measurement of behavior or the NA level were carried out using Mann-Whitney U test. A P-value <0.05 was considered to be significant.
We performed in vivo microdialysis measurements of extracellular monoamines in the prefrontal cortex of 11–13 week-old mice. A guide cannula (AG-4; EICOM, Kyoto, Japan) was implanted into the prefrontal cortex (+1.9 mm anteroposterior and +0.5 mm mediolateral relative to the bregma and −2.0 mm dorsoventral relative to the dura of the skull) under inhalation anesthesia with nitrous oxide, oxygen and isoflurane (2%). Two days after surgery, a dialysis probe (AI-4-01, 1-mm membrane length; EICOM) was inserted through the guide cannula and perfused at a flow rate of 1 μl/min with artificial cerebrospinal fluid (147.0 mM NaCl, 4.0 mM KCl, 2.3 mM CaCl2) or high potassium-containing artificial cerebrospinal fluid (51.0 mM NaCl, 100.0 mM KCl, 2.3 mM CaCl2). Samples were collected every 20 min and injected directly into an HPLC column (EICOMPAK CA-5ODS; EICOM) by an auto injector (EAS-20; EICOM). The concentrations of the monoamines in the dialysate were determined by HPLC with an electrochemical detector (ECD300; EICOM). After the experiments, 1 μl of 0.3% Evans blue dye was microinjected through the cannula to histologically verify the position of the probe, and only data from animals with a correct probe placement were used in the analysis. The results are given as the mean and standard error of the mean (S.E.M.) of the data as described in the Y-maze test.
|
10.1371/journal.pcbi.1004162 | Twisting Right to Left: A…A Mismatch in a CAG Trinucleotide Repeat Overexpansion Provokes Left-Handed Z-DNA Conformation | Conformational polymorphism of DNA is a major causative factor behind several incurable trinucleotide repeat expansion disorders that arise from overexpansion of trinucleotide repeats located in coding/non-coding regions of specific genes. Hairpin DNA structures that are formed due to overexpansion of CAG repeat lead to Huntington’s disorder and spinocerebellar ataxias. Nonetheless, DNA hairpin stem structure that generally embraces B-form with canonical base pairs is poorly understood in the context of periodic noncanonical A…A mismatch as found in CAG repeat overexpansion. Molecular dynamics simulations on DNA hairpin stems containing A…A mismatches in a CAG repeat overexpansion show that A…A dictates local Z-form irrespective of starting glycosyl conformation, in sharp contrast to canonical DNA duplex. Transition from B-to-Z is due to the mechanistic effect that originates from its pronounced nonisostericity with flanking canonical base pairs facilitated by base extrusion, backbone and/or base flipping. Based on these structural insights we envisage that such an unusual DNA structure of the CAG hairpin stem may have a role in disease pathogenesis. As this is the first study that delineates the influence of a single A…A mismatch in reversing DNA helicity, it would further have an impact on understanding DNA mismatch repair.
| When a set of 3 nucleotides in a DNA sequence repeats beyond a certain number, it leads to incurable neurological or neuromuscular disorders. Such DNA sequences tend to form unusual DNA structures comprising of base pairing schemes different from the canonical A…T/G…C base pairs. Influence of such unusual base pairing on the overall 3-dimensional structure of DNA and its impact on the pathogenesis of disorder is not well understood. CAG repeat overexpansion that leads to Huntington’s disorder and several spinocerebellar ataxias forms noncanonical A…A base pair in between canonical C…G and G…C base pairs. However, no detailed structural information is available on the influence of an A…A mismatch on a DNA structure under any sequence context. Here, we have shown for the first time that A…A base pairing in a CAG repeat provokes the formation of left-handed Z-DNA due to the pronounced structural dissimilarity of A…A base pair with G…C base pair, leading to periodic B-Z junction. Thus, these results suggest that formation of periodic B-Z junction may be one of the molecular bases for CAG repeat instability.
| Apart from the ‘canonical’ B-DNA conformation, DNA can also adopt a variety of ‘non-canonical’ conformations such as hairpin, triplex and tetraplex depending on the sequence and environment. It is well known that formation of such unusual non-B-DNA structures during the overexpansion of trinucleotide microsatellites (tandem repeats of 1–3 nucleotide length) is responsible for at least 22 incurable trinucleotide repeat expansion disorders (TREDs) that are mainly neurological or neuromuscular in nature[1,2,3,4,5]. For instance, occurrence of hairpin structure due to the abnormal increase in the CTG repeat length in the untranslated region of DMPK gene causes myotonic dystrophy type-1[6,7]. Likewise, hairpin formation in CAG repeat expansion located in the protein-coding region leads to Huntington’s disorder & several spinocerebellar ataxias[7]. Direct evidence for the role of such hairpin structure in instigating replication-dependent instability has been demonstrated for the first time in human cells with 5’CTG.5’CAG microsatellite overexpanion[8]. Recently, it has been shown that CAG repeat overexpansion in DNA leads to toxicity by triggering cell death[9,10] and thus, warranting a detailed investigation on the hairpin structures formed under such abnormal expansion.
Although diverse mechanisms at DNA, RNA and protein levels have been identified for the progression of TREDs[11], until now, the main focus as potential therapeutic targets has been on RNA and protein levels. In fact, crystal structures of RNA duplex (hairpin stems) containing CUG[12] and CAG[13] repeats that form noncanonical U…U[12] & A…A[13] base-pairs offers useful information as the pathogenic CUG and CAG RNA hairpins have a role in misregulating the alternative splicing by MBNL1[14], leading to neurotoxicity. Though the isosequential DNA also intends to form hairpin structure[15], detailed structural insights about DNA duplex with CAG and CTG repeats that form A…A and T…T mismatches respectively are still inaccessible. With emerging evidence on ‘DNA toxicity’ of CAG repeat overexpansion[9,10], such structural information would facilitate the understanding of underlying mechanisms behind repeat instability at DNA level which is yet another potential drug target. In this context, we aim here to investigate the structure and dynamics of DNA duplex containing CAG repeat using molecular dynamics (MD) simulation technique. Surprisingly, results of the MD simulations indicate that A…A mismatch in a CAG repeat overexpansion induces periodic B-Z junction irrespective of the starting conformation. Thus, we suggest that such an unusual DNA structure of CAG hairpin stem may affect the biological function and may be one of the factors responsible for ‘DNA toxicity’ [9,10].
Role of a single noncanonical A8…A23 pair amidst canonical base pairs (Fig. 1A) is investigated through 300ns MD simulation, prior to the investigation of CAG repeats with periodic A…A mismatch as in Huntington’s disorder. As CAG repeat containing RNA crystal structures[13],[16] exhibit two different glycosyl conformations for A…A mismatch, 2 starting models with N6(A23)…N1(A8) hydrogen bond are considered for MD simulation: one with anti…anti (~250°) and the other with +syn(~79°)…anti(~250°) base glycosyl conformation.
Root mean square deviation (RMSD) calculated over 300ns simulation indicates the existence of three different ensembles (Fig. 1B): the first ensemble persists till ~16.5ns with RMSD centered around 2.8(0.7)Å, the second one persists between 16.5-181ns with a RMSD of 4.7(0.7)Å and the third one persists beyond ~181ns with the highest RMSD of 6.2(0.8)Å.
Intriguingly, a high RMSD of 4.5(0.6)Å observed between 16.5-100ns is associated with a change in glycosyl conformation of mismatched A23 and A8 from the starting anti conformation to -syn conformation.
During the first 16.5ns, A8 and A23 fluctuate between -syn and anti glycosyl conformations. Beyond 16.5ns, both A8 [-38(17)] and A23 [-66(17)] stay in -syn conformation (Fig. 1C). Similar tendency is also seen in the neighboring G24, wherein, it prefers -syn [309 (15°)] conformation beyond 16.5ns (Fig. 1C). Thus, it is clear that A8…A23 mismatch disfavors anti…anti glycosyl conformation and causes distortion in the duplex.
Aforementioned conformational changes in chi are accompanied by transformations in sugar-phosphate backbone at and around the mismatch site. For instance, during the first 100ns simulation, A8G9&G24C25 steps exhibit the characteristics of Z-DNA. The conformational angles (ε,ζ,α,γ) at A8G9 favor (g-,g+,g+,trans) [283(11°), 83(14°), 99(48°), 181(36°)] (Fig. 1D). Similar tendency is seen at G24C25 step with (ε,ζ,α,γ) favoring (293(15°), 89(13°), 79(14°), 199(13°)) conformation (Fig. 1E). These conformational rearrangements lead to transformation from right-handed B to left-handed Z form at the A8…A23 mismatch site leading to the formation of B-Z junction. These changes happen mainly due to the sugar phosphate flipping (S1–S3 Movie), which can clearly be seen from the repositioning of O4’ atoms (Fig. 1F, colored blue) of A8&A23 sugars as well as the sugar-phosphate backbone (Fig. 1F, indicated in arrow).
Strikingly, the effect of left-handed Z-DNA conformation observed between 16.5-100ns is also reflected in the helical twist angle of C7A8.A23G24 step which favor low (negative) twist of −4° (7) (Fig. 1G) flanked by high (positive) twists at the neighboring G6C7 (32 (4°)) & A8G9 (31(6°)) steps (S1 Fig). These, together with the conformational changes at A23…A8 mismatch reflect in the helicity of the duplex, which can be clearly seen from the superposition of average structures calculated over 1-100ps and 14.9-15ns (Fig. 1H). While the former is in B-form conformation, the latter shows a change in helicity leading to local Z-DNA formation. Occurrence of a low negative twist due to local Z-DNA formation in the midst of high twists at G6C7A8G9 stretch leads to local unwinding of the helix as can be seen Fig. 1I. As A23…A8 mismatch site is located exactly in the middle of DNA (Fig. 1A), aforementioned distortions lead to Z-DNA sandwich, viz., a mini Z-DNA is embedded in a B-DNA. Essentially, similar features are observed in B-Z junction formed by L-deoxy guanine and L-deoxy cytosine (S1 Fig).
As the Z-DNA formation happens due to the sugar-phosphate flipping, hydrogen bond between A8&A23 undergoes minor changes (S2 Fig). During the first ~16.5ns, N1(A8)…N6(A23) hydrogen bond persists, whereas, between 16.5-100ns, N1(A23)…N6(A8) hydrogen bond is predominantly favored due to the slight movement of A23 towards the minor groove. Base extrusion at the mismatch site is also observed during 100ns simulation.
B-Z junction induced by A8…A23 mismatch propagates to the neighboring bases (A5 to A11) beyond 181ns (Fig. 1I-J), which reflects in the highest RMSD of 6.2Å (Fig. 1B). Though the chi angle at A8, A23 and G24 remain in -syn conformation (Fig. 1C) as the 1st 100ns simulation (see above), (ε,ζ,α,γ) at the A8G9 step takes up (trans,g-,g-,g+) with C7A8.A23G24 step adopting a slightly higher helical twist of 11.1(9°) (Fig. 1G). However, (ε,ζ,α,γ) at G3C4, G6C7, G9C10, A11G12, G21C22, G24C25 and T26G27 step also favor (g-,g+,g+,trans) (Fig. 2). Additionally, A5G6&C25T26 favor (g-,g-,g+,t) for (ε,ζ,α,γ), while C7A8 takes up (g-,g-,g-,g+). This eventually reflects in the helical twist angle at the central A5G6, G6C7, C7A8, A8G9 & C10A11 adopting lower helical twist (S1 Fig). Notably, (g-,g-,g+,t) conformation for A5G6 and for its complementary C25T26 step results in a negative twist of -10°. Thus, there is an evident increase in Z-DNA stretch at & around the mismatch site during the end of the simulation. It is noteworthy that beyond 181ns, N1(A8)…N6(A23) & N1(A23)…N6(A8) hydrogen bonds are equally favorable, while the canonical C7…G24 and G9…C22 hydrogen bonds flanking the A8…A23 mismatch remain unaffected throughout the simulation (S2 Fig).
Concomitant to above, major and minor groove widths also undergo changes. Unwinding of the helix leads to the expansion of minor groove width at the mismatch site to ~20.1(0.4)Å flanked by comparatively narrower groove widths of 12.7Å&14.9Å on either side at the end of the 300ns simulation.
Akin to A8…A23 mismatch with anti…anti glycosyl starting conformation, the starting model with +syn…anti glycosyl conformation also undergoes significant conformational changes. This can be seen from RMSD (Fig. 3A) that increases to 2(0.3)Å till ~9.1ns and subsequently to 3.1(0.5)Å during 9.1-36ns. It stays ~5.5(0.8)Å beyond 36ns.
Detailed analysis indicates that the increase in RMSD to 5.5Å is due to the conformational preference for local Z-DNA structure at & around the A8…A23 mismatch site to accommodate the mismatch. In fact, an increase in Z-DNA stretch around the mismatch site is seen (Fig. 3B) during the 300ns simulation. One of the marked changes associated with Z-DNA conformational preference is A8 adopting high-anti/-syn (287 (17°)) glycosyl conformation beyond 36ns (Fig. 3C). Conformational changes at A8 beyond 36ns enforce -syn glycosyl conformation for neighboring G9 (248 (26°) to 321(32°)) and G24 (248(25°) to 324 (15°)) (S3A Fig). Other notable changes that happen during the early part of the simulation (~9ns) in seeding Z-DNA conformation are, the preference for -syn glycosyl conformation by G21 (from 249(24°) to 296(25°)) (hydrogen bonded with C10) and A11 (from 257(23°) to 302(32°)) (base paired with T20) that are located in the neighborhood of A8…A23 mismatch site (S3A Fig). Irrespective of the above conformational changes, chi at A23 stays close to the initial +syn (S3A Fig) conformation throughout the simulation. It is noteworthy that a total loss of hydrogen bonds at N1(A8)…N6(A23) & N6(A8)…N7(A23) that happenes due to base extrusion during 30-40ns facilitates B-Z transition (S3B-E Fig).
Yet another interesting observation is the preference for stacked conformation between the mismatched A8&A23 bases (Fig. 4) that is facilitated by the Z-DNA conformation. As a result, there is a total loss of N1(A8)…N6(A23) hydrogen bond as well as N6(A8)…N7(A23) hydrogen bond between the mismatched bases beyond 150ns (S3B Fig). It happens in such a way that ~133ns the hydrogen bond becomes longish, followed by A8 and A23 moving out-of-plane with each other. Subsequently, A8 stacks on top of A23 like an intercalator and stays till the end of the simulation (Fig. 4). During the aforementioned conformational changes, the canonical C7…G24 and G9…C22 that is located above and below the A8…A23 mismatch respectively remain intact (S3B Fig).
Excitingly, aforementioned transformations are accompanied by prevalence for Z-DNA backbone conformation. For instance, when both A8 & A23 are in plane during the first 100ns simulation, -syn conformation for chi at A8, G9, G21, G24 & A11 is concomitant with (ε,ζ,α,γ) adopting (g-,g+,g+,trans) at G9C10 (S4A Fig), A11G12 (S4B Fig) & G24C25 (S4F Fig) steps, while T20G21 (S4C Fig), C22A23 (S4D Fig) & A23G24 (S4E Fig) steps taking up (g-,g-,g+,trans). Consequent to the above sugar-phosphate conformational changes, helical twists at C7A8 (8.5 (12))°, A8G9 (7.4 (8))° and C10A11 (6.1 (9))° (Fig. 3D) steps adopt low twist values in between high twist values (S5B Fig) resulting in a Z-DNA sandwich structure as before.
Stacked conformation of A8&A23 that is formed after 150ns leads to large fluctuation in the helical twist of C7A8 & A8G9 steps, wherein, the C1’…C1’ vector of A8…A23 is nearly perpendicular to the C1’…C1’ vectors of the neighboring canonical base pairs. This is associated with large fluctuation in conformational angle alpha at C22A23 step (S6 Fig). Additionally, (ε,ζ,α,γ) for G9C10, A11G12, G21C22, G24C25, C10A11 & T20G21 steps also favor Z-DNA conformations like (g-,g+,g+,trans) & (g-,g-,g-,trans) (S6 Fig). The general tendency in helical twist associated with the above conformational preference is that A8G9 (8 (11))°, G9C10 (25(5))° and C10A11 (18 (8))° prefer a low twist during the 150-300ns (S5C,D Fig).
It is clear from above that like in the previous situation (Fig. 1), formation of local Z-DNA conformation is propagated to the neighboring bases (from C7 to G12) of A8…A23 mismatch. This eventually reflects in at least 3 steps located in the middle of the duplex taking up lower helical twists (S5B-D Fig). Essentially, this leads to unwinding of the double helix (S5B-D Fig & S7 Fig and S4 Movie), a typical characteristic of B-Z junction (PDB ID: 1FV7). Such unwinding is accompanied by expansion in the major (maximum of 28 Å) and minor (maximum of 20 Å) groove widths. However, at the mismatch site, the minor groove width shrinks to 11.5 Å during the 100ns simulation. It further shrinks to 8 Å, followed by the stacked conformation of A8&A23.
Thus, formation of a local Z-DNA conformation accompanied by unwinding of the helix is evident even with a single A…A mismatch irrespective of the starting conformation.
To investigate the effect of periodic occurrence of A…A mismatch as in the real situation of Huntington’s disorder and several spinocerebellar ataxias, 300ns MD simulation has been carried out for d(CAG)6.d(CAG)6 sequence (Fig. 5A). As before, 2 starting models each with +syn…anti and anti…anti glycosyl conformations are considered for all the six A…A mismatches.
RMSD (~3.3 (0.9) Å) calculated over 300ns MD simulation of (CTG)6.(CAG)6 duplex (Fig. 7A) indicates that the molecule undergoes minimal conformational rearrangement from the starting B-form geometry (Fig. 7B). Strikingly, the overall structure doesn’t show any tendency to adopt Z-form, as can be visualized from Fig. 7C. Instead, it retains the compact B-form geometry.
The helical twist always stays positive (Fig. 7D-F), adopting a trend of high helical twists at GC (38.8(4°)) step compared to CT (23(4°)) and TG (28(5°)) steps over the last 10ns (S19 Fig). Unlike before, both A’s and G’s don’t favor ±syn conformation and have the tendency to retain anti glycosyl conformation (180–270°) (~%70) (Fig. 7G,H). Significant conformational changes in the backbone are also not observed as (ε,ζ) fall profoundly in BI (t,g-) or BII (g-,t) conformation (S20 Fig). Similarly, (α&γ) favor either (g-,g+) or (g-,t). All these together pinpoint B-DNA conformational preference for (CTG)6.(CAG)6 duplex.
Structural information about the distortions caused by A…A mismatch in a DNA duplex is not yet well defined at the atomistic level. The only structure that has been reported so far with A…A mismatch in a DNA is the complex of a DNA duplex and Muts, an E. coli mismatch repair protein, with a significant bending at the mismatch site (PDB ID: 2WTU). NMR and thermodynamic studies of A…A mismatch containing DNA duplex offer controversial results. While some of them suggest that A…A mismatch destabilizes[18,19,20,21] the DNA duplex significantly, the others do not[22]. Physicochemical studies indicate that A…A mismatch in a GAC repeat adopt several distinct conformations in solution including Z-DNA[23,24]. In fact, it has been suggested that A…A mismatch in GAC repeat promotes Z-DNA formation [23].
Understanding the structural role of A…A mismatch is very important in the context of Huntington’s disorder and several spinocerebellar ataxias due to the formation of hairpin structures consisting of noncanonical A…A base-pairs. MD simulations carried out in this context reveal a very exquisite observation that A…A mismatch in a CAG repeat induces change in the helicity from right-handed B-DNA to left-handed Z-DNA. Even a single A…A mismatch tends to form a local Z-DNA structure leading to Z-DNA sandwich (Figs. 1,3). When the A…A mismatches occur in a regular interval, it leads to local left-handed Z-DNA formation at the mismatch site followed by a right-handed DNA at the canonical WC pair site leading to periodic B-Z junctions (Figs. 5,6). Formation of Z-DNA structure is evident from the preference for (±)syn…high-anti/(-)syn glycosyl conformation by A…A mismatch and backbone conformational angles (ε,ζ,α,γ) favoring (g-,g+,g+,t), (g-,g-,g+,t) and (g-,g-,g-,g+) at & around the mismatch site. Additionally, G’s prefer -syn conformation. This results in a low helical twist at the CA and AG steps in the midst of high twist at the GC step, a characteristic of B-Z junction (PDB ID 1FV7).
An intriguing observation is that a single hydrogen bonded noncanonical A…A mismatch induces Z-DNA conformation through ‘zipper mechanism’ [25] assisted by base extrusion, base and/or backbone flipping (Figs. 1,6 and S2,S3&S21 Figs). While the sugar-phosphate backbone flipping is prominent in anti…anti glycosyl conformation, base extrusion and sugar-phosphate & base flipping are favored by +syn…anti conformation to transit from B-to-Z form DNA. Yet another interesting fact is that the above-mentioned Z-DNA formation is a noninstantaneous event, rather it propagates in a stepwise manner (Figs. 5I, 6 (Bottom) and S7 Fig). Though the noncanonical A…A mismatch impels Z-DNA conformation, the canonical base pairs have the prevalence for B-form geometry resulting in B-Z junction. Formation of such B-Z junction can be readily visualized by unwinding of the double helix irrespective of the starting glycosyl conformation (S22 Fig).
A…A mismatch adopts 2 different ‘base flipping’ pathways to undergo transition from +syn…anti to -syn…-syn (Fig. 6) accompanied by sugar phosphate rearrangements. One mode of transition is +syn moving to -syn through cis conformation (via counter-clockwise rotation around glycosidic bond), while the other is via trans conformation (through clockwise rotation around the glycosidic bond). In general, DNA with +syn…anti conformation takes longer time to undergo the B-Z transition, compared to anti…anti conformation.
Reported structural changes provoked by A…A mismatch can be attributed to the higher degree of nonisomorphism between A…A mismatch and the canonical base pairs. This can be visualized from the larger value of residual twist and radial difference [17,26], the measures of base pair nonisomorphism (S23 Fig). In fact, both residual twist (16º) and radial difference (1.6Å) are quite prominent for A…A mismatch with anti…anti glycosyl conformation, but, only residual twist (16º) is significant and the radial difference is negligible (0.2Å) in the case of +syn…anti glycosyl conformation. This may be the reason for the reluctance of A…A mismatch to retain anti…anti conformation and the transition to -syn…-syn being quite fast compared to +syn…anti starting conformation.
In general, the transition from B-to-Z involves complex mechanisms and exhibits a high-energy barrier to transit to Z-DNA conformation. In fact, several mechanisms have been proposed for B-to-Z transition[27] and a recent adaptively biased and steered MD study demonstrates the coexistence of zipper and stretch-collapse mechanisms engaged in transition[28]. However, the mechanistic effect that arises from the intrinsic extreme nonisosterecity of A…A mismatch with the canonical base pairs immediately dictates B-to-Z transition without the influence of any external factors. As the A…A mismatch is single hydrogen bonded, it exhibits enormous flexibility for base extrusion and flipping, facilitating the formation of Z-DNA through zipper mechanism. Interestingly, such a conformational change is not seen in the crystal structure of RNA duplex with A…A mismatch[13]. Thus, it is clear that the effect of A…A nonisomorphism is pronounced in the DNA and not in the RNA.
Several experimental studies have revealed that d(GA) [29], d(GAA) [30], d(GGA) [31] and d(GAC) [23,24] repeats that contain A…A mismatches are prone to adopt parallel homoduplex. Such preponderance for parallel duplex by these sequences may be due to left-handed Z-DNA provoking nature of A…A mismatch, which is a high-energy conformation. Hitherto, this aspect is not realized as there is no DNA duplex structure with A…A mismatch available with any sequence context. Earlier low-resolution 1D NMR studies on DNA duplexes comprising of A…A mismatch[18,19,20,21,22] offer only minimal information with some of them indicating notable destabilization induced at A…A mismatch site[18,19,20,21]. Strikingly, it has been shown by circular dichroism study that CAG repeat spectra resembles GA homoduplex but not CCG and CTG[32]. Propensity of A…A mismatch containing DNA to adopt a parallel DNA duplex is also reported[21]. However, the possibility of CAG repeat expansion to favor parallel duplex can be ruled out as it forms hairpin structure[7,8], which eventually leads to antiparallel orientation for the two strands of the DNA hairpin stem. Thus, DNA hairpin stems containing CAG repeat may adopt local Z-DNA conformation at A…A mismatch site leading to ‘B-Z junction’ as revealed by the current investigation. Our result gains support from earlier surface probing using anti-DNA antibody that demonstrated the presence of Z-DNA structure in CAG & CTG repeat expansions [33]. It can also be recalled that formation of hairpin structure with such Z-DNA stem has been observed earlier in a different sequence context [34,35,36]. Thus, we envisage that such noncanonical ‘B-Z junction’ in CAG repeat expansion may be one of the factors responsible for the newly emerging mechanism of ‘DNA toxicity’ observed in CAG repeat expansion[37].
Thus, for the first time it has been shown here that the A…A mismatch in a DNA duplex with CAG repeat is an inducer of local Z-form conformation through ‘zipper mechanism’ that stems from backbone flipping and base pair extrusion & flipping leading to B-Z junction. Such B-Z junction instilled by A…A mismatch results from the mechanistic effect intrinsic to the nonisoterecity of A…A mismatch with the flanking canonical base pairs. With emergence of evidence on ‘DNA toxicity’ of CAG overexpansion and its role in triggering cell death [9,10], one can envision that occurrence of B-Z junction is the molecular basis for Huntington’s disorder and several spinocerebellar ataxias. This further leads to the speculation that B-Z junction binding protein may have a role in the diseased states. Reported results would further be useful in understanding DNA repair mechanisms involving A…A mismatch, thus adding a new dimension to the role of A…A nonisosterecity on DNA structure.
Initially, (CTG.CAG)5 & (CTG.CAG)6 DNA duplexes containing canonical C…G and G…C base-pairs with ideal B-form geometry are generated using 3DNA[38]. These models are subsequently manipulated to introduce a non-canonical A…A mismatch in the middle of canonical base pairs to generate a 15mer DNA duplex (Fig. 1A) using Pymol (www.pymol.org, Schrödinger, LLC) molecular modeling software. A…A mismatch is modeled so as to form N6(A)…N1(A) hydrogen bond. For the generation of model with periodic A…A mismatches (18mer, Fig. 3A), ‘T’s in the (CTG.CAG)6 duplex are replaced manually with A’s as mentioned above. To establish base-sugar connectivity and to restraint the sugar-phosphate backbone conformation, the models are refined using X-PLOR [39] by constrained-restrained molecular geometry optimization and van der Waals energy minimization. The second conformation for the A…A mismatch, viz., N6(A)…N1(A) hydrogen bond with +syn…anti glycosyl conformation is generated using X-PLOR by applying appropriate restraints. Subsequently, the models are subjected to a total of 1.5μs molecular dynamics simulations (MD) using Sander module of AMBER 12 package [40].
X-PLOR generated duplex models with A…A mismatches and the 3DNA generated canonical (CTG.CAG)6 duplex are solvated with TIP3P water molecules and net-neutralized with Na+ counter ions. Following the protocols described in our earlier papers [17,41,42], equilibration and production runs are pursued for 300ns for the sequences given in Table 1. Simulations are performed under isobaric and isothermal conditions with SHAKE (tolerance = 0.0005 Å) on the hydrogens [43], a 2fs integration time and a cut-off distance of 9 Å for Lennard-Jones interaction. FF99SB forcefield is used and the simulation is carried out at neutral pH. Trajectories are analyzed using Ptraj module of AMBER 12.0. Helical parameters and conformation angles are extracted from the output of 3DNA using in-house programs. Due to the presence of noncanonical base pairs, helical twist angles are calculated with respect to C1’…C1’ vector [17,41,42]. Pymol is used for visualization and MATLAB software (The MathWorks Inc., Natick, Massachusetts, United States) is used for plotting the graphs.
|
10.1371/journal.pbio.2003993 | Semantic representation in the white matter pathway | Object conceptual processing has been localized to distributed cortical regions that represent specific attributes. A challenging question is how object semantic space is formed. We tested a novel framework of representing semantic space in the pattern of white matter (WM) connections by extending the representational similarity analysis (RSA) to structural lesion pattern and behavioral data in 80 brain-damaged patients. For each WM connection, a neural representational dissimilarity matrix (RDM) was computed by first building machine-learning models with the voxel-wise WM lesion patterns as features to predict naming performance of a particular item and then computing the correlation between the predicted naming score and the actual naming score of another item in the testing patients. This correlation was used to build the neural RDM based on the assumption that if the connection pattern contains certain aspects of information shared by the naming processes of these two items, models trained with one item should also predict naming accuracy of the other. Correlating the neural RDM with various cognitive RDMs revealed that neural patterns in several WM connections that connect left occipital/middle temporal regions and anterior temporal regions associated with the object semantic space. Such associations were not attributable to modality-specific attributes (shape, manipulation, color, and motion), to peripheral picture-naming processes (picture visual similarity, phonological similarity), to broad semantic categories, or to the properties of the cortical regions that they connected, which tended to represent multiple modality-specific attributes. That is, the semantic space could be represented through WM connection patterns across cortical regions representing modality-specific attributes.
| One of the most challenging questions in cognitive neuroscience is how semantic knowledge, for example, that “scissors” and “knives” are related in meaning, can emerge from primary sensory dimensions such as visual forms. It is often assumed that in the human brain, semantics are stored in regions of the brain cortex, where distinct types of modality-specific information are transferred to and bind together. We tested an alternative hypothesis—“representation by connection”—in which higher-order semantic information could be coded by means of connection patterns between cortical regions. Combining data from behavior and brain imaging of 80 patients with brain lesions, we applied machine learning to construct the mapping models between the lesion patterns on axonal tracts (white matter) and item-specific object-naming performances. We found that specific white matter lesions produced deficits in object naming associated with the object’s semantic space, but not relevant to its primary dimension. The naming performances of semantically related objects were better predicted from white matter lesion-pattern models. That is, the higher-order semantic space could be coded in patterns of brain connections by linking cortical areas that do not necessarily contain such information.
| One of the most challenging questions in cognitive neuroscience is how abstract knowledge emerges from more basic dimensions of information, such as visual shapes and patterns of motor action. How do we proceed from the visual shape of a pair of scissors to the knowledge that they can be used to cut things and that they are semantically related to an axe, which looks different and is manipulated differently from scissors? Research on the neural basis of semantic memory—the storage of general knowledge about the world—has revealed widely distributed brain regions supporting modality-specific attributes of objects, such as shape, color, and motion (e.g., [1,2]; see review in [3]). Nonetheless, such attribute-specific knowledge and its simple pairings are not adequate to explain the actual semantic space of objects that have quite different sensory/motor attributes but that may nonetheless be considered to be semantically similar (e.g., [4–7]). To achieve such a semantic space, various steps of binding and abstraction are assumed to occur at specific gray matter (GM) regions [6,8–11].
Although past research on semantic representation has focused on the roles of cortical regions, specific white matter (WM) tracts have been found to be necessary for semantic processing, including the left inferior fronto-occipital fasciculus (IFOF), the left uncinate fasciculus (UF), and the left anterior thalamic radiation. Damage to these tracts is associated with semantic deficits in patients [12–17]. WM is classically assumed to relay information [18–20]. In accord with this general notion, these WM tracts that are necessary for semantic processing are assumed to relay distributed information to particular GM regions (e.g., the anterior temporal lobe or angular gyrus) for binding, where concepts are represented and the “deep structures” of semantic space are formed [6,7,21]. The nature of the potential information carried by WM has never been discussed or examined.
Herein, we present results for a new notion that the WM connections, being natural binding structures, provide an alternative basis to achieve semantic representation. Distributed GM regions that represent different attribute dimensions (e.g., shape, color, manner of interaction) of the same object are connected by WM. The WM linking pattern itself would then contain multiple dimensions of information in these GM regions and, importantly, additional information about the manner of mapping among various attributes. The incorporation of these elements has been argued to be necessary for the “higher-order” semantic similarity relationships, which are not explained by attribute-specific spaces, to emerge (e.g., [7, 22]).
To investigate the information coded in WM connections, we extended representational similarity analysis (RSA) [23], a highly productive method that tests the nature of representation in functional magnetic resonance imaging (fMRI) studies of cortical regions [24–26], to lesion data and WM connections. RSA examines the relationship between the representational dissimilarity matrix (RDM) derived from neural patterns and RDMs based on various types of stimulus information as a measure of information representation. The conventional neural RDM is measured by the dissimilarity of brain activity patterns induced by stimulus conditions. Here, we compute the neural RDMs with a machine-learning model using the voxel-wise lesion patterns as features to predict behavioral performance in patients with brain damage (see Fig 1). The performance in picture naming of 100 object items and the structural MRI data of 80 patients were collected. For each WM connection, a training model was built for each item (e.g., scissors) using the support vector machine (SVM) classifier with patients’ voxel-wise lesion patterns as predictive features and the naming performances of that item as labels (0, incorrect; 1, correct). The correlation between the predicted score using the classifier from that item and the actual scores of another item (e.g., axe) was taken as the neural similarity basis of these two items, based on the assumption that if this connection pattern contains certain aspects of information shared by the naming process of these two items, models trained with one item (useful features relevant for such information) should also predict naming accuracy of the other item. Once the neural RDMs are obtained from various WM connections or GM regions using this method, they can be correlated with behavioral RDMs of various object property dimensions, including the semantic RDM and four modality-specific attribute RDMs (shape, manipulation, color, and motion). Neural RDMs that are correlated with the semantic RDM even after controlling for the attribute RDMs are considered to contain “higher-order” semantic information.
Behavioral RDMs for the semantic, shape, manipulation, color, and motion features of 100 objects (20 animals, 20 fruits and vegetables, 20 tools, 20 non-tool small objects, and 20 large non-manipulable objects) were generated using a multi-arrangement method [29]. In this task, 20 college students were instructed to arrange the items by a particular dimension of interest on a computer screen, and the distance among items was derived, resulting in an RDM (see Fig 2A).
The semantic RDM was visually clustered into three domains: animals, fruits and vegetables, and man-made objects (tools, small non-tool objects, large non-manipulable objects; see Fig 2B & 2C). Visualization of the semantic RDM using multidimensional scaling (Fig 2C) further revealed that within each category, words with closer semantics tended to share similar function (e.g., scissors and knife), share certain distinct features, or belong to finer subordinate categories (e.g., peanut and potato). The semantic RDM and the four modality-specific attribute RDMs were intercorrelated to various degrees (Fig 2D; semantic with shape: r = 0.35; with manipulation: r = 0.47; with color: r = 0.23; with motion: r = 0.27; p < 10−9).
Neural RDMs were generated for each of the 688 WM connections (S1A Fig) that were identified through deterministic tractography across 90 automated anatomical labeling (AAL) regions based on the diffusion tensor imaging (DTI) data of 48 healthy controls [27]. To generate the neural RDM for each WM connection, we performed lesion-naming model decoding using voxel-wise lesion patterns and item-level naming responses. For 80 patients with brain damage, lesion patterns in each WM connection (with each voxel in the WM connections labeled as “lesion” or “intact”) for each patient were obtained by overlapping the manually traced lesion mask (converted to the MNI) space) with the WM mask (see Fig 1). A total of 680 out of 688 WM connections with adequate lesion coverage (see Materials and methods; see also S1E Fig for the lesion distribution map) were included in the following analyses. The patients’ naming performances for each of the 100 pictures were collected (performance distribution in S1B Fig).
WM neural RDMs were generated using item-based lesion-naming prediction models. For 197 connections, the lesion-naming models had successful within-item prediction averaged across all items (Bonferroni p < 0.05; diagonal in Fig 1D). That is, they yielded successful naming prediction models and were the connections that we considered in the following analyses. Of these connections, 185 were located in the left hemisphere and 12 in the right hemisphere (S1C Fig). For each of these WM connections, we computed the correspondence between the predicted scores using SVM classifiers built using the training patients’ lesion patterns and the naming scores of one item and the actual naming score of another item in the testing samples across testing iterations. This between-item correlation was taken as the similarity value for this item pair in the neural RDM, based on the assumption that if this connection pattern contains certain aspects of information shared by the naming process of these two items being captured by the SVM model, models trained with one item should also predict naming accuracy of the other item. Worth clarifying is that this procedure does not depend fully on the correlation between the actual naming accuracies across item pairs but also to what degree the potentially shared underlying properties for their naming process are supported by each WM connection (as captured by the SVM models). For example, for connections supporting phonological processing, the SVM models may pick up phonological properties and result in higher correlation between phonologically related pairs; those supporting semantic processing may pick up semantic properties and result in correlation between semantically related pairs. The resulting 100 × 100 (-item) lesion-naming prediction similarity matrix was transformed to be the neural RDM of this connection (1-prediction similarity, Fig 1D).
Using RSA, the correlations between the WM neural RDMs and the semantic RDM were assessed. Significantly positive correlations were obtained in 60 WM connections (r = 0.03–0.11, false discovery rate [FDR] q < 0.05; see S1D Fig). These WM connections connected widely distributed regions across the left hemisphere, and approximately half (31/60) of the connections had at least one of the connected nodes located in the temporal lobe. The most densely connected regions (degree z-score > 1) were the middle temporal gyrus (MTG), superior temporal gyrus (STG), orbital part of middle frontal gyrus, inferior parietal lobule (IPL), and precentral gyrus.
What about semantic effects that could not be explained by modality-specific attributes, peripheral factors, or broad semantic categorical effects? We controlled for the effects of all four modality-specific attributes, two peripheral variables (the early visual and phonological) and semantic category matrix (labeling within-category pairs 1 and between-category pair 0) using partial correlation. The semantic effect was consistently significant in eight WM connections (r = 0.03–0.07, FDR q < 0.05; Fig 3A–3C). Table 1 presents the detailed statistical results before and after, including these variables as covariates. These eight connections were considered to represent (relatively) higher-order semantic space. Five of them were located in the left ventral visual pathway and connected occipital regions (middle occipital gyrus, calcarine sulcus, and lingual gyrus) and temporal regions (STG, MTG, superior anterior temporal lobe [ATL], and middle ATL). The three remaining WM connections were located in the right hemisphere, connecting the postcentral gyrus with the thalamus, lingual gyrus, and parahippocampal gyrus. These reconstructed connections are shown in Fig 3B and S2 Fig.
To examine the degree to which the semantic effects we observed on these WM connections reflect effects of broad semantic category, we also checked the RSA effect of the category matrix (correlating the neural RDM and the category RDM) and found that none of these connections had significant effects of the semantic category (p > 0.05, Table 1).
To consolidate the main results above, we further performed validation analyses to test the following concerns: (1) The WM mask we adopted was constructed using DTI data acquired on a scanner with a low magnetic field (1.5 T) and 32 directions. Was the WM connection construction accurate and unaffected by crossing-fiber problems? (2) To maximize power, we included patients with multiple etiologies (84% stroke and 16% traumatic brain injury [TBI]) and lesion distributions (37.5% lesion in the left hemisphere only, 43.8% lesion in bilateral hemispheres, and 18.8% in the right hemisphere only). Were the results systematically affected by disease type or hemispheric differences?
What types of representations are linked by the WM connections that represent the semantic space? Do the WM connections simply relay semantic information that has already been encoded in the GM nodes, or do they contain information that cannot be accounted for by representation in the GM nodes? We tested the representational contents of the seven GM nodes that were connected by the five higher-order semantic WM connections whose effects remained robust in the validation tests (see Table 1). Four GM regions had successful within-item naming prediction and were considered in the RSA analysis: superior ATL, middle ATL, MTG, and STG. The neural RDM for each GM node was constructed using the same method as with the neural RDMs of the WM connections.
We found that the higher-order semantic representation in the five semantic WM connections cannot be simply explained by GM information (Fig 3D; S1 Table): when correlating the GM neural RDMs with the semantic RDM (controlling for peripheral and categorical matrices), only the superior ATL reached significance (r = 0.04, FDR q < 0.05). However, this effect could be explained by modality-specific attribute representations. After controlling for the four modality-specific attribute matrices, none of the four GM nodes significantly correlated with the semantic RDM at either the conventional threshold (FDR q < 0.05) or a less stringent threshold (uncorrected p < 0.05, see S1 Table). Additionally, when testing the higher-order semantic representation in the five WM connections by further adding the neural RDMs of the two GM nodes being connected as additional confounding variables, the results remained unchanged (see Table 1). We further constrained our WM connection mask with a WM mask constructed by T1 segmentation (conducted using SPM8 in MNI T1 template, default parameters) to offer a clear WM boundary, i.e., containing only WM voxels. We then recomputed the higher-order semantic RSA in these WM connections using only the voxels within the WM mask and found that the effects in all five WM connections remained significant (FDR q < 0.05, r = 0.03–0.07, SD = 0.01).
If not semantic, do these GM nodes code modality-specific attributes? We correlated the neural RDM of each GM node with each of the four modality-specific attribute RDMs (shape, manipulation, color, and motion; Fig 3D & S1 Table; the three control matrices—low-level visual, phonological, category—were controlled for). The superior ATL, MTG, and STG were significantly correlated with the shape and manipulation RDMs (shape: r = 0.04–0.08, manipulation: r = 0.12–0.16, FDR q < 0.05). The middle ATL was significantly correlated with the shape and color RDMs (shape: r = 0.04, color: r = 0.06, FDR q < 0.05).
Finally, we conducted a whole-brain analysis across all 90 AAL GM nodes. In addition to superior ATL, the neural RDMs of the left IPL, precentral gyrus, and postcentral gyrus were significantly correlated with the semantic RDM (r = 0.04–0.05, FDR q < 0.05), but none of these or any other GM regions retained significance after controlling for the four modality-specific attribute matrices (FDR q < 0.05).
To test the potential WM basis of semantic representation, we developed a structural-property-pattern-based RSA approach by applying machine learning to lesion and behavioral data in patients to derive item-based neural RDMs for WM connections. We found that a set of WM connections connecting occipital/middle temporal regions and anterior temporal regions represented a semantic space that was not explained by broad semantic categories or the effects of modality-specific attributes and, hence, was addressed as higher-order semantic representation. Such semantic effects were not fully explained by the properties of the GM nodes that were connected. Although the neural RDM of a connecting node—the superior ATL—correlated with the semantic RDM, such effect diminished after controlling for modality-specific attributes. Instead, these GM nodes tended to represent modality-specific attributes, including shape and manipulation in the superior ATL, MTG, and STG and shape and color in the middle ATL.
First, it should be noted that we inferred semantic effects to be higher-order when they were not explained by linear combinations of the classical modality-specific attributes for objects. The potential effects of some untested modalities or certain nonlinear combinations across various modalities could not be fully excluded. Also, subjectively judged semantic distance might be a rather composite measure that is driven by multiple semantic dimensions, which may have different neural bases (e.g., [30]). Under the current (conventional) operation, these WM connections that represent higher-order semantics tend to lie in several major pathways that have been associated with semantic processing using univariate lesion-behavior correlation or intraoperative stimulation [12,16,21,27,31]. These connections partly belong to IFOF, and the inferior longitudinal fasciculus (ILF) (the overlapped voxels with the Johns Hopkins University WM template: IFOF [32%], ILF [71%], and minimally on the minor forceps [6%] and superior longitudinal fasciculus [8%]). Lesion or atrophy in IFOF is associated with semantic deficit severity in patients with stroke and in patients with semantic dementia [12,27,32]. A similar result was also found with ILF in semantic dementia [16,33]. Additionally, direct intraoperative stimulation of IFOF induces semantic errors [34,35]. Our current findings based on multivariate RSA demonstrate that the organization of specific connections among these large WM tract bundles represent the fine-grained semantic space. Items closer in semantic space are represented by more similar WM patterns in these specific connections. Note that it is well known that patients’ specific naming errors may vary from session to session [36]. The WM lesion pattern observed here is likely associated with some aspects of semantic space rather than with specific items. The damage of such specific aspects of semantic space would result in noisy/impaired representation for a range of items sharing that space, resulting in potentially different outputs at different time points. Such semantic space was nonetheless much finer than broad semantic categories, however, as the RSA results were robust after controlling for the categorical matrix. It is also well known that patients may make different types of errors, such as phonological and semantic paraphasias, which may be originated from different cognitive stages. Our approach here pulled all types of naming errors together, and the RSA results of correlating the neural RDM with different RDMs (semantic versus phonological/visual) presumably reflect the neural basis of different error types, which should be directly examined in future research.
What is the relationship between the WM representations and the nature of the GM regions that they connect? First, we indeed observed that one of the seven linked GM regions was related to semantic space—superior ATL. The finding that lesion-pattern-behavior (neural) RDM in the superior ATL correlated with semantic space before regressing out the effects of modality-specific attributes converges nicely with the accumulated evidence about the cortical representation of semantics from fMRI and neuropsychological studies. ATL is the region with the strongest atrophy in patients with semantic dementia, which is marked by semantic deficits [6,7,31,37,38] and is sensitive to multiple modalities of object attributes [39,40]. Unlike the WM connections related to higher-order semantic space, the semantic effect in the superior ATL could be explained by the effects of modality-specific attributes. Worth noting is that ventral ATL was not scrutinized because it was not a node in the AAL parcellation we used but was included in the fusiform and inferior temporal nodes. What should be highlighted, however, is that the positive effects of higher-order semantic representation in the WM connections are significant and are not simply inheriting the properties of the connected GM nodes. Several higher-order semantic WM connections observed here connected ATL with other regions, inviting further questions about whether it is the integrity of ATL or of the ATL-related WM connections that make stronger contributions to the semantic deficits in semantic dementia patients.
While our results certainly do not argue against the possibility that there are specific GM regions supporting semantic representation, we found that the GM nodes being connected by the WM connections obtained here tended to represent multiple modality-specific object properties. Of the four GM regions we could test, the MTG, STG, and superior ATL represented shape and manipulation properties, and the middle ATL represented shape and color properties. These results converge nicely with the fMRI literature studying the sensitivity of these regions for object attributes. For instance, the effects of various attributes were recently tested using parametric modulation analyses [2], which found that the posterior MTG was sensitive to both shape and manipulation knowledge. Coutanche and Thompson-Schill [39] found that the ATL codes the integration of color and shape, and Peelen and Caramazza [40] found that the ATL codes both manipulation and location. The STG was sensitive to motion properties in Fernandino et al. [2] but not in our study, perhaps due to different parcellation scales regarding the finer structure within this region. Note that many studies about the attribute-specific property representations have revealed results in sensory and motor cortices (e.g., shape in the lateral occipital/temporal cortex: [26,41]; color in the ventromedial occipital cortex such as lingual gyrus: [42–44]). However, these regions could not be tested in our data given their chance-level lesion-naming prediction performance, which could either be due to low lesion distribution in these regions (see S1E Fig for lesion distributions) or because the specific dimensions they represent are unnecessary for object picture-naming behavior. It may also be the case that higher-order semantic space is formed by binding multiple, rather than single, pairs of attributes. Consistent with this speculation, it has been shown that computation simulation models with a convergent architecture, in which intermediate units code multiple types of dimension pairings, were better at capturing the “deep” structure of conceptual space and promoting generalizations across semantically related items that were not apparently similar along single dimensions [22].
What is the mechanism of coding higher-order semantic information in WM that connects multiple modality-specific attributes? One potential mechanism could be through synchronized firing of specific sensory and motor patterns for objects. Consider when people use a pair of scissors: the neurons that represent the attributes across various modalities—e.g., shape, haptics, ways of grasping and manipulating it, seeing the consequence of using it (things being cut)—fire together. Such functional co-activation across a wide range of attributes occurs often when we see or use scissors, which enhances the structural connection between neurons within and across dimensions of the same object. WM provides a basis for such synchronization between distant cortical regions [45]. These synchronizations also lead to the building and tuning of WM connections, because neuronal activity traveling through axons can affect the properties of myelin sheaths in the active circuit; for example, electrical activity in the axon induces myelination [46,47]. This interactive process results in the WM basis of a multidimensional representation of “scissors,” which is closer in the higher-order semantic space to concepts such as “axe” or “paper.” The formation and modulation of the WM microstructure underlying these representations can be affected by our experiences, which is the basis of acquiring new concepts and of the coloring of existing concepts. Ample evidence describes how WM is affected by experience. Early-life experiential deprivation in animals and humans leads to decreased myelin sheath thickness and WM volume [48,49], whereas these parameters increase when the organism is placed in a rich experiential environment [50]. Reading training [51] and music practice [52,53] during childhood lead to increased fractional anisotropy in WM. The acquisition of motor skills changes the WM microstructure [54,55]. The exact relationship between WM microstructure and the functional coupling between cortical regions for various representational dimensions warrants further studies.
A final methodological note is that the approach we developed here—building neural RDMs using machine learning with structural lesion data and condition-specific performances—could be easily adapted to other cognitive issues and all kinds of brain structural integrity measurements, including DTI indices (e.g., fractional anisotropy, mean diffusivity) or voxel-based morphometry measures for both patient and healthy populations. For the current study, we chose to focus on manually traced lesion on the T1 image (with reference to T2) because it captures the structural damage in our specific patient group (mostly stroke) in a most straightforward fashion. RSA, an approach that connects major branches of systems neuroscience—brain-activity measurement, behavioral measurement, and computational modeling [23]—could now be extended to an additional branch, i.e., brain structural measurement.
In conclusion, using a structural-property-pattern-based RSA approach, we found that the WM structures mainly connecting occipital/middle temporal regions and anterior temporal regions represent fine-grained higher-order semantic information. Such semantic relatedness effects were not attributable to modality-specific attributes (shape, manipulation, color, and motion) or to the representation contents of the cortical regions that they connected and were above and beyond the broad categorical distinctions. By connecting multiple modality-specific attributes, higher-order semantic space can be formed through patterns of these connections.
Eighty patients with brain damage participated in the present study. The patient group (60 males, 20 females) was recruited from the China Rehabilitation Research Center with at least 1 month post-onset (mean = 6.09; SD = 11.69; range: 1–86 months) and premorbidly right-handed. The majority suffered from stroke (n = 67) and others suffered from TBI (n = 13). The patients’ mean age was 45 years (SD = 13; range: 19–76 years) and mean years of formal education was 13 (SD = 3; range: 2–19). Twenty additional college students (10 males; mean age = 22.9, SD = 2.45, range = 19–27) participated in the multi-arrangement experiment for the behavioral RDMs. This study was approved by the Institutional Review Board of the State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University (IORG0004944), adhering to the Declaration of Helsinki for research involving human subjects. All participants gave informed written consent.
Each subject was scanned using a 1.5T GE SIGNA EXCITE scanner with an 8-channel split head coil at the China Rehabilitation Research Center. We collected two types of images: (1) high-resolution 3D T1-weighted MPRAGE images in the sagittal plane with a matrix size = 512 × 512, voxel size = 0.49 × 0.49 × 0.70 mm3, repetition time (TR) = 12.26 ms, echo time (TE) = 4.2 ms, inversion time = 400 ms, field of view (FOV) = 250 × 250 mm2, flip angle = 15°, and slice number = 248; and (2) FLAIR T2-weighted images in the axial plane with a matrix size = 512 × 512, voxel size = 0.49 × 0.49 × 5 mm3, TR = 8,002 ms, TE = 127.57 ms, inversion time = 2 s, FOV = 250 × 250 mm2, flip angle = 90°, and slice number = 28. To improve the image quality, the T1 image was scanned twice. The two scans were then co-registered and averaged for the following analyses. All imaging data can be found at the Open Science Framework database (URL: https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb).
We used structural-property-pattern (lesion)-based RSA to investigate semantic and modality-specific attribute representation in WM connections and GM regions. Similar to the conventional RSA, which is a highly fruitful method to research the neural representation in cortical regions using functional imaging data, the structural-property-pattern (lesion)-based RSA computes the relationship between the neural RDMs and behavioral or theoretical RDMs. The main difference is that the neural RDMs in this study were constructed by machine-learning models based on performances on neuropsychological tests and patients’ brain structural lesion patterns. The main rationale for this neural similarity measure is that if a WM connection pattern contains certain aspects of information shared by the naming process of two items (e.g., some semantic features), models trained with one item should also be able to predict naming accuracy of the other item to some degree. We first extracted the lesion features, balanced item labels by bootstrapping, input the lesion features and balanced labels into SVM training and testing to obtain the neural RDM, and used permutation to estimate the significance level of the neural RDM. The full pipeline is shown in Fig 1 and the details for each of these steps are described below in turn. The scripts of the full pipeline can be found at https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb.
The neural RDMs were correlated with behavioral RDMs using Spearman correlation. Specifically, for each WM connection, its neural RDM (a 100 [-item] × 100 [-item] matrix) and the semantic RDM (a 100 [-item] × 100 [-item] matrix) were both converted to a 1 × 4,950 vector. Correlation was computed on these two vectors (4,950 pairs of values). The r values were used to determine the extent of specific information encoded in the WM connections/GM regions. The FDR (q < 0.05) was used for multiple comparison correction. To investigate the higher-order semantic effects beyond modality-specific attributes, partial correlation analyses were performed between the semantic RDM and neural RDMs, with the modality-specific attribute RDMs (and the peripheral and categorical matrices) as nuisance variables. As explained in the “Behavioral RDM Construction” session, we adopted two ways of treating missing values in the modality-specific attributes (e.g., animal items were not rated on “manipulation” property)—setting it to be 1 (most dissimilar with other items on this modality) or to “NaN” (missing value). The RSA mapping procedure was implemented using a custom MATLAB function.
|
10.1371/journal.pbio.1000136 | Long-Term Relationships between Synaptic Tenacity, Synaptic Remodeling, and Network Activity | Synaptic plasticity is widely believed to constitute a key mechanism for modifying functional properties of neuronal networks. This belief implicitly implies, however, that synapses, when not driven to change their characteristics by physiologically relevant stimuli, will maintain these characteristics over time. How tenacious are synapses over behaviorally relevant time scales? To begin to address this question, we developed a system for continuously imaging the structural dynamics of individual synapses over many days, while recording network activity in the same preparations. We found that in spontaneously active networks, distributions of synaptic sizes were generally stable over days. Following individual synapses revealed, however, that the apparently static distributions were actually steady states of synapses exhibiting continual and extensive remodeling. In active networks, large synapses tended to grow smaller, whereas small synapses tended to grow larger, mainly during periods of particularly synchronous activity. Suppression of network activity only mildly affected the magnitude of synaptic remodeling, but dependence on synaptic size was lost, leading to the broadening of synaptic size distributions and increases in mean synaptic size. From the perspective of individual neurons, activity drove changes in the relative sizes of their excitatory inputs, but such changes continued, albeit at lower rates, even when network activity was blocked. Our findings show that activity strongly drives synaptic remodeling, but they also show that significant remodeling occurs spontaneously. Whereas such spontaneous remodeling provides an explanation for “synaptic homeostasis” like processes, it also raises significant questions concerning the reliability of individual synapses as sites for persistently modifying network function.
| Neurons communicate via synapses, and it is believed that activity-dependent modifications to synaptic connections—synaptic plasticity—is a fundamental mechanism for stably altering the function of neuronal networks. This belief implies that synapses, when not driven to change their properties by physiologically relevant stimuli, should preserve their individual properties over time. Otherwise, physiologically relevant modifications to network function would be gradually lost or become inseparable from stochastically occurring changes in the network. So do synapses actually preserve their properties over behaviorally relevant time scales? To begin to address this question, we examined the structural dynamics of individual postsynaptic densities for several days, while recording and manipulating network activity levels in the same networks. We found that as expected in highly active networks, individual synapses undergo continual and extensive remodeling over time scales of many hours to days. However, we also observed, that synaptic remodeling continues at very significant rates even when network activity is completely blocked. Our findings thus indicate that the capacity of synapses to preserve their specific properties might be more limited than previously thought, raising intriguing questions about the long-term reliability of individual synapses.
| Synapses are widely believed to constitute key loci for modifying the functional properties of neuronal networks, possibly providing the basis for phenomena collectively referred to as learning and memory [1],[2]. Indeed, an overwhelming body of literature supports the notion that synapses are “plastic”, that is, change their functional characteristics in response to specific activation patterns. The hypothesis that activity-dependent changes to synaptic characteristics constitutes a key mechanism for modifying neuronal network function also implies, however, that synapses, when not driven to change their characteristics by physiologically relevant stimuli, should retain these characteristics over time. Otherwise, physiologically relevant modifications to network function would be gradually lost due to stochastic, spurious changes or spontaneous drift. Thus, it might be expected that the capacity of synapses for directed change—synaptic plasticity—should be accompanied by a tendency to retain their characteristics at all other times, a phenomenon we will refer to here as “synaptic tenacity”.
The advent of molecular imaging techniques and the ability to study the molecular dynamics of specific molecules are gradually leading to the realization that synapses are not static, rigid structures; rather, they are made of multimolecular protein ensembles that exhibit significant dynamics at time scales of seconds to hours. Such dynamics include the recruitment and dispersal of regulatory constituents, lateral diffusion, endocytosis and exocytosis of postsynaptic neurotransmitter receptors, cytoskeletal dynamics and spine “morphing”, loss, incorporation, and turnover of scaffold molecules, and the interchange of synaptic molecules, multimolecular complexes, and synaptic vesicles among neighboring synapses (reviewed in [3]–[11]). When considering the bewildering dynamics exhibited by synaptic molecules, it becomes apparent that the long-term tenacity of synaptic structure and, by extension, synaptic function is not at all an obvious outcome. Yet to date, very little is known on the long-term tenacity of individual synapses [12].
Despite the molecular dynamics of synaptic constituents, most central nervous system (CNS) synapses appear to be quite persistent, although some degree of synapse formation and elimination is observed, depending on brain region, type of synapse, animal age, and imaging techniques [13]–[20] (reviewed in [8],[21]). Interestingly, however, even persistent synapses, when examined over long time scales (days), seem to exhibit considerable morphological changes (for example, [14],[15],[17],[22],[23]; see also [18]). In most of these studies, it was surmised that the observed changes in synaptic morphology represented structural manifestations of synaptic plasticity processes.
In most of the aforementioned studies, synapses were visualized by means of volume-filling fluorescent dyes (mainly enhanced green fluorescent protein [EGFP] or its spectral variants) and identified on the basis of typical pre- and postsynaptic morphological features (i.e., axonal varicosities and dendritic spines, respectively), whereas functionally relevant reporters, such as synaptic vesicle, postsynaptic receptor, active zone, or postsynaptic density (PSD) molecules were rarely used. Furthermore, even though manipulations aimed at altering network activity were performed in some of these studies, actual network activity was not recorded. Thus, the actual relationships between synaptic tenacity, synaptic remodeling, and network activity over these long time scales remained unknown.
To evaluate the tenacity of individual synaptic structures over behaviorally relevant time scales and differentiate between activity-dependent and activity independent-synaptic remodeling, an experimental system is needed in which both structural dynamics of individual synapses and electrical activity can be monitored continuously and simultaneously at sufficiently high temporal resolutions for very long periods. At present, this is an extremely challenging requirement, in particular where in vivo studies are concerned. We therefore developed a novel system, based on networks of rat cortical neurons in primary culture, that allowed us to continuously follow and record the structural dynamics of individual PSDs over time scales of minutes to weeks while concomitantly recording (and manipulating) network activity in the same preparations. We find that the vast majority of PSDs in this preparation undergo significant, continuous remodeling over time scales of many hours and days. The direction and extent of PSD remodeling are strongly affected by network activity levels, but remodeling does not cease upon suppression or elimination of activity. Our findings, described below, thus indicate that the tenacity exhibited by individual synapses over time scales of days is rather limited and may indicate that structural (and by extension, functional) properties of individual synapses experience significant drift over long durations.
In order to examine the tenacity of individual synapses, a system was needed that would allow us to record the structural dynamics of individual synaptic structures while concomitantly recording network activity in the same preparations and to do so continuously for many days. The experimental system we developed for this purpose was based on primary cultures of cortical neurons obtained from neonatal rats and plated on substrate-integrated multielectrode array (MEA) dishes [24]–[27]. To allow for the use of high numerical aperture oil-immersion objectives (that typically have short working distances), we used special MEA dishes made of very thin glass (180 µm) that are ideally suited for high-resolution imaging. Each dish contained 59 electrodes arranged in an 8×8 grid with interelectrode distances of 200 µm. Although the flat 30-µm diameter electrodes were opaque, the leads were transparent, resulting in minimal optical obstructions in the imaged regions (Figure 1A).
PSDs were visualized by expressing an EGFP-tagged variant of the PSD molecule PSD-95 (PSD-95:GFP). PSD-95 [28],[29] is a major postsynaptic scaffold protein that is thought to cluster postsynaptic NMDA receptors at postsynaptic sites. Most importantly, PSD-95, through interactions with transmembrane AMPA receptor regulatory proteins (such as stargazin) is believed to dictate the number of AMPA receptors found within the postsynaptic membrane (reviewed in [30]; see also [31]–[33]). Fluorescently tagged PSD-95 was used previously to study excitatory synapse formation (for example, [22],[34]–[39]), PSD turnover (for example, [40]–[42]), and PSD remodeling [22],[43], and was shown to faithfully represent PSD architectural rearrangements [43]. The particular EGFP-fusion protein used here was characterized extensively [35],[44]; When expressed in cultured hippocampal neurons, it was shown to localize correctly to postsynaptic structures associated with functional presynaptic sites, colocalize with the AMPA receptor subunit GluR1, and only negligibly affect gross synaptic characteristics [35]. In the current study, we used a third-generation lentiviral expression system [45] to express this fusion protein after an extensive series of preliminary experiments showed that this method was greatly preferable over more common transfection methods (calcium phosphate, cationic lipids, and electroporation): Lentivirus-based expression was easily titratable, resulted in low and constant expression levels, and importantly, unlike the aforementioned transfection methods, did not affect network activity properties or reduce cell numbers. Transduction was performed on day 5 in vitro leading to PSD-95:GFP expression in a small (10 to 50) number of neurons in each dish (and, occasionally, in nonneuronal cells). As shown in Figure 1, PSD-95:GFP assumed a punctate appearance, with the puncta commonly located at the tips of dendritic spines.
To allow for long-term (many days) combined optical/electrophysiological recordings from these preparations, a commercial MEA headstage/amplifier was installed on a custom-built confocal laser scanning (inverted) microscope (CLSM) equipped with a robotic XYZ stage. In each experiment, the MEA dish was covered with a custom-built cap and placed in the headstage/amplifier connected to the CLSM's robotic stage. The MEA dish and oil-immersion objective were heated to 37°C, and a sterile mixture of 5% CO2, 95% air was streamed into the MEA dish. An ultraslow perfusion system was used to exchange the media at very low rates (two volumes per day). Images were collected automatically at 30-min intervals from four to 12 fields of view (or sites; ∼95×70 µm in size), with each site representing a portion of the dendritic arbor of a different neuron (Figure 1). Seven to 26 Z-sections were collected at each site, beginning at a predetermined offset above the upper glass surface. To correct for focal drift, the focal plane of the upper glass surface was located automatically before collecting each image stack [35]. Network activity in the imaged networks was recorded from all 59 (extracellular) embedded electrodes (Figure 2). For each electrode, waveforms of individual action potentials were stored digitally and then converted into series of discrete events (Figure 2D). Under these conditions, preparations were routinely maintained on the microscope stage, recorded from and imaged continuously for many days and even weeks (Figure 1C and 1D).
The synaptic identity of PSD-95:GFP puncta was verified by labeling active presynaptic compartments in live neurons with fluorescent antibodies against the lumenal domain of the synaptic vesicle protein synaptotagmin-1 ([46]; see Materials and Methods for further details). As shown in Figure S1, >80% of PSD-95:GFP puncta were juxtaposed against functional synaptic vesicle recycling sites, in good agreement with prior measurements performed in cultured hippocampal neurons using the styryl dye FM 4–64 (83%; [35]). The high degree of colocalization along with the fact that labeling was based entirely on spontaneous activity strongly indicates that most PSD-95:GFP puncta represent bona fide glutamatergic synapses that are activated by spontaneous network activity.
MEA dishes allowed us to sample network activity from 59 locations in the network, but due to the presence of multiple neurons near each electrode, the identity of neurons from which activity was recorded remained obscure. Furthermore, due to the random nature of lentiviral infection, neurons expressing PSD-95:GFP were not necessarily located over any particular electrode. It was thus necessary to verify that the activity recorded through the electrodes faithfully represented the activity of those neurons expressing PSD-95:GFP and followed by time-lapse microscopy. To that end, we took advantage of the fact that most activity in networks of dissociated cortical neurons occurs in the form of synchronized bursts (for example, [25],[27],[47]–[49]; see also Figure 2A–2C). By using fluorescent calcium indicators and synchronized MEA recordings, we found that practically all network bursts were time-locked to calcium transients measured by line scanning in the somata of PSD-95:GFP-expressing neurons (Figure S2; 27 neurons, four separate experiments). These experiments strongly indicate that the characteristics of network activity recorded through the MEA faithfully represent, at least to a first approximation, the activity of PSD-95:GFP-expressing neurons. Furthermore, the tight correlation between network bursts and calcium transients suggests that these neurons respond well to excitatory synapse activation, implying that PSD-95:GFP expression does not severely impair glutamatergic synapse functionality.
In summary, the system described here allowed us to follow structural dynamics of individual and functional glutamatergic synapses at relatively high temporal resolutions and over many days while concomitantly recording network activity in the same preparations.
We initially set out to the examine the long-term stability of PSD-95:GFP puncta under baseline conditions, that is, without experimentally manipulating network activity. To that end, we used preparations maintained in culture for at least 17 d. At this stage, networks of rat cortical neurons are considered to be relatively mature, and beyond the phase of extensive dendrite extension and synapse formation. The preparations were mounted on the CLSM and followed by combined time-lapse imaging and electrophysiological recordings as described above. At the end of each experiment, spike counts per electrode were extracted from the electrophysiological recordings, and the numbers of PSD-95:GFP puncta and their respective fluorescence intensities were extracted from maximal intensity projections of Z-stacks for each time point and each site (see Materials and Methods).
These experiments resulted in two highly consistent observations. The first concerns network activity. As noted above, cultured networks of dissociated cortical neurons typically develop complex patterns of spontaneous activity composed of asynchronous action potentials and synchronized bursts (Figure 2A–2C). These forms of activity were observed here as well, as soon as recording was initiated. However, over the first 1–2 d, we consistently observed significant elevations of spontaneous activity levels (Figure 3A). These elevations reflected both increases in the number of “active” electrodes as well as increases in the frequencies of action potentials measured from individual electrodes (Figure S3). Intermittently, periods of “superbursts” (i.e., bursts of bursts; [49]) were recorded (see below), giving rise to significant variability in action potential counts from one minute to the next. This gradual increase in network activity at the beginning of the experiment was observed in practically all experiments regardless of preparation age, indicating that it was due, somehow, to the environmental conditions introduced during the experiments (see below).
The second consistent observation concerns the morphological complexity of postsynaptic structures along dendritic segments. In practically all experiments, we observed a gradual increase in the total number of PSD-95:GFP puncta, mainly (but not exclusively) due to increased density of PSD-95:GFP puncta along existing dendritic segments, at both spine tips and shafts (Figures 1C and 3B). Moreover, the population of PSD-95:GFP puncta changed from one that was relatively uniform in terms of fluorescence intensity to one that contained both very large, bright puncta as well as many small, dim puncta (Figures 1C, 3C, and 3D). Interestingly, the broadening of the puncta intensity distribution was usually transient, and was partially reversed when network activity reached relatively high levels after 1–2 d (Figure 3A, 3C, and 3D). As the majority of PSD-95:GFP puncta were typically juxtaposed against functional presynaptic boutons (Figure S1; see also [35]), the observed increases in PSD-95:GFP puncta number reflected, in all likelihood, increased numbers of glutamatergic synapses. Furthermore, given that PSD size, PSD-95:GFP fluorescence, spine head dimensions, AMPA receptor number, and glutamate-induced synaptic currents, are well correlated [16],[50]–[54] changes in PSD-95:GFP content, probably reflected changes in PSD size and possibly in synaptic strength [8],[53].
The (post)synaptic remodeling described above resulted in dendrites assuming morphological characteristics more akin to those of dendrites in vivo. Yet we could not rule out the possibility that these morphological changes were actually reflecting pathological processes induced by the environmental conditions during experiments or damage inflicted by continuous imaging. To examine the possibility that the experimental conditions were detrimental to neuronal vitality, we used the same system to follow the development of less mature networks in which vigorous growth and synapse formation are known to occur, because here, pathological phenomena such as growth cessation, axon/dendrite retraction, and synapse elimination, are clearly recognizable. To that end, preparations were mounted on the CLSM starting from day 9–10 in vitro, maintained in the environmental conditions described above, and imaged at higher frequencies (every 10 min instead of 30) for about 1 wk. Dendritic development in these experiments appeared to proceed as expected: new branches were added, synapses were formed at high rates, and network activity levels increased 10- to 20-fold (Figure S4 and Video S1). In none of these experiments (n = 7) did we observe signs of damage. In fact, these experiments resulted in exciting and, to the best of our knowledge, unprecedented recordings of dendritic development and synapse formation that will be described elsewhere. These experiments, therefore, do not support the possibility that the experimental conditions used here adversely affect neuronal viability, and lead us to conclude that the synaptic remodeling described above is not secondary to pathological processes.
In the aforementioned experiments, we observed significant changes in PSD-95:GFP puncta number and fluorescence intensity, which, in all likelihood, reflected changes in glutamatergic synapse number and PSD size. This remodeling occurred concomitantly with significant changes in network activity, which pointed to the possibility that the two phenomena might be causally related. It should be noted, however, that unlike network activity, that generally increased over time (Figure 3A), the initial broadening of the PSD-95:GFP fluorescence intensity distribution (and the gradual increase in mean puncta fluorescence) was usually followed by a second phase during which the intensity distribution partially recovered (as did mean puncta fluorescence; Figure 3C and 3D) and thereafter remained stable for days, indicating that the potential relationships between the two phenomena are not straightforward.
To determine whether the observed changes in PSD-95:GFP puncta number and fluorescence intensity were dependent on changes in network activity, we repeated the experiments described above except that here, spontaneous network activity was blocked by adding tetrodotoxin (TTX) about an hour after the experiments were started. As shown in Figure 3E–3H, blocking network activity did not block the initial broadening of the PSD-95:GFP fluorescence intensity distribution. However, the second phase (the partial recovery of the intensity distribution) was completely lost. Instead, the distributions of PSD-95:GFP puncta fluorescence intensities continued to broaden, and mean puncta fluorescence continued to increase (Figure 3G and 3H). In addition, the number of puncta did not increase over time, and in fact, gradual decreases in puncta numbers were observed (Figure 3F).
These experiments indicate that the initial broadening of PSD-95:GFP puncta fluorescence intensity distribution is not driven by activity. Rather, it seems to be driven by the exposure to environmental conditions during experiments. Given that ambient temperature and atmospheric conditions were identical to those in the incubators in which preparations were maintained, the most likely “culprit” is the slow perfusion. Indeed, these phenomena are not observed if perfusion is not applied (unpublished data). On the other hand, in the absence of perfusion, the long-term viability of these preparations was drastically impaired. Interestingly, media turnover rates (∼0.15%/min) were one to two orders of magnitudes lower than cerebrospinal fluid (CSF) turnover rates in the intact rat brain (1% to 16%/min; [55]), indicating that perfusion rates were not excessively high.
In contrast to the initial broadening of PSD-95:GFP fluorescence intensity distributions, the subsequent constriction of fluorescence intensity distributions was clearly dependent on network activity. This dependence indicates that an increase in activity levels is associated with a reduction in mean PSD size. This finding is consistent with the concept of “synaptic scaling” [56]–[58], that is, the adjustment of synaptic strength to match neuronal activation levels. Interestingly, following the initial broadening and subsequent constriction, PSD-95:GFP puncta fluorescence intensity distributions remained relatively stable as long as activity levels did not change significantly (as exemplified in Figure 3C). These observations are consistent with the possibility that PSD size is generally stable, with changes in activity followed by uniform and gradual scaling of PSD size (multiplication by a scalar, for example). However, as shown next, this does not seem to be the case.
The analysis described so far indicates that PSD size distribution remains rather stable as long as activity levels are not drastically altered. However, this analysis was performed at the population level, and thus did not provide information on the long-term stability of individual PSD sizes. To quantify the stability of individual PSDs, we developed software for tracking identified PSD-95:GFP puncta in long time series of image stacks, and used it to quantify the fluorescence of individual PSD-95:GFP puncta at a temporal resolution of 30 min over several days. Although PSD-95:GFP puncta were relatively stable over time scales of several hours, some puncta exhibited considerable dynamics (lateral movements, merging, and splitting) as previously described [22],[35],[36]. We therefore limited our analysis to PSD-95:GFP that could be identified and tracked reliably throughout the experiments, excluding puncta that became obscured by, merged with, or split from other puncta, but not excluding puncta that simply appeared or disappeared during the experiments. The fluorescence of all tracked puncta at all time points was then measured, and these data were compared to network activity during the same period.
Figure 4 shows such an analysis performed for one neuron of the experiment of Figure 3A–3D. Five (out of ∼200) puncta tracked over the entire experiment are shown in Figure 4B. Note that it is not always apparent that the puncta shown in each time series are indeed the same ones. However, only one frame out of 20 sequential frames is shown in this figure (10-h intervals), whereas imaging was performed at 30-min intervals, allowing for very reliable tracking of discernable objects (see also Figure 1D). Plotting the fluorescence of these five puncta over >90 h (Figure 4D) revealed that some puncta exhibited significant changes in their fluorescence over this period, whereas the fluorescence of others was more stable. Yet, when the fluorescence of all tracked puncta is rendered for long stretches of time (days), the instability of individual PSD-95:GFP puncta becomes strikingly apparent (Video S2). These observations indicate that the seemingly static size distributions of Figure 3 are, in fact, population steady states, with individual synapses within this population undergoing continual and extensive remodeling.
To examine the dependence of changes in PSD size on initial PSD size, changes in the fluorescence of individually tracked puncta at the end of consecutive, 7-h time windows, were plotted as a function of their fluorescence at the beginning of each time window. To minimize the effects of short-term fluctuations, data were first “smoothed” with a five–time point (2-h) kernel. As shown in Figure 4E, significant changes in puncta fluorescence over time were observed for all puncta, regardless of their initial size. Interestingly, however, as activity levels increased, a relationship developed between initial puncta fluorescence and subsequent changes in fluorescence: Bright puncta tended to become dimmer, whereas very dim puncta tended to become brighter, as if activity was driving the convergence of PSD sizes to some optimal value. These relationships could be approximated reasonably well by linear regression fits to the data. In should be noted, however that the R2 values of these linear fits were not very high, suggesting that the direction and magnitude of PSD size change were only partially determined by their instantaneous size.
To further examine the dependence of the aforementioned relationship on network activity, identical experiments were performed in which spontaneous network activity was blocked abruptly by adding TTX 40 to 70 h after the experiments were started. Significant changes in puncta fluorescence over time were still observed in the presence of TTX, and such changes were observed for small and large puncta alike (Figure 5 and Video S3). Strikingly, however, relationships between initial PSD-95:GFP puncta fluorescence and subsequent changes in fluorescence were lost. This is indicated by the fact that the slopes of linear regression lines fit to these data approached zero. Similar findings were observed for a total of six neurons (Figure 5G). As expected, TTX addition led to a broadening of the PSD-95:GFP puncta fluorescence distribution (Figure S5A). Furthermore, this was seen for both the entire PSD-95:GFP population and for the smaller population of tracked PSD-95:GFP puncta (Figure S5B), indicating that the population of tracked puncta faithfully represented the entire PSD-95:GFP puncta population.
Relationships between PSD-95:GFP puncta fluorescence and subsequent changes in puncta fluorescence were also examined by manipulating network activity with diazepam, a coagonist of GABAA receptors in widespread clinical use. As shown in Figure 6, bolus additions of diazepam at two different concentrations (2.5 and 25 µg/ml) led to temporary and recoverable reductions in network activity levels (Figure 6C). Despite the reduction in activity levels, changes in puncta fluorescence still occurred to similar degrees, although the dependence of such changes on initial PSD-95:GFP puncta fluorescence was reduced or lost, in particular at higher diazepam concentrations (five experiments, tracked synapse data from four neurons).
Several points are worth further emphasis. First, changes in puncta fluorescence continued to occur even when activity was suppressed or eliminated altogether. In fact, fluorescence intensity differences measured for individual PSD-95:GFP puncta between consecutive images distributed similarly for TTX-treated and untreated preparations (see Figure 7A). Second, even in active networks, instantaneous PSD size was not the sole determinant of subsequent changes in PSD size. This is evident, not only from the relatively low R2 values of the linear regression fits, but also from the fact that the distribution of PSD sizes in active networks did not continue to constrict with time (Figure 3C and see also Figure S5A), indicating that activity-driven convergence of synaptic size distribution is balanced by processes that act to broaden PSD size distribution. Third, changes in the strength of relationships between instantaneous PSD size and subsequent changes in PSD size were slow to develop, and occasionally were only observed several hours after activity levels had changed significantly. Finally, although reductions in network activity levels were generally associated with increases in PSD size (Figure 5), we practically never observed (except one time point for one cell) a positive relationship between PSD size and the extent to which PSD size changed following reductions in activity levels. Put differently, once activity was blocked, changes in PSD size, in absolute terms, were similar for small and large PSDs alike (see Figure 5, 90–97 h, for example). This observation strongly argues against the possibility that increases in PSD size following reductions in activity levels can be viewed as simple multiplicative scaling.
The long-term recordings of synaptic remodeling described so far indicate that (1) spontaneous network activity maintains distributions of synaptic sizes within rather constrained boundaries; (2) reductions in network activity levels result in a broadening of synaptic size distributions, increases in mean synaptic size, and reductions in synapse numbers; (3) most synapses exhibit significant changes in size over time; (4) in active networks, changes in synaptic size are partially dependent on momentary synapse size: large synapses tend to become smaller, whereas small synapses tend to become larger; and (5) when activity is blocked or significantly suppressed, synapses continue to change their sizes, but the direction and extent of these changes become independent of momentary synapse size.
We hypothesized that the phenomenological relationships between network activity and synaptic remodeling described above could be explained by the following set of rules: (1) synapses continuously undergo spontaneous, activity-independent changes (drift) in their size; (2) activity acts to reduce the size of large synapses on the one hand, and increase the size of small synapses on the other; (3) new, small synapses are continually formed at a constant rate; and 4) synapses whose size is reduced beneath some threshold are eliminated.
To examine whether this set of rules could, at least in principle, explain the phenomena described above and produce synaptic size distributions similar to those measured experimentally, we created a simple numerical model in which sizes and fates of individual synapses were updated over time according to these four rules (see legend of Figure 7 and Materials and Methods for further details). Figure 7C–7F shows simulations seeded with the initial puncta counts and intensities measured in the experiments of Figures 4 and 5. As these figures show, the numerical simulations recapitulated the major phenomenological relationships between network activity and synaptic size distributions (Figure 7C and 7E). Furthermore, they predicted correctly the changes in synaptic counts measured in these experiments (Figure 7D and 7F). These findings indicate that relationships between synaptic size distributions and network activity levels can be accounted for, at least in principle, by the four simple rules described above without a need for explicit “scaling” or compensatory mechanisms invoked by reductions in network activity levels (although this does not preclude their existence).
Although the simulated data generally approximated the experimental measurements quite well, they failed to account for a small population of particularly large synapses observed in active networks (visible as blue dots in the upper regions of the histograms of Figure 7C and 7E). We do not think this is related to imperfect simulation parameters but rather to the existence of at least one other “rule” that drives the formation of large synapses in highly active networks. Indeed, we noted that relatively large synapses tended to appear during periods of particularly high levels of synchronous activity, often associated with “superbursting” [48], observable as periods of significant variability in action potential counts from one minute to the other. In fact, a comparison of particularly bright (1.5 standard deviations above mean puncta fluorescence) PSD-95:GFP puncta appearance rates to burst rates revealed strong temporal relationships between these two phenomena (Figure S6A), relationships observed in practically all experiments in which superbursting occurred. In one striking example, tracking particularly bright puncta backward in time revealed that 18 out of 18 such bright puncta exhibited dramatic increases in fluorescence with the onset of superbursting or seemed to appear de novo (Figure S6B, S6C, and S6E; see also Figure 6, puncta 1 and 4). It is likely that these large PSD-95:GFP puncta represent synapses that underwent forms of potentiation associated with spine-head enlargement (e.g., [23],[53],[59]–[64]). In fact, we suspect that high levels of synchronized activity, rather than high activity levels per se are also the driving force behind the aforementioned tendency of small puncta to grow larger in highly active networks (see [65] and discussion below). As we did not have access to the precise firing timings of individual pre- and postsynaptic partners, this possibility remains somewhat speculative. Interestingly, however, as our system did allow us to follow these synapses for relatively long periods, we were able to compare their tenacity to that of the rest of the synaptic population. These observations indicated that newly enlarged synapses did not necessarily fair much better than other synapses in terms of their long-term tenacity, even in the presence of TTX (Figure S7; see also [23],[59]), although their lifespans were typically longer (unpublished data).
Individual CNS neurons, even those in our cell culture preparations, typically receive thousands of excitatory synaptic inputs of varying strengths. The experiments described so far reveal that individual PSD-95:GFP puncta exhibited significant remodeling over time, probably reflecting significant changes in synaptic strengths. Given the high levels of spontaneous activity in these networks, it might have been expected that this remodeling is driven to a large extent by network activity. However, as shown above, remodeling does not cease upon suppression or elimination of network activity. To what degree are the relative weights of excitatory inputs of a given neuron reconfigured by ongoing activity? How stable does this configuration remain when the physiologically relevant driving force, i.e., activity, is removed?
To evaluate the impact of synaptic remodeling on the synaptic configurations of individual neurons, we defined a measure for quantifying the degree to which PSD-95:GFP puncta belonging to a given neuron changed their sizes relative to each other over time (irrespective of global changes in puncta fluorescence). The basic idea was to sort the synapses formed on a given neuron according to their sizes. Then, at each subsequent time point, the same synapses were sorted again according to their new sizes. The degree to which each synapse changed its rank relative to its original rank was then determined, and finally, all rank changes for all synapses were summed and normalized to give a value between 0 and 1. This measure (denoted Mt) was taken to represent the relative remodeling for that time point and for that particular neuron. In practice, Mt was calculated according to the following equation:where n is the population of tracked PSD-95:GFP puncta, rt is its rank at time t, and r0 its rank at time t = 0. Mt will approach 1.0 if the rank of each synapse at time t is furthest away from its rank at time t = 0, and will approach 0.67 if the ranks at time t bear no relationships beyond chance to the ranks at t = 0.
Figure 8A shows how the relative remodeling measure changes over 30 h before and 30 h after the addition of TTX (four different neurons, 505 PSD-95:GFP puncta). As shown in this figure, relative remodeling occurs about twice as fast in active networks as compared to networks in which activity was blocked. This effect is clearly activity dependent and not time dependent because no differences in relative remodeling rates are observed in two consecutive 30-h time windows when TTX is not added (Figure 8B). It is important to stress, however, that remodeling continues at significant, albeit slower rates even in the presence of TTX. This remodeling is not an artifact of imaging-related noise because practically no change in M was observed in control experiments performed in exactly the same experimental conditions using paraformaldehyde-fixed, PSD-95:GFP-expressing neurons (Figure 8A and 8B). Measurements made in a smaller number of neurons followed for longer periods (70 h, or about 3 d) further show that relative remodeling in this system is extensive, even in the presence of TTX, and that relative remodeling does not plateau within this time frame (Figure 8C).
These findings indicate that ongoing activity drives significant changes in the synaptic reconfigurations of individual neurons. Just as important, however, these findings also show that substantial “drift” in synaptic configurations occurs even in the absence of network activity, the physiologically relevant driving force.
Here, we describe experiments aimed at evaluating the long-term tenacity of individual glutamatergic synapses. To that end, we developed a novel system that allowed us to continuously follow and record the structural dynamics of synapses formed between rat cortical neurons in primary culture, over time scales of minutes to weeks, while concomitantly recording and manipulating network activity in the same preparations. We found that in spontaneously active networks, the range and distribution of synaptic sizes was maintained within rather constrained boundaries. Yet, when synapses within these populations were followed individually, the majority exhibited considerable changes in size over time scales of hours and days, and it became evident that the apparently static size distribution was in fact a steady state of synapses undergoing continual remodeling. Further analysis revealed that the extent and direction of this remodeling was partially dependent on momentary synaptic size, with large synapses exhibiting a tendency to grow smaller, and small synapses a tendency to grow larger. Blocking network activity did not stop synaptic remodeling, but changes in synaptic size became independent of momentary synaptic size. This undirected and unconstrained “drift” of synaptic size was associated with a broadening of synaptic size distributions, and gradual reductions in synaptic numbers. From the perspective of the single neuron, our experiments showed that activity drives changes in the relative weights of its excitatory inputs, but also revealed that these weights exhibit significant “drift” even in the absence of any network activity.
These findings point to several potentially important conclusions: First, they suggest that synaptic size exhibits continuous and significant “drift” over time (many hours to days) even in the absence of activity, indicating that the structural, and by extension, the functional tenacity of synapses is somewhat limited over long time scales. Second, they suggest that activity acts to partially direct this drift, promoting the convergence of synaptic sizes on some “optimal” size distribution. Third, although our findings do support previous reports that activity blockade is associated with increased synaptic size, they do not support the notion of multiplicative scaling [66]. Finally, our findings support the widespread belief that network activity drives synaptic remodeling and alters the relative weights of synapses formed on a given neuron. However, they also indicate that these relative weights undergo significant spontaneous, activity-independent changes as well.
The experiments described here were based on several techniques: networks of dissociated cortical neurons, MEA substrates, automated multisite confocal microscopy, fusion proteins of synaptic proteins, lentiviral expression vectors, and automated image analysis. Although most of these techniques are in common use, it was their unique combination that allowed us to follow synaptic remodeling and relate it to network activity over relatively long time scales. Of particular note is the use of MEA substrates fabricated on very thin glass (ThinMEAs) that allowed the use of high numerical aperture objectives, resulting in both high-resolution images and very efficient light collection. In fact, control experiments performed in paraformaldehyde-fixed neurons revealed that photobleaching rates did not exceed 10% per day (48 time points per day, 15 focal planes per time point). This was undoubtedly an essential factor in our ability to image neurons at relatively high rates for such prolonged periods.
Another key technique was the use of an ultraslow perfusion system. This system maintained cell viability in a remarkable fashion: unlike typical long-term experiments carried out at physiological temperatures, where some rundown is usually observed after 12–24 h, we observed no signs of rundown even after 2 wk of continuous imaging. As the perfusion medium was identical to the normal growth medium, it would seem that medium replacement was the critical factor. Furthermore, we found that slow exchange rates were imperative, as rapid medium replacement was detrimental as well. A byproduct of the slow perfusion was a gradual increase in network activity resulting in very high and complex spontaneous spiking patterns. At present, we do not know why this occurs, but this phenomenon was instrumental in exposing the effects of activity and accentuating the effects of activity suppression.
A potential concern is the use of an exogenous form of PSD-95 fused to EGFP (an ∼30 kDa polypeptide). We cannot exclude the possibility that the addition of EGFP interferes, perhaps in subtle ways, with the interactions of PSD-95 with its endogenous binding partners, with implications on PSD remodeling dynamics. Furthermore, PSD-95:GFP overexpression was previously shown to affect synaptic properties and even occlude forms of activity-induced synaptic plasticity [67]–[71]. The severity of such effects, however, probably depends on overexpression levels. For example, in the first study mentioned above, PSD-95:GFP expression levels were reported to be many fold greater than endogenous PSD-95 levels, whereas the use of Sindbis or Semliki forest viral vectors for PSD-95:GFP expression in others might have resulted in similar situations [72]. In agreement with this possibility, other studies using fluorescently tagged variants of PSD-95:GFP did not detect significant effects on synaptic properties [54],[73],[74]. It should be noted that overexpression levels in our hands were low and similar to those we have previously reported (∼27%; [35]), and at these expression levels, effects on postsynaptic and presynaptic properties were very small [35]. Finally, in the current study PSD-95:GFP puncta did exhibit both “homeostatic” forms of synaptic remodeling as well as synchronous activity-driven remodeling, in agreement with studies based on immunohistochemistry, electron microscopy, relatively inert reporter molecules such as EGFP, or live imaging of AMPA receptors (for example, [53],[63],[73],[75]–[78]). Given the low PSD-95:GFP expression levels here and the fact that the aforementioned forms of synaptic plasticity were not occluded in our system, it seems unlikely that the phenomena described here are solely artifacts of PSD-95:GFP overexpression, although, as mentioned above, we cannot exclude the possibility of the introduction of some quantitative inaccuracies [79].
A broader concern relates to the fact that the study was performed in dissociated cell culture. Although it was this very fact that allowed us to concomitantly record synaptic remodeling and network activity as described above, we cannot ignore the possibility that the limited tenacity exhibited by synapses here is somehow related to this experimental system. We could mention the fact that many phenomena pertaining to synaptic dynamics described in cell culture were also observed in vivo (compare, for example, [80] with [54]). Nevertheless, it would be prudent not to take the absolute values provided here too literally. It is also important to note that the experiments were performed in neurons that are relatively immature (3–4 wk in vitro) as compared to those in the mature rat brain. Given that several measures of synaptic dynamics subside with age [15],[16],[54],[74], the absolute rates of synaptic remodeling reported here might overestimate those that occur in the mature brain. Yet it is worth noting in this regard that in vivo imaging indicates that spine-head remodeling is quantitatively similar in young and adult mice [15].
Finally, it is worth stressing that MEA recordings, unlike single-cell recordings used in other studies concerning activity-induced synaptic remodeling (for example, [53],[63]), do not allow one to directly relate changes in the structural properties of an imaged synapse to changes in the strength of the connection it mediates (see also [81]). However, the integrated system described here did allow us to study the remodeling of very large numbers (thousands) of individual postsynaptic densities over times scales of minutes to weeks and to explore how these dynamics are affected by measured levels of network activity, and hence expose long-term phenomena that become apparent only at the population level. Thus, in spite of the potential drawbacks raised above, the advantages offered by this system are substantial.
Previous studies, initially in cell culture and later on in vivo as well, have shown that reductions in activity levels are followed by general increases in the strength of excitatory glutamatergic synapses, whereas enhanced activity levels have opposite effects (reviewed in [56]–[58]). It was suggested that these changes represent “homeostatic” mechanisms that serve to stabilize neuronal activity levels. Furthermore, it was shown that activity blockade-induced increases in synaptic strength were best explained by a scaling of synaptic strengths by the same multiplicative factor [66]. Finally, several studies [73],[78] have shown that synaptic scaling has a predominant, although not exclusive, postsynaptic component, manifested as an increase in the number of AMPA-type glutamate receptors localized to postsynaptic compartments.
Our findings that suppression of activity broadens the distribution of PSD-95:GFP puncta fluorescence intensities and increases mean PSD-95:GFP fluorescence (Figures 3 and 5) are in general agreement with the aforementioned studies (see also [77],[82]). Our findings, however, offer an alternative explanation for this phenomenon. First, these findings indicate that activity exerts a positive control on synaptic size, with changes in synaptic sizes inversely related to momentary synaptic size. Second, the findings indicate that the suppression of activity removes this positive control, resulting in unrestricted “drift.” These findings are consistent with the possibility that mean synaptic size thereafter increases because this drift is asymmetrically constrained: synapses grow both larger and smaller in apparently random fashion, but synapses that become too small are eliminated (see also [17]), further biasing the mean synaptic size toward larger values (Figure 7). Therefore, there might be no a priori need for mechanisms actively invoked by reductions in activity levels (although these might exist), that act to adjust synaptic size accordingly. Although the difference between negative signaling and the removal of positive signaling is somewhat semantic, it is a simpler explanation and thus, perhaps, more appealing.
An interesting finding was the observation that high levels of network activity were associated with gradual increases in PSD-95:GFP puncta numbers, whereas suppressed activity levels were associated with gradual decreases in PSD-95:GFP puncta numbers (see also [82],[83]). These phenomena could have resulted from changes in synapse formation rates, changes in synapse elimination rates, or both [40]. However, a third possibility exists: that high levels of activity, and in particular synchronous activity, promote the stabilization and growth of newly formed synapses, echoing the prominent role of synchronous activity in nervous system development [84]. Conversely, in the absence of such activity, the tendency of new synapses to be stabilized would be reduced, and thus more synapses would be lost. Indeed, recent work from De Roo and colleagues [23],[74] indicates that the vast majority of new synapses are transient, but activity, and in particular rhythmic activity, leads to the stabilization of new spines, possibly by promoting synaptogenic interactions with nearby axons [71] (see also [85]–[87]) and to their eventual enlargement. Given that most activity in our preparations was in the form of synchronized bursts, new synapses were probably more likely to be stabilized in active networks, as compared to networks in which activity was suppressed. By way of extension, synchronous activity may serve to promote the enlargement of a subset of small synapses—new and preexisting alike (Figure S6)—reducing the likelihood that their size will drift below some critical threshold resulting in their loss, and thus providing some explanation for the tendency we observed for small synapses to grow larger in active networks [65]. Furthermore, periods of highly synchronized activity often led to the appearance of large PSDs (Figure S6), in agreement with studies showing that synchronous activity and specific stimulation patterns lead to significant spine-head enlargement (for example, [23],[53],[59]–[64],[79]). Unfortunately, as mentioned above, MEA recordings did not allow us to relate changes in the structural properties of an imaged synapse to the specific activity patterns it experienced. Interestingly, however, we were able to follow the fate of such synapses, and we noted that in many cases, this enlargement was transient (Figure S7), in good agreement with similar observations recently made in cultured hippocampal slices [23]. We thus do not find strong evidence that the tenacity of such synapses is significantly different from that of other synapses.
High activity levels also promoted reductions in the size of large synapses (Figure 4). Given that synaptic size seems to be inherently limited [88], it is conceivable that larger synapses may be more sensitive to forces that limit synaptic size, and, by extension, by activity-driven, size-limiting forces. The exact nature of such activity-driven forces is unknown, but given the relatively long time scale of their actions, they are likely to involve protein exchange [54],[80],[89],[90], protein degradation [91], and local protein synthesis [92], although mechanisms activated on shorter time scales may also be involved [93]. Regardless of their specific nature, such processes would drive continuous change in synaptic sizes and lead to continuous reconfiguration of synaptic weights (Figure 8).
The long-term recordings of individual PSD-95:GFP puncta described here indicate that synaptic sizes exhibit spontaneous changes over time scales of many hours and days. This is probably not an artifact of cell culture, as fluctuations in spine-head size from one day to the next were observed in vivo (for example, [14],[15],[17],[18]). Interestingly, in a study published very recently, Yasumatsu and colleagues [94] analyzed fluctuations in the volumes of individual dendritic spines in cultured rat hippocampal slices. In common with previous studies, a volume-filling dye (EGFP) and two-photon microscopy were used to image dendritic spines once a day for several days. However, this study also examined the effects of NMDA-type glutamate receptor antagonists on spine volume fluctuations and showed that such fluctuations are observed also in the presence of such antagonists. As network activity levels were not measured, actual activity levels remained unknown, and thus, relationships between spine size fluctuations and network activity remained speculative. Perhaps the fact that no synaptic homeostasis-like processes were detected may indicate that basal activity levels were rather low to start with. Nevertheless, these findings further support the possibility that synaptic remodeling might persist even in the absence of activity and that the findings reported here are not solely artifacts of dissociated cell culture.
Yasumatsu et al. [94] also proposed that distributions in spine size, rates of spine formation and elimination, and the long-term persistence of large spines might be explained by such fluctuations according to a mathematical model based on Brownian motion (or “random walk”) with drift and reflecting boundaries. Indeed, we also found that generally similar models can adequately describe synaptic size distributions in relatively quiescent networks (Figure 7). However, we found that such models fall short of providing satisfactory accounts of synaptic size distributions in highly active networks, and that these seem to be governed by additional principles, such as network synchronicity levels (Figure S6). Most importantly, the mere persistence of large spines (proposed to function as “write-protected” devices) would not guarantee synaptic configuration stability, in particular when considering that these exhibited the largest activity-independent volume fluctuations [94]. Thus, significant synaptic reconfiguration seems to be an unavoidable consequence of activity-independent synaptic remodeling.
When considering the extensive molecular dynamics occurring in the minute synapse, the limited long-term tenacity of individual PSDs might not seem surprising. Perhaps it may be unrealistic to expect that a biological structure with the dimensions of a PSD, composed of dozens to hundreds of copies of individual proteins, whose function depends on even smaller numbers of molecules and is located at extreme distances from the soma, will maintain its structural and functional properties with pinpoint precision for many months and years [12]. Consequently, and assuming that our findings are not unique to the experimental system used here, several tentative conclusions may be drawn.
The first conclusion concerns the significance of individual synapses in “encoding” some functional characteristic of a neuronal network. Obviously, the weight of individual excitatory synapses in determining postsynaptic neuron output is rather limited to start with (see, for example, [95]). However, our findings indicate that the reliability of these weights, already challenged by fluctuations on short time scales (release failures, quantal fluctuations), is further challenged by processes occurring over longer time scales. This finding indicates that if indeed persistent changes in network function are realized by changes in synaptic weights [1],[2], the weights of many synapses might need to be changed to produce an enduring alteration of network function.
The second conclusion concerns the possibility that the limited tenacity of synapses, the “drift” they exhibit, is actually a fundamentally important feature of synapses, in particular in developing networks. One could imagine that synaptic drift and the consequential drift of network function constitutes an “exploratory” process that allows networks to explore the space of synaptic configurations in search of appropriate synaptic input levels or functionally “useful” or “successful” configurations. Following this logic, configurations that produce desirable results would be stabilized, perhaps by diffuse neuromodulatory systems activated by salient stimuli or “rewards,” input that is obviously absent from our preparations. Although entirely speculative at this point, it is intriguing to consider the possibility that the principles of diversity and selection, arguably the most universal principles of evolving biological systems, might also be fundamental principles in the processes that govern synaptic remodeling.
Primary cultures of rat cortical neurons were prepared in a similar manner to that described previously for hippocampal preparations [35]. Cortices of 1–2-d-old Sprague-Dawley rats were dissected, dissociated by trypsin treatment followed by trituration using a siliconized Pasteur pipette. A total of 1–1.5·×106 cells were then plated on thin-glass MEA dishes (MultiChannelSystems MCS), whose surface had been pretreated with Polyethylenimine (Sigma) to facilitate cell adherence. Cells were initially grown in medium containing minimal essential medium (MEM; Sigma), 25 mg/l insulin (Sigma), 20 mM glucose (Sigma), 2 mM l-glutamine (Sigma), 5 µg/ml gentamycin sulfate (Sigma), and 10% NuSerum (Becton Dickinson Labware). The preparation was then transferred to a humidified tissue culture incubator and maintained at 37°C in a gas mixture of 5% CO2, 95% air. Half the volume of the culture medium was replaced three times a week with feeding medium similar to the medium described above but devoid of NuSerum, containing a lower l-glutamine concentration (0.5 mM) and 2% B-27 supplement (Invitrogen).
FU(PSD-95:EGFP)W was assembled from FUGW [45], GW1-PSD95-EGFP-N3 (a generous gift by David Clapham; [44]), and pEGFP-N3 (Clontech Laboratories) in the following manner: PSD-95 was cut out of GW1-PSD95-EGFP-N3 using Nhe1 and HindIII. pEGFP-N3 was linearized using HindIII and SmaI. Linearized pEGFP-N3 and the PSD-95 fragment were ligated, and PSD-95:EGFP was subsequently cut out using XbaI and NheI. FUGW was linearized and the GFP fragment removed using XbaI. The process was completed by the ligation of the remaining FUGW and PSD-95:EGFP.
Lentiviral particles were produced using a mixture of FU(PSD-95:EGFP)W and the Lentiviral packaging vector mix of the ViraPower four-plasmid lentiviral expression system (Invitrogen). HEK293T cells were cotransfected with a mixture of FU(PSD-95:EGFP)W and the three packaging plasmids: pLP1, pLP2, and pLP\VSVG. Transfection was performed in 10-cm plates when the cells had reached 80% confluence, using 3 µg of FU(PSD-95:EGFP)W, 9 µg of the packaging mixture, and 36 µl of Lipofectamine 2000 (Invitrogen). Supernatant was collected after 48 and 72 h, filtered through 0.45-µm filters, aliquoted, and stored at −80°C. Transduction of cortical cultures was performed on day 5 in vitro by adding 5–15 µl of the filtered supernatant to each MEA dish.
The thin-glass MEA dishes used here contained 59 30-µm-diameter electrodes arranged in an 8×8 array, spaced 200 µm apart. The dishes contain 59 electrodes, rather than 64, because the corner electrodes are missing, and one of the remaining leads is connected to a large substrate-embedded electrode designed for use as a reference (ground) electrode. The flat, round (30-µm diameter) electrodes are made of titanium nitride, whereas the tracks and contact pads were made of transparent indium tin oxide (Figure 1A).
Network activity was recorded through a commercial 60-channel headstage/amplifier (Inverted MEA1060; MCS) with a gain of X1024 and frequency limits of 1–5,000 Hz. The amplified signal was multiplexed into 16 channels, amplified by a factor of 10 by a 16-channel amplifier (Alligator Technologies), and then digitized by an A/D board (Microstar Laboratories) at 12 K samples/s per channel. Data acquisition was performed using AlphaMap (Alpha-Omega). Most data were stored as threshold-crossing events with the threshold set to −30 µV. Electrophysiological data were imported to Matlab (MathWorks) and analyzed using custom-written programs.
Scanning fluorescence and brightfield images were acquired using a custom-designed confocal laser scanning microscope based on a Zeiss Aviovert 100 using a 40× 1.3 NA Fluar objective. The system was controlled by software written by one of us (NEZ) and includes provisions for automated, multisite time-lapse microscopy [35]. MEA dishes containing networks of cortical neurons were mounted in the headstage/amplifier that was attached to the microscope's motorized stage. The MEA dish was covered with a custom-designed cap containing inlet and outlet ports for perfusion and air, a reference ground electrode (a circular platinum wire), and a removable transparent glass window. The MEA dish was continuously perfused with feeding medium (described above) at a rate of 5 ml/d by means of a custom-built perfusion system based on an ultra-low-flow peristaltic pump (Instech Laboratories) using an imbalanced set of silicone tubes. The tubes were connected to the dish through the appropriate ports in the custom-designed cap. A mixture of 95% air with 5% CO2 was continuously streamed into the dish at very low rates through a third port, with flow rates regulated by a high-precision flow meter (Gilmont Instruments). The base of the headstage/amplifier and the objective were heated to 38°C and 36°C, respectively, using resistive elements, separate temperature sensors, and controllers, resulting in temperatures of approximately 37°C in the culture medium.
EGFP was excited by using the 488-nm line of an argon laser. Fluorescence emissions were read through a 500–545-nm bandpass filter (Chroma Technology). Time-lapse recordings were usually performed by averaging six frames collected at each of seven to 26 focal planes spaced 0.8–1 µm apart. All data were collected at a resolution of 640×480 pixels, at 12 bits/pixel, with the confocal aperture fully open. To increase experimental throughput, we collected data sequentially from up to 11 predefined sites, using the CLSM robotic XYZ stage to cycle automatically through these sites at 30-min time intervals. Focal drift during the experiment was corrected automatically by using the CLSM autofocus feature [35].
Fura Red labeling of virally transduced cells was performed at the end of several long-term experiments. A total of 1 µl of Fura Red (Invitrogen) from a 2 mM stock solution in DMSO was diluted into 800 µl of culture medium drawn from the preparation dish, subsequently returning the mixture into the dish and mixing gently. After 30 min of incubation, line scan imaging through the neurons somata was performed at a rate of 54 lines/s. At the beginning of each scan cycle, a TTL signal was generated by the microscope and recorded to one of the four free channels of the electrophysiology data acquisition system for temporally aligning imaging and electrophysiological recordings performed during the same period. Fura Red was excited at 488 nm, and the emission was read through a 565-nm long-pass filter (Chroma Technology).
The synaptic identity of PSD-95:GFP puncta was verified by labeling active presynaptic compartments with fluorescent antibodies against the lumenal domain of the synaptic vesicle protein synaptotagmin-1 [46]. These antibodies are taken up by synaptic vesicles whenever they undergo cycles of exocytosis and endocytosis, leading to their accumulation at active synaptic vesicle recycling sites. Labeling was performed as follows: monoclonal anti-synaptotagmin-1 antibodies (Synaptic Systems) were labeled with Alexa-647-tagged Fab fragments using the Zenon antibody labeling kit (Invitrogen). One microgram of the primary antibody was diluted in 18 µl of sterile PBS followed by the addition of 10 µl of IgG labeling reagent. The mixture was incubated in the dark for 5 min, after which, 10 µl of blocking solution was added followed by a second, 5-min incubation. The anti-synaptotagmin-1–Fab mixture was diluted in 500 µl of medium drawn from the MEA dish, and subsequently, the entire volume was returned to the dish and gently mixed. Time-lapse imaging was then continued for 6–12 h. Fluorescent puncta were observed approximately 20–30 min after the mixture application using the 633-nm line of helium-neon laser for excitation and a 640-nm long-pass emission filter (Chroma Technology).
TTX (Alomone Labs) or diazepam (Teva) were diluted in 100 µl of medium drawn from the culture dish. The mixture was subsequently returned to the dish and mixed gently. The final concentration in the dish was 1 µM (TTX) or 2.5/25 µg/ml (diazepam). The addition of TTX to the dish was also followed by the addition of TTX to the perfusion medium.
All imaging data analysis was performed using custom-written software (“OpenView”) written by one of us (NEZ). For the purpose of this project, major portions of the software were rewritten to allow for automated/manual tracking of objects in 3-D time series of confocal images. Boxes of 8×8 pixels were centered on fluorescent puncta, and mean pixel intensities within these boxes were obtained from maximal intensity projections of Z-section stacks. For measuring distributions of puncta intensities (such as those of Figure 3), boxes were placed programmatically at each time step using identical parameters, but no tracking of individual puncta was performed. For tracking identified puncta, all puncta were initially boxed and then a smaller number of puncta (typically 200) were selected randomly and thereafter tracked. Automatic tracking was based on weighted comparisons of vicinity (in X, Y, and Z), intensity, and most importantly, “constellations,” that is, punctum location relative to neighboring puncta within a radius of 50 pixels. The reliability of automatic tracking was reasonable, but not perfect, and therefore, all tracking was verified and, if necessary, corrected manually.
All data were exported to Matlab and analyzed using custom-written algorithms. Images for figures were processed by linear contrast enhancement and low-pass filtering using Adobe Photoshop and prepared for presentation using Microsoft PowerPoint.
Changes in the intensity of a given PSD-95:GFP punctum were modeled as follows: at each time point t, the stepwise change in fluorescence Δf(t) was calculated as a weighted (w, 1 − w) sum of a predictable, activity dependent component Δfp(t) and a random component Δfr(t):
The relative weights of the two components were set to depend on the normalized spike rate S(t) (0 to 1):
The predictable component was calculated as follows:
The random component was determined by randomly choosing a step from a pool of >20,000 measured steps (in TTX, starting 5 h after application; Figure 7A). New synapses were added at fixed rates. The initial size of new synapses was determined by randomly choosing a synapse from a pool of >600 experimentally measured new synapses (Figure 7B). A threshold was set based on the dimmest PSD-95:GFP puncta that could be detected in our system, and synapses whose fluorescence levels fell below this threshold were eliminated. Simulations were seeded using initial values of puncta counts and intensities measured in real experiments. All constants were maintained between all simulations, except c1, c2, and c3.
|
10.1371/journal.pgen.1006094 | Binding of the Fkh1 Forkhead Associated Domain to a Phosphopeptide within the Mph1 DNA Helicase Regulates Mating-Type Switching in Budding Yeast | The Saccharomyces cerevisiae Fkh1 protein has roles in cell-cycle regulated transcription as well as a transcription-independent role in recombination donor preference during mating-type switching. The conserved FHA domain of Fkh1 regulates donor preference by juxtaposing two distant regions on chromosome III to promote their recombination. A model posits that this Fkh1-mediated long-range chromosomal juxtaposition requires an interaction between the FHA domain and a partner protein(s), but to date no relevant partner has been described. In this study, we used structural modeling, 2-hybrid assays, and mutational analyses to show that the predicted phosphothreonine-binding FHA domain of Fkh1 interacted with multiple partner proteins. The Fkh1 FHA domain was important for its role in cell-cycle regulation, but no single interaction partner could account for this role. In contrast, Fkh1’s interaction with the Mph1 DNA repair helicase regulated donor preference during mating-type switching. Using 2-hybrid assays, co-immunoprecipitation, and fluorescence anisotropy, we mapped a discrete peptide within the regulatory Mph1 C-terminus required for this interaction and identified two threonines that were particularly important. In vitro binding experiments indicated that at least one of these threonines had to be phosphorylated for efficient Fkh1 binding. Substitution of these two threonines with alanines (mph1-2TA) specifically abolished the Fkh1-Mph1 interaction in vivo and altered donor preference during mating-type switching to the same degree as mph1Δ. Notably, the mph1-2TA allele maintained other functions of Mph1 in genome stability. Deletion of a second Fkh1-interacting protein encoded by YMR144W also resulted in a change in Fkh1-FHA-dependent donor preference. We have named this gene FDO1 for Forkhead one interacting protein involved in donor preference. We conclude that a phosphothreonine-mediated protein-protein interface between Fkh1-FHA and Mph1 contributes to a specific long-range chromosomal interaction required for mating-type switching, but that Fkh1-FHA must also interact with several other proteins to achieve full functionality in this process.
| Specific chromosomal interactions between distal regions of the genome allow for DNA transactions necessary for normal cell function, but the protein-protein interfaces that regulate such interactions remain largely unknown. The budding yeast Fkh1 protein uses its evolutionarily conserved phosphothreonine-binding FHA domain to regulate a long-range DNA transaction called mating-type switching that allows yeast cells to switch their sexual phenotype. In this study, another conserved nuclear protein, the Mph1 DNA repair helicase, was shown to interact directly with the FHA domain of Fkh1 to regulate mating-type switching. The Fkh1-Mph1 interaction required two phosphorylated threonines on Mph1 that were dispensable for many other Mph1-protein interactions and other Mph1 chromosomal functions. Thus a discrete protein-protein interface between two multifunctional chromosomal proteins helps define a long-range chromosomal interaction important for controlling cell behavior.
| The Saccharomyces cerevisiae Fkh1 (forkhead homolog 1) protein is a member of the FOX (forkhead box) family of proteins defined by their winged-helix DNA binding domains. The FOX family proteins are best known for their transcriptional roles in regulating the cell cycle and differentiation [1]. For example, the Fkh1 paralog, Fkh2, controls the cell-cycle regulated transcription of the CLB2-cluster genes required for the proper execution of M-phase events [2–12]. Fkh1 appears to play an accessory role here, as deletion of both FKH1 and FKH2, but not either gene alone, causes severe cell-cycle dysfunction. However, its molecular functions and the mechanisms by which Fkh1 participates in this process remain poorly understood [3,13]. Accumulating evidence indicates that Fkh1 and 2 also play a transcription-independent role in regulating the timing profile for DNA replication origin activation [14,15]. In addition, Fkh1 has a unique role not shared with Fkh2 in recombination-mediated mating-type switching [16,17], but the molecular mechanisms of this Fkh1 function are not completely understood.
Mating-type switching allows haploid cells of one mating-type to switch to the other, consequently enabling two neighboring haploids to mate and undergo sexual reproduction [18]. Mating-type switching is a critical aspect of yeast biology and evolution that has been used as a model to better understand the repair of double-strand breaks (DSBs) through homologous recombination [19]. During mating-type switching, a DSB is generated by the HO endonuclease at the MAT locus that contains either a- or alpha- mating-type genes. This break is repaired through homologous recombination using donor template sequences located at the silent mating-type loci, HML or HMR, at the opposite ends of the same chromosome as MAT (Fig 1A) [19,20]. HML and HMR contain a repressed copy of alpha (HMLα) or a genes (HMRa), respectively. Productive mating-type switching requires the proper choice between these two donor loci so that the opposite mating-type gene is inserted at MAT. Thus MATa cells favor recombination with HMLα ~90% of the time, while MATα cells choose HMRa as a donor ~90% of the time (Fig 1A). The choice of mating-type donor, that is the directionality of mating-type switching, does not depend on the mating-type genes themselves, but on the protein-DNA complex that forms at a regulatory cis-element called the recombination enhancer (RE), a chromosomal region located between the MAT and HML loci [21]. Fkh1 has been shown to regulate the directionality of mating-type switching by binding to RE in MATa cells and establishing a strong preference for HMLα for repair (Fig 1B) [16]. The forkhead associated (FHA) domain of Fkh1 is sufficient for this function as a LexA-Fkh1-FHA domain fusion is fully functional in regulating donor preference if RE is replaced with LexA binding sites [22].
FHA domains are present in many proteins involved in chromosomal functions and serve as protein-protein interaction modules that specifically recognize phosphorylated threonine residues [24–28]. This property of FHA domains and the involvement of the Fkh1 FHA domain in donor preference during mating-type switching support a model in which the Fkh1 FHA domain controls the directionality of mating-type switching through direct interactions with a phosphorylated protein partner(s) (Fig 1B). This model posits that the presumed partner protein(s) likely binds the DSB at MATa, and through an interaction with Fkh1 bound to RE, localizes HMLα, the donor locus, near the DSB, allowing for efficient strand invasion to occur [22]. Currently, the identities of this Fkh1 partner protein(s) is unknown, and the possible roles of this protein(s), or the Fkh1 FHA domain, in Fkh1’s other cellular roles are also unknown.
To address these issues, we performed a 2-hybrid interaction screen that identified five Fkh1-interacting proteins. Domain analyses revealed that Fkh1 interacted with each of these proteins via its FHA domain. Mutation of key residues within this domain revealed that it was important for Fkh1’s role in cell-cycle regulation, though no single interacting partner could account for this role. In addition, our genetic analyses indicate that functions of the FHA domain outside of its phosphopeptide binding activity contribute to Fkh1’s cell cycle role. Focusing on one Fkh1 binding partner, Mph1, we found that its loss altered donor preference during mating-type switching. Using multiple approaches, we defined a peptide within Mph1 that interacted directly and efficiently with purified Fkh1 in vitro and in a manner that depended on the phosphorylation state of two threonines within the peptide. Mph1 also interacted with Fkh1 in cells and this interaction required the same threonines that mediated the Fkh1-Mph1-peptide interaction. Alanine substitutions of the two threonines in Mph1 (mph1-2TA) caused a defect in donor preference during mating-type switching similar to that caused by mph1Δ. However, mph1-2TA cells did not share other cellular defects caused by mph1Δ, such as sensitivity to MMS or an elevated rate of mutation. Because MPH1 could only partially explain Fkh1-FHA’s role in mating-type switching, we examined the role of a second Fkh1-interacting protein identified in our screen, encoded by YMR144W. A ymr144WΔ also altered mating-type switching directionality, and ymr144WΔ mph1Δ reduced the efficiency of this process beyond that of either mutation alone. We have named this gene FDO1 for Forkhead one interacting protein involved in donor preference. Thus we have delineated a specific cellular role for Fkh1 and Mph1 mediated by an FHA-phosphothreonine interaction, and provided evidence that Fkh1-FHA bound to the RE likely must recognize several proteins at the DSB for full function in mating-type switching directionality.
To identify proteins that interact with Fkh1, we used a 2-hybrid interaction screen in which a Fkh1-Gal4 DNA binding domain (Fkh1-GBD) fusion protein served as bait and a library of Gal4 activation domain (GAD) fusions served as prey [29]. This Fkh1-GBD fusion protein contained the entire Fkh1 coding sequence except for its forkhead DNA binding domain, as this domain was replaced with GBD. Five proteins were identified as positive interactors from this screen (Table 1). These included the DNA helicase Mph1 that is involved in recombinational repair, the Gln3 and Ure2 proteins involved in transcriptional control, and the two uncharacterized proteins with unclear functions [30–34]. Mph1, Ure2, and Fdo1 (formerly Ymr144w) were identified in a previous proteomic screen as proteins that co-purified with a Fkh1-FLAG fusion protein [35], verifying the effectiveness of our screen.
To define how Fkh1 interacts with the proteins identified in our screen, we tested which regions of Fkh1 interacted with Mph1, the yeast homolog of the human FANCM helicase [36,37]. The Fkh1-Mph1 interaction was of particular interest because both proteins are implicated in recombinational repair, though each protein also has other functions [16,19,22,36,37]. Our 2-hybrid screen identified the C-terminal region of Mph1 (amino acids 762–993, henceforth referred to as Mph1-Ct), which has been shown to act as a regulatory domain on this protein, providing interaction sites for numerous proteins that regulate its function [38–41]. To define the region of Fkh1 that interacts with Mph1-Ct, we tested several GBD constructs containing different regions of Fkh1 (Fig 2A) and found that amino acids 50–202 of Fkh1, the majority of which is comprised of the FHA domain, was sufficient for interaction with Mph1-Ct (Fig 2B). Conversely, the fkh1(Δ50–202) mutant did not interact with Mph1-Ct. Thus, the region of Fkh1 containing amino acids 50–202 (henceforth referred to as Fkh1-FHA) was necessary and sufficient to interact with Mph1-Ct.
Next, we examined whether the predicted phosphothreonine binding ability of Fkh1-FHA was required for binding Mph1. To this end, we performed homology modeling (Fig 2C and 2D) using published structures of multiple FHA domains as template (see S1 Fig). Of the homology models generated, the one using the well-characterized N-terminal FHA domain of the checkpoint protein Rad53 [42] as template yielded the highest quality model (S1 Fig). Using this information, as well as additional secondary structure prediction [46] of the regions not modeled, we generated a structure-based sequence alignment of the Fkh1 and Rad53 FHA domains. Upon generation of the homology model and alignment, we found that the FHA domain of Fkh1 is ~50 amino acids larger than previous studies have reported [5,22], as it contains two extra predicted β-strands in addition to the 11 β-strands which comprise the core FHA domain fold [47] (Fig 2C–2E). In addition, this approach allowed for identification of several amino acids predicted to be on or near the phosphopeptide binding surface of Fkh1 (Fig 2D, homology model, and Fig 2E, structure-guided alignment). Five of these residues (Fig 2E, boxed) form the phosphothreonine binding pocket and are conserved among FHA domains [48]. In addition, multiple residues within loops two, three, and four of this domain can make direct contacts with phosphopeptide binding partners in other FHA domains and are less well conserved, allowing different FHA domains to have distinct binding specificities [47]. We note that the predicted phosphopeptide binding surface of Fkh1 FHA is predominantly positively charged, suggesting a preference for binding to a peptide with negatively charged residues (S3 Fig).
Based on this structural and alignment information we engineered several single amino acid substitutions in Fkh1-FHA and assessed their ability to interact with Mph1-Ct in 2-hybrid assays. We found that several amino acids predicted to be on the phosphopeptide binding surface, as well as a more distal residue (S155), were important for interaction with Mph1 (Fig 2D-red, Fig 2E-highlighted yellow and S2A Fig). For example, Fkh1 R80 is conserved in all FHA domains and the analogous residue in Rad53 makes direct contact with its partner peptide [42,48]. Substitution of alanine for Fkh1 R80 abolished the interaction between Fkh1-FHA and Mph1-Ct (Fig 2D and 2E and S2A Fig). In contrast, amino acid substitutions in several amino acids predicted not to be on the phosphopeptide binding interface of Fkh1-FHA had no effect on the Fkh1-FHA-Mph1-Ct 2-hybrid interaction, including substitutions within the extended loop two (Fig 2D-black, Fig 2E-underlined, and S2B Fig). Taken together, these mutagenesis studies suggest that the predicted phosphopeptide-interaction surface of the FHA domain of Fkh1 is important for interaction with Mph1.
To test whether the FHA domain of Fkh1 is also involved in interacting with other proteins recovered from our 2-hybrid screen, we examined their binding to Fkh1-FHA and the mutant constructs described above in the 2-hybrid assay (S2A and S2C Fig). Fkh1-FHA was necessary and sufficient to interact with Ecm30(1005–1183), Gln3(20–189) and Ure2(84–354) (S2A Fig). In addition, with only a few exceptions for assays with Gln3, the amino acid substitutions that abolished Fkh1-FHA-Mph1-Ct binding also abolished the interaction with these other proteins. Finally, a region containing the FHA domain of Fkh1 was necessary but not sufficient to interact with Fdo1, suggesting the involvement of additional regions for their interaction (S2C Fig). Thus Fkh1 can interact with a number of distinct proteins via its conserved FHA domain.
To understand the biological functions of protein interactions observed with the Fkh1 FHA domain, we investigated whether this domain was required for the functions shared between Fkh1 and 2, namely the regulation of the cell cycle and colony morphology. Deletion of both FKH1 and FKH2, but not either gene alone, causes cell-cycle dysfunction that leads to a pseudohyphal-like growth that produces rough, chalky colonies that scar solid agar medium [3–7]. While the FHA domain of Fkh2 is important for FKH2 function [9,10], the role of the Fkh1 FHA domain in FKH1 function in these phenotypes has not been reported. Therefore, we determined whether mutant versions of Fkh1 examined above (referred to as fkh1-m) resulted in these defects in a fkh2Δ background (Fig 3A). We note that all the examined fkh1-m proteins were expressed at levels similar to that of wild type Fkh1 (Fig 3B), indicating that any observed defects are not due to a loss of Fkh1 protein.
By examining spore clones generated from diploids heterozygous for both fkh1-m and fkh2Δ, we first confirmed previous findings that fkh1Δ fkh2Δ and fkh1-dbdΔ fkh2Δ yeast grew slowly and produced a colony that scarred the agar medium (Fig 3C) [3]. We also found that a fkh1 allele lacking the FHA domain coding region (Δ50–202, fkh1-fhaΔ), when combined with fkh2Δ, produced the same phenotype as fkh1Δ and fkh1-dbdΔ (Fig 3C). Thus, this N-terminal region including the Fkh1 FHA domain (residues 50–202) was important for Fkh1’s role in cell cycle regulation. The single residue substitution alleles examined, fkh1-R80A, fkh1-S110A and fkh1-R111A produced smaller colonies when combined with fkh2Δ, indicating that these single amino acids were also essential for wild-type Fkh1 function in this assay (Fig 3C). Each of these residues is predicted to be critical for the phosphopeptide binding function of the Fkh1 FHA domain. The remainder of the fkh1-m alleles examined in this assay caused no discernible defect when combined with fkh2Δ (Fig 3C). However, most of the alleles did reduce mitotic growth rates in liquid culture when combined with fkh2Δ, suggesting a defect in functions that overlap with Fkh2 (Fig 3D). The different effects of fkh1-fhaΔ versus the fkh1-m alleles suggest that Fkh1 residues 50–202 have functions beyond phosphopeptide binding activity in cell cycle regulation. Regardless, most single amino acid substitutions predicted to reduce or abolish FHA phosphopeptide binding activity caused mitotic growth defects, supporting a role for the Fkh1 FHA domain in Fkh1’s overlapping roles with Fkh2 in the yeast cell cycle.
The data presented above supported the hypothesis that Fkh1’s role in cell-cycle regulation is mediated through the Fkh1 FHA domain’s interaction with one or more partner proteins. To test if any of the putative partners defined in the 2-hybrid screen were important for this role, we examined whether deletions of genes encoding these proteins phenocopied a fkh1-fhaΔ or the fkh1-m alleles, such as fkh1-R80A, using the same genetic logic as in Fig 3A. A complete deletion of the protein coding regions for MPH1, ECM30, GLN3, URE2 or FDO1 did not reduce colony size when combined with a fkh2Δ, the diagnostic for Fkh1 function in this assay (Fig 3C). A ure2Δ did slow colony formation after dissection, but this effect did not require a fkh2Δ mutation. Therefore, no single Fkh1 interaction partner identified in the 2-hybrid screen could explain how the FHA domain contributed to Fkh1’s overlapping role with Fkh2 in cell-cycle regulation and morphology.
An important transcription-independent function of Fkh1 lies in the regulation of recombination-mediated mating-type switching [16,22]. Only one Fkh1-interaction partner identified in our 2-hybrid screen, Mph1, has an established role in recombinational repair [36,37,50]. Therefore, we focused on gaining a better molecular understanding of the Fkh1-Mph1 interaction. First, we confirmed this interaction using co-immunoprecipitation. Fkh1 was recovered in an immunoprecipitation with anti-FLAG antibodies only in cells expressing Mph1-FLAG (Fig 4A). Conversely, Mph1-FLAG was recovered in an immunoprecipitation with anti-Fkh1 antibodies only in cells expressing Fkh1 (Fig 4B). We found that this co-immunoprecipitation interaction depended on the region containing the FHA domain of Fkh1 (Fig 4B), validating our 2-hybrid results. In addition, 2-hybrid assays using different GBD-Mph1 fusions showed that amino acids 762–993 of Mph1 were both necessary and sufficient for its interaction with Fkh1-FHA, a result consistent with our finding in the original 2-hybrid screen (Fig 5A). Moreover, a smaller Mph1 fragment composed of amino acids 751–810 was sufficient to interact with Fkh1-FHA, albeit to a weaker extent than Mph1-Ct (amino acids 762–993), while Mph1 lacking this region was unable to bind the Fkh1 FHA domain (Fig 5A).
Previous studies of FHA domains [24,25,48] and the alignment and mutagenesis described in Fig 2 led to the prediction that the Fkh1 FHA domain binds partner proteins through contact with a phosphothreonine residue. To test this idea, we used the 2-hybrid assay to examine if any threonine in Mph1 was required for binding Fkh1. We focused on the overlapping 49 residues between Mph1(751–810) and Mph1(762–993), which contained only two threonines (Fig 5B). Substitution of alanine for both of these threonines (T776AT785A), but not either single T→A substitution, abolished the Mph1-Fkh1 interaction (Fig 5C). This finding was confirmed by co-immunoprecipitation, as Fkh1 failed to pull down mph1-T776AT785A in an immunoprecipitation experiment (Fig 5D). Both assays suggest that the Fkh1-Mph1 interaction required one of two threonines (T776 and T785) within Mph1. These residues are located within a highly acidic region of Mph1. The modeled structure of Fkh1-FHA showed a strongly positively charged concave surface, mainly formed by R80, K107, R111, K112, and R132 (S3 Fig), all of which were required for binding Mph1, suggesting Fkh1 uses this lysine-arginine-rich region to help recognize Mph1 through electrostatic interactions. The Mph1-Ct region serves as a regulatory hub on the Mph1 multifunctional helicase, directing its interactions with several partner proteins, including a subunit of the Smc5/6 complex (Smc5), the large subunit of RPA (Rfa1), and a subunit of the histone fold complex (Mhf2) [38–41]. To determine whether T776 and T785 were involved in these previously reported interactions, 2-hybrid assays were performed with the same series of Mph1 variants examined for interaction with Fkh1. Mph1-T776AT785A was able to interact with all three tested proteins (Fig 5C). Thus T776 and T785 directed a specific interaction between Fkh1 and Mph1 that was distinct from Mph1’s interaction with several other protein partners.
To better establish how Fkh1-FHA interacted with Mph1 we performed 2-hybrid assays in which T776 and/or T785 of Mph1 were replaced with aspartic acid or glutamic acid (Fig 5E). These negatively charged residues can act as phosphomimetics, and thus it was possible that if the role of these two threonine residues were fulfilled via their phosphorylation, that T→D or E substitutions would support the Fkh1-Mph1 2-hybrid interaction via electrostatic contributions alone. However, substitution of these threonines with aspartic acid or glutamic acid, but not the single substitutions, abolished interaction with Fkh1, indicating that T→D or E substitutions were as disruptive to the Fkh1-Mph1 interaction as the T→A substitutions we examined (Fig 5E). These data provide evidence that the threonine residue identities are particularly important, supporting the conclusion that the Fkh1 FHA domain is interacting with this region of Mph1 via classical FHA-phosphothreonine peptide contacts and not merely electrostatic interactions.
Many FHA domains (including the Rad53 N-terminal FHA domain) display a preference for particular amino acids at the pT +3 residue, while other FHA domains have a preference for particular amino acids at other positions [47,52]. As a first step toward understanding the binding preferences of the Fkh1 FHA domain we looked at how substitution of alanine for residues surrounding the two threonines in Mph1 affected Fkh1 binding. We found that substitution of alanine for any of these residues alone did not abolish Fkh1 binding, consistent with the finding that any single T→A substitution (T776A or T785A) did not abolish the Fkh1-Mph1 interaction. However, substitution of alanine for residues surrounding T776 in combination with a T785A substitution did reduce the Fkh1 2-hybrid interaction (Fig 5F). In particular, substitution of alanine for the aspartate at position 774, the serine at position 775, or the glutamate at position 777 in combination with T785A reduced or abolished the Fkh1-Mph1 interaction. Thus the region surrounding T776, including residues D774, S775 and E777, contributed to the Fkh1-Mph1 interaction. We used the same approach to define important residues surrounding T785, analyzing alanine substitutions in combination with T776A (Fig 5G). These data provided evidence that the region surrounding T785, most notably residue E786 but also to a lesser degree residue S782 and E784, contributed to the Fkh1-Mph1 interaction. These data provide additional evidence that this region of Mph1 contains two separate and independent FHA-binding motifs and that both motifs have similar features, including a preference for glutamic acid at the pT+1 position.
Next, we tested whether Fkh1 interacted directly with Mph1 through the region containing T776 and T785 and if this interaction was controlled by phosphorylation of these threonines. To this end, recombinant Fkh1-6xHis was purified from E. coli and its ability to bind an 18-residue peptide representing Mph1(772–789) was assessed by fluorescence anisotropy (Fig 6). The peptide that was phosphorylated on both T776 and T785 bound purified Fkh1 efficiently, with a Kd of 2.2 μM, well within range of other FHA-phosphopeptide interaction affinities [47]. The non-phosphorylated version of the peptide bound Fkh1 with a >100-fold reduced affinity (Kd of 270.8 μM). In addition, and consistent with the effects observed in the 2-hybrid assays in Fig 5C, mono-phosphorylated forms of the peptide (i.e. containing phosphorylation on only T776 or T785) also bound Fkh1, albeit with modestly reduced affinities. These data support the conclusion that Mph1 contained two independent Fkh1-FHA binding motifs, each having a similar affinity for Fkh1.
After establishing that the Fkh1-Mph1 interaction was mediated by the FHA domain of Fkh1 and one of two phosphothreonines on Mph1, we assessed whether this interaction was important for Fkh1’s role in mating-type switching. Fkh1 regulates donor preference during mating-type switching by directly binding to the recombination enhancer (RE) and promoting recombination between an HO-induced DSB at MAT and the donor locus HML. In a previous study, the N-terminal region of Fkh1 containing the FHA domain was shown to be sufficient to direct RE function [22]. This point was elucidated by engineering a strain in which RE was replaced with LexA binding sites and a LexA-Fkh1-FHA fusion protein was expressed [22]. In this Fkh1-dependent assay, the a-mating-type genes located at HMR were replaced by MATα sequences that contained a unique BamHI restriction site (HMRα-B), such that repair of a DSB generated by the HO endonuclease at MATa will always result in a MATα cell, and those using the HMRα-B donor sequence can be cut by BamHI, while those using HMLα cannot. Thus donor preference can be examined by testing the relative abundance of the two different repair products through a PCR reaction that amplifies MATα sequences followed by a BamHI restriction digest (Fig 7A). Consistent with a previous finding [22], HML was the preferred donor, as it was used as template for repair in >90% of cells, while in a strain containing a mutant version of LexA-FHA containing the R80A substitution (LexA-FHA-R80A), recombination between MATa and HML was reduced to less than 20% (Fig 7B). We found that mph1Δ reduced the function of RE, as HML now acted as the donor in <80% of cells (Fig 7B). While this level of reduction was not equivalent to that caused by loss of Fkh1-FHA function, it was highly reproducible. Moreover, mph1-2TA phenocopied the effect of the mph1Δ allele and reduced HML usage to <80%. Additionally, mph1-2TA did not reduce HML preference further in strains expressing LexA-FHA-R80A, providing additional genetic evidence that the Fkh1-Mph1 interaction contributed to donor preference during mating-type switching. The helicase activity of Mph1 is not responsible for this activity, as a helicase defective mutant of MPH1 (mph1-Q603D) did not alter donor preference as drastically as deletion of MPH1 or the mph1-2TA allele, although it did have a statistically small effect. This donor preference defect caused by mph1-2TA was specific to this allele because, unlike mph1Δ cells, mph1-2TA cells did not exhibit sensitivity to MMS (Fig 7C) or an increase in mutation rate (Fig 7D). Thus the mph1-2TA allele caused a specific functional defect in Mph1’s role in regulating RE function while leaving at least two other known roles for Mph1 intact.
The reduction in HML usage in mph1-2TA strains is less than that in cells expressing LexA-FHA-R80A, suggesting there must be other Fkh1 partners required for its role in mating-type switching. To address a role for additional Fkh1-FHA partner proteins, we examined the switching profile in cells lacking Fdo1. We found that deletion of FDO1 reduces HML usage to ~80%, a 10% reduction relative to the wild type control similar to the level of reduction caused by deletion of MPH1 (Fig 8A). Interestingly, in contrast to the Mph1-Fkh1 interaction, the Fkh1 FHA domain was not sufficient for interaction with Fdo1 (S2C Fig). However, further examination of this interaction by 2-hybrid showed that, in the context of full length Fkh1, the fkh1-R80A mutation reduced the Fkh1-Fdo1 interaction, strongly suggesting that the established phosphothreonine binding function of the FHA domain was necessary for the Fkh1-Fdo1 interaction as it was for the Fkh1-Mph1 interaction (Fig 8B). To test whether the defects in donor preference caused by deletions of MPH1 and FDO1 were additive, we also examined mating-type switching in mph1Δ fdo1Δ cells. HML usage was reduced in these cells to a greater degree than in cells containing either single mutation, suggesting that Mph1 and Fdo1 contribute independent Fkh1-FHA binding interactions to control Fkh1-regulated donor preference.
This study provided evidence that Mph1 was a direct Fkh1-FHA phosphoprotein partner relevant to Fkh1’s role in regulating the directionality of mating-type switching. This Fkh1-Mph1 interaction was mediated through a small peptide within the C-terminal regulatory region of Mph1 that contains two threonines each capable of directing interactions with the Fkh1 FHA domain. Mutagenesis studies show that these two threonines likely act as two independent Fkh1-FHA binding motifs, as both threonines must be substituted with alanine to abolish binding by 2-hybrid. Additionally, the amino acid sequences surrounding the two threonines are similar and highly acidic. Both motifs have a glutamic acid residue at the pT+1 position, and mutational analyses indicated that this residue was important for each motif to direct binding of the Fkh1 FHA domain to Mph1. While the 2-hybrid data cannot exclude the possibility that the +1 glutamic acid is required for phosphorylation of the relevant threonine and not directly involved in Fkh1-FHA binding, they nevertheless indicate that a TE signature is relevant to each motif’s independent ability to direct an Mph1-Fkh1-FHA interaction. These observations underscore that there are two redundant Fkh1-FHA binding motifs built into this small region of Mph1. Because a mutant incapable of phosphorylation on these threonines, mph1-2TA, behaved as an mph1Δ in a mating-type switching assay, but not in other commonly used assays that assess MPH1 function, we propose that the Fkh1-Mph1 interaction helps establish the long-range chromosomal interaction essential for donor preference during mating-type switching.
While our data were consistent with the model for Fkh1 bound to the recombination enhancer (RE) guiding the HML locus to the DSB at MAT [22], they also raised an important new question. In particular, why does loss of Fkh1-FHA function cause a much larger defect in RE function compared to mph1-2TA (or mph1Δ), both of which abolish Fkh1-FHA-Mph1 interactions? The simplest explanation is that Mph1 is only one of several proteins bound to the DSB at MAT that the Fkh1 FHA domain uses to locate this lesion. It makes sense for Fkh1 to bind several different proteins at the DSB with relatively weak affinities—in this way the RE remains close to MAT long enough to increase the opportunity for strand invasion into HML. At the same time Fkh1 is not bound so tightly to any one partner or the DSB region itself to inhibit strand invasion and the protein/DNA remodeling necessary to drive the recombination event. Therefore, we propose that there must exist other Fkh1-FHA partner proteins at the HO-induced DSB at MAT that contribute to the RE’s ability to direct the MAT locus to HML. The multi-partner model for Fkh1 FHA function in donor preference may represent a general mechanism by which Fkh1 FHA performs its other biological functions in transcription and replication. This type of mechanism may allow for relatively high specificity but low affinity (and thus potentially highly dynamic) interactions that may be important to these complex chromosomal processes.
Based on this idea and data reported in a previous study, the CK2 kinase likely phosphorylates many Fkh1-interacting proteins involved in donor preference [22]. In this regard we note that, consistent with our observation of an interaction in asynchronous cells and within the 2-hybrid context, CK2 constitutively phosphorylates target proteins [54]. Additionally, the amino acid sequence surrounding both relevant Mph1 threonines are consistent with a CK2 target [54]. When these phosphorylated proteins come together at a DSB, perhaps with other proteins phosphorylated in a more regulated manner by other kinases, they collectively serve to define the DSB for Fkh1-FHA. Consistent with this proposal, a deletion of FDO1, a gene encoding another Fkh1-FHA interaction partner identified in our screen, also reduced donor preference to a degree similar to that of mph1-2TA (or mph1Δ). Moreover, a deletion of both genes to create an fdo1Δ mph1Δ cell reduced preference for HML to a degree greater than deletion of either gene alone. However, a substantial amount of Fkh1-FHA-dependent donor preference remained intact even in cells carrying null mutations in both of these genes, suggesting that another protein or proteins at the DSB must interact with Fkh1-FHA. Many proteins, in addition to Mph1, bind to DSBs and would be good candidates for additional Fkh1-FHA interaction partners that regulate donor preference [55–57]. While mating-type switching is a specific form of homologous recombination, it is clear that DSB repair in diploids also requires a search for homologous regions by the DSB [58]. It will be interesting to learn whether this more generalized process uses similar protein-protein interactions to stabilize chromosomal interactions that serve to juxtapose homologous regions.
Our data provided evidence that the Fkh1 FHA domain may be controlling most, if not all, Fkh1-mediated biology in yeast. Indeed, many fkh1-fha single residue substitution (fkh1-m) mutants abolished interaction with all protein partners uncovered here and reduced Fkh1’s ability to function in cell-cycle regulation with Fkh2, even though deletion of no single gene encoding an interaction partner had an effect. Based on the results with donor preference, it seems likely that multiple different Fkh1-FHA interaction partners will be needed to fully explain Fkh1-FHA’s role in cell cycle regulation. A deletion of the entire FHA domain of Fkh1 (fkh1-fhaΔ) phenocopied a fkh1Δ mutation in cell cycle regulation as measured by both mitotic cell division rates and pseudohyphal-like growth and agar scarring when combined with a fkh2Δ allele. Because the established role of FHA domains is to bind phosphopeptides, it was perhaps unexpected that amino acid substitutions in the FHA domain predicted to abolish FHA-phosphopeptide interactions only slowed mitotic cell division in fkh2Δ cells without causing pseudohyphal-like growth. The Fkh1 FHA domain may play roles in Fkh1 function in addition to phosphopeptide binding by providing as yet undefined interaction surfaces for other regulators of transcription. Alternatively, the fkh1-fhaΔ allele used in this study lacked coding information for an additional ~30 amino acids outside of the alignment-defined FHA domain that may provide surfaces for additional protein-protein interactions. Regardless, these data raise new questions about whether Fkh1’s roles in regulating cell proliferation rate and suppressing pseudohyphal growth are completely separable, or whether a certain threshold of reduced transcription/altered transcriptional regulation must be met before pseudohyphal growth is also observed.
Our data provided evidence that several Fkh1-FHA interaction partners that can direct Fkh1 cellular roles remain unidentified. As we have shown, determining the role of any particular Fkh1-protein interaction is difficult through mutation of Fkh1-FHA itself, as the same FHA residues participate in multiple Fkh1-protein interactions and Fkh1 processes. For this reason, it will be important to identify other Fkh1-FHA-partner proteins and engineer mutations that specifically abolish their ability to interact with Fkh1, as we did for Mph1 in this study, to isolate the discrete mechanisms and pathways influenced by Fkh1.
Strains used in this study were derived from the Saccharomyces cerevisiae strain w303 unless otherwise noted. Standard methods were used for yeast growth, strain and plasmid construction. Strains used in this study are listed in S1 Table. Plasmids are listed in S2 Table. Random mutagenesis of pGBDU-C1 plasmids was performed as described in [59]. Lack of interaction alleles were identified by replica plating from non-selective media to media selective for 2-hybrid interaction and identifying colonies that were no longer viable. Mutants identified by random mutagenesis were confirmed by directed mutagenesis and 2-hybrid assays.
2-hybrid assays were performed in the PJ69-4A strain as described in [29]. The strain contains two reporter genes, HIS3 and ADE2. The original screen was performed using a Fkh1-GBD fusion protein in which the entire DNA binding domain was precisely replaced with the GBD. This GBD-Fkh1 fusion activated transcription of the HIS3 reporter gene. Therefore colonies harboring potential Fkh1-interacting partners were identified on minimal media lacking both histidine and adenine.
A predicted structure for the Fkh1 FHA domain was generated using the N-terminal Rad53 FHA domain as a template using SWISS-MODEL [42–45]. Amino acids 72–170 were modeled. A structure-based sequence alignment of the N-terminal Rad53 FHA domain (Rad53-1) and the Fkh1 FHA domain was generated using a combination of the Rad53 crystal structure (PDB 1G6G) [42] and structural predictions of the Fkh1 FHA domain based on a combination of the homology model and secondary structure predicted using JPred [46]. Electrostatic potential was generated by PyMol v 1.7 [60].
Heterozygous fkh1-m/+ fkh2Δ/+ diploids expressing Fkh1 mutants were dissected and scanned after three days growth. Agar scarring was assessed by gently patching haploid strains onto YPD and washing with H2O after three days. Growth curves were generated by growing to saturation in YPD media, diluting to an OD600 of 0.1 in a 96-well plate, and monitoring growth by measuring the OD600 every three minutes over a 24 hour period in a Biotek Synergy 2 plate reader shaking at 30°C. Doubling times were calculated by exponential regression of data generated from growth curves during log-phase [61].
Cell extracts for western blotting were prepared as described in [62]. Cell extracts for co-immunoprecipitation were prepared by breaking cells by the glass bead method in CoIP buffer (50 mM HEPES pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% TX-100, protease inhibitors (Calbiotech)). Lysates were then diluted 1:1 in CoIP buffer and incubated with the appropriate antibody. Beads were washed with CoIP buffer without detergents followed by washes with the same buffer with 200 mM NaCl.
Co-immunoprecipitation of Fkh1 and FLAG-tagged Mph1 (modified from [63]) were performed using Anti-FLAG antibodies (ANTI-FLAG M2 Affinity Gel, Sigma) or Protein A sepharose-linked anti-Fkh1 antibodies [64]. The starting extract and immunoprecipitated proteins were examined by protein immunoblotting using either anti-FLAG (ANTI-FLAG M2 monoclonal, Sigma) or anti-Fkh1 antibodies. Orc1 detected with an anti-Orc1 antibody [49] served as a loading control.
C-terminally His-tagged full length Fkh1 protein was expressed from a pET28b expression vector in Rosetta E. coli. E. coli were broken with modified B-PER (Thermo Fisher) diluted 1:1 in wash buffer (50 mM Tris pH 7.0, 5 mM MgCl2, 5 mM ATP, 10% glycerol, 1M NaCl, 5 mM BME, 20 mM imidazole, protease inhibitors (Calbiotech)) with 1 mM EDTA. His-tagged Fkh1 protein was purified using nickel chromatography (Qiagen) and eluted in buffer (Wash buffer with 200 mM NaCl, 500 mM imidazole, and without ATP). Peptides (synthesized by the University of Wisconsin-Madison and the Tufts University Core Facility) were labeled on the N-terminus with 5-carboxy fluorescein and an aminohexanoic acid linker. Peptides (constant final concentration of 3 nM) were mixed with titrations of purified Fkh1-6xHis protein in binding buffer (50 mM HEPES pH 7.0, 200 mM KCl, 10% glycerol, 5 mM BME, 1 mM EDTA, 1 mM EGTA, 5 mM MgOAc, 0.02% NP-40, protease inhibitors (Calbiotech)). Polarization at each concentration was measured in triplicates in 384-well polystyrene black microplates (Thermo Fisher Scientific #262260) by a Biotek Synergy H4 multimode plate reader (light source: xenon flash, offset from top: 7 mm, sensitivity: 60%, excitation: 485/20 nm, emission: 528/20 nm, both parallel and perpendicular, normal read speed). Fraction bound (Fb) at each concentration was calculated based on the corresponding polarization values (P): Fbc = (Pc—Pmin) / (Pmax—Pmin), where Pmin is the polarization value of the no-protein control and Pmax is the polarization value of the saturation value for that peptide. Dissociation constants (Kd) were derived by KaleidaGraph (version 4.1.3) using the following equation:
Fb = [protein] / ([protein] + Kd)
Mutation rates were calculated by fluctuation analysis as in [65]. Briefly, single colonies were inoculated into minimal media lacking arginine and grown overnight, diluted 1:10,000 and aliquoted into a 96-well plate. Cells were then incubated, without shaking, at 30°C for 2 days. 24 of the 96 samples were pooled and plated in triplicate to determine the number of viable cells. The remaining 72 samples were spotted onto 10x canavanine plates (minimal media lacking arginine + 0.6 g/L canavanine). Mutation rate was analyzed using FALCOR by the Ma-Sandri-Sarkar maximum likelihood method in which the data are fit to the Luria-Delbrück distribution [53]. For MMS assays, cells were grown to mid-log phase, diluted so that the OD600 is 0.5 and 10-fold serial dilutions were spotted onto YPD plates containing the indicated concentration of MMS. MMS plates were poured fresh on the day of each experiment. Plates were imaged three days after plating.
Donor preference during mating-type switching was determined by a PCR-based method as described in [22]. Briefly, cells were grown in YP-lactate medium to mid-log phase. Expression of the HO endonuclease was induced by addition of 2% galactose and incubated for one hour. Induction was stopped by the addition of 2% glucose and the cells were allowed to recover for 24 hours. DNA was then isolated using quick genomic DNA extraction [66] and PCR was used to amplify MATα sequences using primers Yalpha105F and MAT-dist4R [22]. 700 ng of PCR DNA was then cut with BamHI and the resulting digest was run on an agarose gel. Relative densities of the different bands were determined using ImageJ [67], and donor preference (as HML usage) was calculated using the formula MATα / (MATα+MATα-B).
|
10.1371/journal.pcbi.1004136 | Global Mapping of DNA Conformational Flexibility on Saccharomyces cerevisiae | In this study we provide the first comprehensive map of DNA conformational flexibility in Saccharomyces cerevisiae complete genome. Flexibility plays a key role in DNA supercoiling and DNA/protein binding, regulating DNA transcription, replication or repair. Specific interest in flexibility analysis concerns its relationship with human genome instability. Enrichment in flexible sequences has been detected in unstable regions of human genome defined fragile sites, where genes map and carry frequent deletions and rearrangements in cancer. Flexible sequences have been suggested to be the determinants of fragile gene proneness to breakage; however, their actual role and properties remain elusive. Our in silico analysis carried out genome-wide via the StabFlex algorithm, shows the conserved presence of highly flexible regions in budding yeast genome as well as in genomes of other Saccharomyces sensu stricto species. Flexibile peaks in S. cerevisiae identify 175 ORFs mapping on their 3’UTR, a region affecting mRNA translation, localization and stability. (TA)n repeats of different extension shape the central structure of peaks and co-localize with polyadenylation efficiency element (EE) signals. ORFs with flexible peaks share common features. Transcripts are characterized by decreased half-life: this is considered peculiar of genes involved in regulatory systems with high turnover; consistently, their function affects biological processes such as cell cycle regulation or stress response. Our findings support the functional importance of flexibility peaks, suggesting that the flexible sequence may be derived by an expansion of canonical TAYRTA polyadenylation efficiency element. The flexible (TA)n repeat amplification could be the outcome of an evolutionary neofunctionalization leading to a differential 3’-end processing and expression regulation in genes with peculiar function. Our study provides a new support to the functional role of flexibility in genomes and a strategy for its characterization inside human fragile sites.
| High DNA helix torsional flexibility characterizes sequences which are enriched in fragile sites, loci of peculiar chromosome instability inside human genome often associated with cancer genes. AT-rich flexible islands are suggested to be the determinants of chromosome fragility; however, the origin of their occurrence in cancer genes and the mechanism of chromosome breakage remain unknown. Here, we study DNA flexibility in budding yeast chromosomes. We found that flexibility is conserved in yeast species. Flexibile peaks identify 175 ORFs, mapping on their 3′-end untraslated region. (TA)n repeats of different extension shape the central structure of peaks and co-localize with polyadenylation signals. ORFs with peaks have decreased mRNA stability and prevalent regulatory functions. Our findings support the functional importance of flexibility peaks. They suggest that functional processes may be also at the origin of flexibility peaks presence inside cancer genes in human fragile sites. Definition of role of flexible sequences in genomes may help to understand the processes implied in cancer gene rearrangements.
| DNA conformational flexibility is a function of the dsDNA sequence that defines how the molecule can bend or exhibit a torsion (twist motion) about its axis.
Flexibility is important in DNA supercoiling and shows particular significance in DNA-protein interaction. The relationship of flexibility with the nucleosome occupancy and DNA looping along the genomes determines its key role in many biological functions including the DNA regulation during transcription and replication and DNA repair [1].
The presence of areas of high DNA flexibility at the twist angle has been reported in several unstable regions of human genome, such as fragile sites. Fragile sites are regions peculiarly prone to DNA breakage, usually in conditions of replicational stress; the common fragile sites often map in association with genes involved in tumorigenesis, such as FHIT, WWOX; their instability causes cancer-specific recurrent deletion and translocation breakpoints [2]. While their molecular basis remains elusive, the identification in a number of them of AT-rich flexible islands, capable of forming stable secondary structures has suggested that flexible regions are good candidates for determinants of chromosome fragility [3, 4]. Effects on DNA stability through a structural interference with replication and a block of fork progression have been indicated as possible action mechanisms of flexible sequences [5]. Stalled forks and mitotic entry before replication completion have been indeed shown to be related to chromosome breakage in fragile regions [6]. New results, however, enlighten that also functional aspects are implied in chromosome fragility. Mapping of fragile sites in different cell type confirmed that their setting is tissue dependent and so epigenetically determined [7]. Consistently, fragile sites expressed in human lymphocytes show correlated breakage and are enriched in genes involved in immunity and inflammation, cell-type specific processes [8].
Experimental direct evidence for the role of flexibility in genomic instability has been obtained by using a genetic assay in yeast, where the insertion of a short AT-rich sequence that spans the peak of highest flexibility of the human fragile site FRA16D has been demonstrated to be able to increase chromosome breakage [9]. A support to this model comes from the observation in human genome that AT-rich flexibility peaks also lie at breakpoints of chromosome rearrangements involving the LCR22A-D region of 22q11.2 chromosome, a highly unstable segmental duplication implied in constitutional genomic diseases. [10].
In this paper we approach the problem of biological meaning of DNA helix flexibility by analysing budding yeast chromosome sequences. Yeast has a very compact genome which however comprises a large number of eukaryotic typical genomic elements. A very favourable condition is the large availability of genome-wide data concerning the structural and functional aspects. To this aim, we developed a computer program that predicts the flexibility of the DNA helix by measurements of the twist angle between consecutive base pairs, implementing the TwistFlex software previously developed [11] for the analysis of human fragile sites [3, 12] and its adaptation to fast long sequences analysis.
We present here a high resolution map of twist-angle deviation for the complete genome of Saccharomyces cerevisiae [13]. We determined the presence of 183 flexibility peaks. We defined peaks as segments of genome with twist flexibility above a fixed threshold (i.e. twice the standard deviation). We mapped the location of the flexibility peaks within the yeast genome using the SGD [14] and data reported in literature, both uploaded into the UCSC Genome Browser [15]. Flexibility peaks appear on the 3′UTR of 175 ORFs in S. cerevisiae, which share common features. The connection between flexibility peaks and ORFs could be the evolutionary outcome of modified canonical polyadenylation elements, leading to a differentiated 3′-end processing and gene expression regulation.
The analysis of the first comprehensive map of twist flexibility values reveals the presence of 183 peaks which are 250bp long on average (longest 975bp, shortest 188bp). In the following, peaks shall be denoted by peakIV-16, meaning the 16th peak within chrIV. Their chromosomal map shows no enrichment at specific chromosome arms or at centromere or telomere positions/regions (Fig. 1). The longest chromosomes (chrIV, chrVII, chrXII and chrXV) contain the largest number of peaks, showing a general good correlation between peaks’ distribution and chromosome content (see Table 1 in S1 File). However, peaks do not follow a regular pattern but show regions of intense presence as well as empty regions; the different distances between peaks are reported in Fig. 1 (inset).
The chromosomal map suggests that peaks may be positioned at some specific target sites. First, we compared peaks’ location to ORFs; then, to major genomic annotations. The results, reported S1 Table, show that most of flexibility peaks (170 peaks out of 183, 92.9%) are positioned within interORF regions (Fisher test: p < 10−16). Out of the remaining peaks, 11 lie inside ORFs, one peak lies on a telomere (peakI-2) and one peak lies on a rRNA locus (peakXII-12).
In S. cerevisiae compact genome the interORF regions make up only 27% of the genome length. Of them, 26% are upstream of two divergently transcribed genes and 49% are upstream of one gene and downstream of another, so including putative promoters; finally, 25% are downstream of two convergently transcribed genes, presumably containing only terminators [16].
The inspection of the interORF regions containing flexibility peaks reveals that 67 peaks (39, 4%) lie at interORF regions between converging genes, 77 peaks (45, 3%) lie between genes with unidirectional transcription, only 26 peaks (15, 3%) lie between two genes with divergent transcription (see S1 Table). This is not coherent with 1:2:1 ratio distribution of the yeast genome, making the difference statistically significant for the converging regions (Fisher test: p = 2, 959 × 10−5) as well as for the diverging ones (Fisher test: p = 2.201 × 10−3).
The distribution and position of genes along chromosomes are basic genomic features known to play a role in the regulation of gene transcription and translation; this is of particular importance in yeast compact genome due to its dense arrangement of genes and short intragenic regions. For example, genes that are divergently expressed may share promoter and transcription factors and show similar regulation and functional relationship; similarly, convergent genes may share terminators or 3′-transcribed regions [17]. In this context, the observed prevalent position of flexibility peaks suggests that they could represent structural regulatory signals.
We take advantage of measurement of promoter, 5′UTR, 3′UTR and terminator regions of a large number of yeast genes reported by Tuller et al. [17] to analyze the possible co-localization of any of these regions with flexibility peaks. According to the cited authors, promoters and terminators were considered the sequences intermediate between the different untraslated regions; for only a few ORFs without measure data, the average length of 5′UTR and 3′UTR were reported. We found that all peaks lying between convergent genes, except 4 peaks, co-localize with the 3′UTR of one ORF or of both ORFs, as in the cases of very large peak extension or 3′UTR partial overlap (Fisher test: p < 10−15). Peaks lying between genes with unidirectional transcription co-localize with 3′UTR in 64 cases (Fisher test: p < 10−15). To sum up, peaks on a 3′UTR region are 127 and ORFs with a peak in 3′UTR are 175. Finally, peaks between divergent genes co-localize with 5′UTR in 18 cases (Fisher test: p < 10−15). Peaks’ features are reported on S1 Table.
The presence of shared sequences inside peak sequences was searched by a ClustalW2 alignment analysis, that however give no significant results. Differently, a Repeat Masker analysis revealed that all peaks were characterized by (TA)n or similar AT-rich repeats (Fig. 2).
(TA)n repeats show a predominant presence and characterize all peak types except the 11 peaks lying inside ORFs, all of which contain (TTA)n. Repeats show a great length variability and comprise stretches of uninterrupted dinucleotide TA sequences mixed with degenerated TA sequences (from 23 to 89bp). For this reason, in the following we shall refer to all types of AT-rich sequences as to tandem repeats, indifferently.
3′UTR is a regulatory region; in yeast several distinct but interacting elements compose the 3′-end forming signals: the polyadenylation efficiency element (EE), the positioning element (PE) and the near-upstream/near-downstream elements (w.r.t. cleavage site). EE is the upstream signal including mainly TATATA (consensus sequence: TAYRTA). PE occurs 16 to 27nt downstream and the best word for this element is AATAAA (consensus sequence: AAWAAA); however, it is commonly described only as A-rich, since many functional sequences are characterized only by their adenosine content. The near-upstream element, as well as the near-downstream, is characterized as T-rich [18].
The EE promotes the recruitment of other polyadenylation factors by binding, upon transcription of RNA, the trans-acting factor Hrp1, that also plays important roles in mRNA export, mRNA surveillance and nonsense mediated decay. The TAYRTA sequence provides the greatest effect on 3′-end processing with the T/U at the first and fifth positions being the most critical for function; on a large-scale analysis (1017 yeast nuclear transcripts) more than half of 3′UTR (52%) contained this optimal EE sequence [19]; in more cases, transcripts contain several consecutive copies of EE sequence [20]. Owing to these reported TA-rich EE structures, we searched evidence for a general relationship between the tandem repeats (corresponding to flexibility peaks) and EE elements.
In literature, the sequence for the 3′-end of the GAL7 or MRP2 genes have been made available [20] and authors mapped in detail major poly(A) sites and expanded EE elements (TA)8. We found that the EE elements co-localize with an under-threshold flexible region (i.e. a genomic region where flexibility is enhanced, but does not reach the peak threshold). Similar results have been obtained for the expanded EE element detected within the 3′UTR of FBP1 gene, constituted by a (TA)14 repeat [21], again co-localizing with an under-threshold flexible region; this last element is of special interest because it has been experimentally shown to be a very potent polyadenylation element in both strand orientations. The expanded EE has been suggested [22] to affect polyadenylation offering several overlapping binding sites to Hrp1 or allowing its association/disassociation at multiple binding sites. Thus, we speculated that all the flexibility peaks that are positioned at 3′UTR might have the potential to serve as EEs, with an expansion linked to functional features, where the determinant for complex 3′-end formation could be just the DNA/RNA secondary structure due to helix flexibility.
Ozsolak et al. [23] have obtained very informative data in a map of poly(A) cleavage sites in yeast genome generated by a direct RNA sequencing.
For each poly(A) intense cleavage site (i.e. scored at least 945 by authors of [23]), we calculated the distance from midpoint of repeats in nearest peak. There are 2874 intense sites (out of 34444) which are closer than 500nt from a repeat within a peak. As shown by Fig. 3, intense poly(A) sites occur in a highly position-specific manner, prevalently within a distance range of 5nt to 25nt from repeats: 91.7% of them are closer than 100nt and 73.8% are closer than 25nt. If we limit this analysis only to (TA)n, then 75% are closer than 25nt. Poly(A) intense cleavage sites usually are present as multiple and clustered elements inside range [0-25nt] from repeats. Almost all peaks in convergent and unidirectional intergenic regions match to intense poly(A) signals. The authors of [23] read weak and isolated signals as indicative of a low transcriptional activity; this occurs only in nine peaks, so it is nearly negligble.
Moreover, we inserted on UCSC Genome Browser the position of characterized positioning elements (PE, whose consensus sequence is AAWAAA) and of efficiency elements (EE, whose consensus sequence is TAYRTA), defined for both strands through the Yeast Genome Pattern Matching [24]. The analysis of repeats position and of strand direction of signals highlights a peculiar organization of 3′UTR extremity or of its extension. In unidirectional intergenic regions, the repeat sequence covers the extremity of mapped 3′UTR or lies slightly outside it, bordering the downstream poly(A) signals; the EE element is found in multiple copies, all overlapping the repeat sequences. The PE element, when present, may be positioned either upstream the EE (within the 3′UTR), or downstream the complete 3′-end forming signal, as well as in both positions within the same 3′UTR. Examples include the 3′-ends of genes IME1 (peakX-5), DBF4 (peakIV-14) or CDC53 (peakIV-5) (see supporting S1 file, figure 1).
In the convergent intergenic regions, ORFs often overlap their 3′UTR; here, the repeat sequence and the concomitant EE element may lie either inside only one or inside both 3′UTRs, thus bordering poly(A) signals on both sides; the repeat/EE sequence represents a central element from which the poly(A) reads depart in divergent direction, forming a complex overlapping polyadenylation signal. Examples are the peculiar 3′-ends of the convergent gene pairs TSR1 and RAD59 (peakIV-9, see Fig. 4), as well as ERV15 and AME1 (peakII-10), SNC1 and MYO4 (peakI-1), or DIG2 and PHO8 (peakIV-27) (see supporting S1 file, figure 2).
Interestingly, also in most divergent intergenic regions we found very clear poly(A) signals inserted into to the typical organization repeat/EE/poly(A) previously described for 3′-ends; due to lack of 3′-ends in these regions, this is unexpected. Sometimes the 3′-end signals lie on 5′UTR with sense or antisense orientation as respect to the adjacent ORF, as it happens for the region within the divergent PUF3 and YEH1 genes (peakXII-3); in other cases signals are distant from ORFs without any overlap with its components, as for region of peakX-3 within the divergent TDH2 and MET3 genes (see supporting S1 file, figure 3). These findings clearly indicate the presence of termination signals in absence of annotated transcriptional units; therefore, peaks which are positioned at 3′UTR may also mark non coding RNA genes, that frequently may be antisense transcripts. A large quantity of antisense transcripts has been reported by both Ozsolak and Nagalakshmi studies [23, 25] and they are estimated to cover in yeast the 80% of annotated ORFs. Antisense transcripts are in lower amount and so are characterized by a low number of 3′-end signals; this motivates the presence of weak signals in peaks which are not positioned at 3′-end of ORFs.
Finally, concerning peaks lying inside an ORF, we remark that we found poly(AAT) codons coding for poly-Asn region of polypeptide—instead of poly(A) signals.
On conclusion, TAYRTA elements, closely adjacent to cleavage site, have a non-canonical position in the peak-associated 3′UTRs. To explore the concomitant occurrence of further polyadenylation elements we performed a search for motifs by a MEME analysis [26], carried out on 183 peak regions. We identified, as expected, a TATATATATATATATATGTATAT motif (MEME statistical significance E-value = 4.6 × 10−585) in 145 peaks and a ATTATTATTATTATTATTATTATTATT motif (MEME statistical significance E-value = 3.7 × 10−119) in 32 of them. Moreover, performing an analogous analysis on flexible regions±100 (i.e. peak regions, comprehensive of additional 100nt upstream and downstream), we found that in 183 sites the novel A/T-rich motif CTTCTTTTCTTC (MEME statistical significance E-value = 1.8 × 10−12) was found (see summarizing Fig. 5). This last motif seems to have some function since it again occurs in all interORF peak regions.
Overlapping 3′UTRs are common in many genomes for genes orientated in a tail-to-tail manner. They have been described in yeast, where they may depend on the dense arrangement of genes and possibly to cause transcriptional interference [27]. It is credible that, similarly, for unidirectional genes, failure to terminate transcription at the end of first gene will result in inhibition of the next gene [28] and that this interference type could act as a regulatory system for the differential expression of adjacent gene pairs or for the sense-antisense transcription [29]. This suggests that the flexible elements inside 3′UTR could characterize genes with specific types of termination, where peculiar signals are required possibly to regulate a programmed RNA interference.
Following the rationale that functional elements show a relative evolutionary conservation, we determined the conservation rate of flexible sequences in four other sequenced Saccharomyces sensu stricto species (S. bayanus, S. paradoxus, S. mikatae, S. kudriavzevii). For this analysis a dataset by Scannell et al. [30, 31], containing the alignment of 4298 intergenic regions, was analysed. Out of the 170 flexible sequences (excluding those inside ORFs), 131 regions (77%) conserve a flexibility peak exceeding the fixed threshold in at least one species and 70 regions (41%) in all species; in most cases of conservation failure, under-threshold flexible regions were observed. Conservation of peaks is particularly strong for the convergent and unidirectional intergenic regions. Out of the 67 convergent ones, 55 regions (82, 1%) conserve the flexibility peak in at least one species and precisely 53 in S. paradoxus, 52 in S. mikatae, 50 in S. kudriavzevii and 49 in S. bayanus (see S2 Table). Consistently, 51 out of the 55 conserved flexible sequences are in regions with conserved synteny maintaining convergent transcription. The unidirectional regions conserving a flexibility peak in at least one species are 67 (81, 8%), all maintaining unidirectional transcription. Differently, the peak conservation in divergent intergenic regions is significantly under-represented (50%; Fisher test: p = 0.002).
Of interest, the sequence alignments may show that conservation of peaks does not derive from the identity of intergenic sequence but is frequently consequent to a different organization of a high number of tandem repeats, as visible in the alignments of intergenic regions of peakIV-14 -unidirectional intergenic region between DBF4 and DET1- and peakIV-9 -convergent intergenic region between RAD59 and TSR1 (see supporting S1 file, figure 4 and figure 5). These findings are indicative of an evolutive differentiation among species with a substantial conservation of flexibility peaks, even when there is a weak sequence conservation among the four genomes. Notably, 38 conserved flexibile ORFs (22 in converging and 11 in unidirectional transcription) were found to belong to the list of ohnologs i.e. paralogous genes arising from whole genome duplication [32] (see S2 Table); in all cases, except one, only one member of ohnolog pair carries a flexibility peak in 3′UTR. Usually, the pair members of ohnologs underwent sequence modifications related to functional changes of different extent. Consequently, the peak sequence on one onholog may be a peculiar modification linked to functional divergence between pair members, possibly leading to sub- or neo-functionalization, which are processes already defined in yeast for a number of duplicated genes [33].
The gene order arrangement has an evolutionary meaning [34]. In yeast, for instance, adjacent genes are co-expressed to a significantly higher level than expected [35]; moreover, many highly co-expressed gene pairs take part in the same cellular processes [36]. Accordingly, the conservation of flexibility peaks in convergent or unidirectional pattern may be related to the peculiar structural or functional aspects of gene pairs expression.
The 3′UTR regulates mRNA levels or stability via RNA-protein interactions with mRNA degradation machinery. mRNA stability is a key regulatory step controlling gene expression and ultimately affects protein levels and function. Notably, long- and short-lived transcripts appear to have systematic differences in the EE, suggesting peculiar roles of this poly(A) signal in mRNA stability [37]. Therefore we checked whether the ORFs with peak in 3′UTR could be related with a differential mRNA stability. We took advantage of data about mRNA half-lives derived by Wang et al. [38] coming from mRNA decay profiles measured by microarrays following transcriptional shut-off. Results were searched for the 175 ORFs with peak in 3′UTR compared with all other ORFs; they show that these ORFs are characterized by significant lowering of both poly(A) half-life (t-test: p < 2.5 × 10−2) and overall half-life (t-test: p < 1 × 10−2), indicating their production of unstable mRNAs (see Fig. 6). According to current models for major decay pathways, in yeast poly(A) shortening precedes the decay of the entire transcript and is a rate-limiting step [39]. Differential degradation of mRNAs can play an important role in setting the basal level of mRNA expression and how that mRNA level is modulated by environmental stimuli. It has been suggested that there is a general relationship between the stability of an mRNA and the physiological function of its product. Accordingly, mRNAs involved in central metabolic functions are generally relatively long-lived, whereas those involved in regulatory systems turn over relatively rapidly [38]. Consistently, flexibility peaks inside 3′UTR may be proposed to be part of the regulatory machinery of short-lived mRNAs.
The prevalent occurrence of unstable transcripts for ORFs with peak in 3′UTR has obvious implications for their possible regulatory roles within specific pathways. A functional analysis of all such 175 ORFs (listed in S3 Table) was carried out by identifying the Gene Ontology (GO) terms, using the YeastMine search engine [40]. The search reveals enrichment for 72 GO Biological Process (p < 1.1 × 10−2) as well as for 14 GO Molecular Function categories (p < 2.6 × 10−2), as reported in S3 Table. The first 10 GO BP terms (i.e. with lowest p-value) are identified for a range of 31 to 86 ORFs per GO term, with a mean value of 62.3 ORFs per GO term. The GO MF term “binding” is identified for 101 ORFs.
Many GO terms concerned correlated processes or functions; so, they were processed by the web server REVIGO [41], using the default settings, in order to reduce their redundancy and summarize them in representative subsets the GO lists. The outcomes for Biological Process GO terms (visualized as treemap in supporting S1 file, figure 6, top) point out the presence of ORFs with role in cell cycle, phosphorus/organic cyclic compound/ nitrogen compound metabolism, phosphorylation reproduction, growth, response to acid, signaling. The 175 ORFs include genes expressing key components of cell cycle progression and regulation: TUB2 and TUB3 encoding α and β tubulins, CLB4 and PHO80 encoding cyclins, CDC53 and APC9 encoding respectively the cullin structural protein of SCF complexes and a subunit of the Anaphase-Promoting Complex/Cyclosome; moreover, AME1, RAD24, RAD59 and SWE1 involved in checkpoint maintenance, the FUS3, DIG2 and SLT2 encoding MAP-kinases and their regulator BMH1 encoding the major isoform of 14-3-3 proteins. Further IME1, encoding a master regulator of meiosis and its convergent gene UME6, the key transcriptional regulator of early meiotic genes; moreover MFA1, encoding the essential mating pheromone a-factor, STE50 the major protein involved in mating response. Finally, ASG1, TSR1, ICT1, YAP1, PHO80, FRT1 and HAA1, regulators involved in the stress response. In accordance with the prevalent regulatory functions revealed for Biological Process GO terms, the REVIGO outcomes for Molecular Function GO terms point out the presence of numerous ORFs with role in binding and in phosphatase and kinase activities (visualized as treemap in supporting S1 file, figure 6, bottom).
All these findings confirm the general involvement of ORFs with peak in 3′UTR in regulatory systems as well as their characterization by unstable transcripts. Moreover, these results seem to be coherent with the picture where regulatory function of genes is related to short half-life [38].
In budding yeast, the ability of genes to respond to environmental changes has been related to nucleosome occupancy in 5′-ends and 3′- ends [42, 43]. Nucleosome free regions or nucleosome depleted regions (NFR or NDR) were observed at regulatory regions such as gene TSS and TTS, affecting binding of regulatory proteins, nucleosome ordering inside genes and transcriptional plasticity [44, 45]. Since AT-rich sequences in defined contests have nucleosome-disfavoring property, we evaluated whether the AT-rich sequence in flexible peaks in 3′UTR could play a regulatory role by determining specific nucleosome positioning; thus, we analyzed the co-localization of peaks with NDR, obtained from [46]. We found that large distances occur between each peak and nearest segment with high nucleosome depletion (Fig. 7 in supporting S1 file), indicating that AT-rich peak regions and NDR are not associated elements. A manual inspection was then performed on nucleosome occupancy of all peaks localized in 3′UTR of convergent genes, to be sure to consider only transcriptional terminators. Data on experimental nucleosome occupancy, reported by [47], together with nucleosome coverage predicted by a model based on in vitro sequence data, were available through the SwissRegulon server [48, 49]. We found that no peak shows altered nucleosome coverage. These are unexpected results, as many papers describe nucleosome depletion in yeast gene 3′-end termination. Anyway, they contribute to circumstantiate the flexibility peak’s action, by suggesting that flexible peak may exert exert its function on polyadenylation by affecting phases not directly dependent on local chromatin structure, for example by modulating the nascent mRNA structure.
Considering the gene function of peak associated ORFs, it is of interest that 14 of such ORFs have human orthologs involved in Mendelian diseases, detected from the Database of Human Disease Orthologs [50]; among these are the YPL164C gene, whose human ortholog gene MLH3 encodes the DNA Mismatch Repair Protein Mlh32 associated to HNPCC or Hereditary nonpolyposis colorectal cancer, the YOL071W (SDH5) gene whose human ortholog SDHAF2 (alias PGL2) is associated to familial paragangliomas 2 and the YPL204W gene whose human ortholog CSNK1A1 is associated to familial adenomatous polyposis. A complete description of the human ortholog genes related to diseases is reported in S4 Table, including, besides genes, related diseases and detailed references, the chromosome band localization and the coincidental occurrence of common fragile sites. We highlight that the map position of the human ortholog genes for eleven yeast genes is coincidental with that of known fragile sites [51]; moreover most of orthologs are implied in cancer development. These findings support the relationship between peak associated ORFs and fragile sites.
We remark also the presence of NIT3 among flexible yeast genes, a gene encoding one of two proteins that in S. cerevisiae have similarity to the mouse and human Nit protein, interacting with the human Fhit tumor suppressor. Indeed, the FHIT gene spans FRA3B, the most common human fragile site characterized for the presence of clusters of high flexibility peaks [52]. The FHIT gene has been suggested to have biological effects similar to NIT and to share with it signaling pathways [53].
In this paper we sistematically study the presence of flexibility peaks in S. cerevisiae genome and explore their functional role.
Peaks show a strong co-localization with tandem repeats inside the 3′UTR region of a number of ORFs and in particular with clusters of poly(A) signals. The peculiar architecture of repeats and poly(A) signals inside peaks suggests that they could mark terminations in ORFs characterized by specific requirements in RNA cleavage. Consistently, we characterize the peak presence in ORFs as prevalently lying in regions where convergent transcription occurs. Peaks show a general conservation among different Saccharomyces yeast species, but with a sequence variation in orthologous genes and a clear differentiation between paralogous genes, suggesting that they could be the result of an evolutive differentiation. We provide evidence that ORFs with peak in 3′UTR have transcripts with lower half-life, item considered peculiar of genes involved in regulatory systems with high turnover. More, we show that ORFs with peak in 3′UTR share a number of common functions in biological processes such as cell cycle regulation or stress response. From these findings we infer that flexibility peaks could play a functional role as regulatory elements of gene expression for a peculiar set of genes. A regulation based on flexible sequences has not so far experimental foundation. However, we must consider that, while the impact of 3′-end sequence on gene expression is well established, the understanding of how its effect is encoded in DNA is limited. Polyadenylation is critical for many aspects of mRNA metabolism, including mRNA sytability, translation and transport. PolyA signals act as substrate for cleavage and polyadenylation, for which RNA structure is also a critical determinant [54]. Then, RNA binding proteins regulate almost all post-transcriptional stages [55]. Specific sequence motifs in 3′UTR have been identified in yeast implied in stabilization [56] and stress response [57]. In particular, an increased AT-content upstream the polyadenylation site has been shown to modulate protein expression dynamics [58]. Thus, AT rich tandem repeats and strand flexibility may be crucial in determining the interaction with polyadenylation factors, the mRNA structure and the accessibility of binding sites to multiple regulators. The notion that enriched tandem repeats in S. cerevisiae could guide transcriptional modulation has been established for genes carrying very variable tracts of repeats in promoter; the involved genes have the general feature of interacting with the cell environment and so requiring rapid response changes [59, 60]. Gene regulation differs greatly among related species, constituting a major source of phenotypic diversity. This issue assumes relevant significance for gene evolution and tandem repeats have been considered able to drive transcriptional divergence and to confer evolvability to gene expression [61]. The variable repeat-based component of peaks inside 3′UTR may have similar origin and evolution. Tandem repeats are intrinsically prone to variation having often units lost or gained by replication slippage [62]: Thus, long repeat stretches could be derived from the well-known polyadenylation enhancement elements; their potential in modulating gene expression regulation (termination efficiency and transcript half-life) may have been the feature that determined their fixation in peculiar genes.
These findings on yeast genome may be relevant for the knowledge of the relationship between flexibility peaks and human genome instability. Common fragile sites are chromosome regions prone to breakage upon replication stress. To date, 22 fragile sites, among the 230 mapped in human lymphocytes, are known at molecular level but the molecular basis of fragility remains unknown. They extend over megabase-long regions, tend to overlap very large genes and share a delayed completion of DNA replication. Recently, delayed replication has been correlated with a paucity of initiation events [63, 64]. Notably, the authors found that FRA3B and FRA16D, the most active fragile sites in human lymphocytes, have low levels of fragility in fibroblasts, where instead other sites show very high fragility; cell-type-specific replication programs characterize the commitment to fragility at different loci in each cell-type, indicating that fragility is epigenetically defined.
These findings are consistent with the view that fragile sites serve a function; this is supported by a number of indirect but relevant observations, the first of which is the conservation of fragile sites in synteny regions in the mouse and human genomes in all cases analyzed so far. The second one is their enrichment in genes related to cell cycle regulation, apoptosis or similar processes involved in cancer development [65]. More in detail, chromosomal fragile sites FRA3B and FRA16D, carrying the FHIT and WWOX genes respectively, that are genes playing a major role in apoptosis, show correlated expression and association with failure of apoptosis in lymphocytes from cancer patients [66]. In the same perspective, all fragile sites belong to networks of correlated breakage, comparable to gene expression pathways activated in response to damage stress; in particular the correlated fragile sites, analyzed in lymphocytes, are enriched in genes involved in immunity and inflammation, that are cell-type specific processes of lymphocytes [8].
Coherently with the above described functional aspects, flexibility peaks in yeast occur in ORFs involved in cell cycle control or stress response, where flexible sequences seemed to play a regulatory role in gene expression. While yeast is a unicellular and quite simple organism, many processes are highly conserved; it is conceivable that conservation may concern the specific mechanisms that differentiate the expression of peculiar gene classes. In higher eukaryote evolution, these mechanisms may have been used in the commitment of the different genes to stress response, that is cell and tissue specific [67].
In this view, the regulatory role of flexibility peaks inferred for yeast genes could be actual also for human fragile genes, even if not necessarily involving 3′-end termination process. The extent of this correlation will be determined by a comparable genome-wide analysis on human sequence DNA flexibility.
We refer to complete Saccharomyces cerevisiae RefSeq genome as obtained and annotated on SGD (SacCer2 assembly).
Experiments on conformational analyses of DNA require large numbers of conformations to be sampled. The conformation of DNA and its sequence dependence are mainly determined by the chemical structures of the base pairs and their interactions. The computational model by Sarai et al [68] examines DNA flexibility on the basis of base pairs interactions and the results agree with available experimental observations. The algorithm StabFlex is used to calculate potential local variations in the DNA structure that are expressed as fluctuations in the twist angle (degrees, deg). It is a reimplementation of the TwistFlex software [11] and it is targeted to analyze very large sequences.
The calculation of twist fluctuations is made for overlapping windows along a given sequence (window length L = 100bp, window shift s = 1bp). Within each window the flexibility is calculated for consecutive dinucleotide steps, and the average value of all steps in the window is assigned to the midpoint dinucleotide step. The flexibility is measured in degrees (deg) in the range [7 deg;16 deg].
An example of the output data is given in Fig. 4. Peaks emerge spontaneously as short genomic regions where signal is extremely high. They are marked by arrows in the top picture. The complete flexibility data for a genomic region are plotted as a quantized signal and each flexibility value refers to 100bp, as shown in the bottom zoomed snapshot.
Fig. 7 (top picture) shows the normalized distribution of windows flexibility values for all 16 chromosomes of yeast genome. As shown in Fig. 7 (bottom picture), for large flexibility values (greater than 12deg) the distribution is no longer Gaussian. The non-Gaussian tail identifies flexibility peaks, as follows. First, we pre-selected regions with outstanding flexibility values, deviating significantly from the average (not lower than 𝓢 = mean+2×stand dev, which is 12.1 for all chromosomes). That value 12.1 may be read as the point where Gaussianity is lost (see inplot in Fig. 7). Regions correspond to the genomic sequence covered by overlapping consecutive windows simultaneously exceeding 𝓢. Second, such regions whose maximal flexibility value exceeds threshold θ = 13.8 are defined flexibility peaks. The threshold has been fixed as in literature [12, 52].
Peaks have been denoted by peakIV-16, meaning the 16th peak within chrIV.
The statistical significance of properties and classifications has been assessed by means of Fisher’s exact test and t-test. Fisher’s exact test is used in the analysis of 2 × 2 contingency tables built for categorical data that result from classifying objects in two different ways; it is used to examine the significance of the association (contingency) between the two kinds of classification. A t-test is a statistical hypothesis test in which the test statistic follows a Student’s t distribution if the null hypothesis is supported. It can be used to determine if two sets of data are significantly different from each other, and is most commonly applied when the test statistic would follow a normal distribution if the value of a scaling term in the test statistic were known. For both tests, specific R programs have been designed and implemented by the authors.
Differently, when external classifications have been used, statistical significance has been imported with the results. This applies to motifs found by MEME and to GO terms’ enrichment. As stated by the authors in [26], MEME usually finds the most statistically significant (low E-value) motifs first. The E-value of a motif is based on its log likelihood ratio, width, sites, the background letter frequencies, and the size of the training set. The E-value is an estimate of the expected number of motifs with the given log likelihood ratio, and with the same width and site count, that one would find in a similarly sized set of random sequences.
Concerning GO terms, as stated in [69], there are a number of different tools that provide enrichment capabilities. Tools differ in the algorithms they use, and the statistical tests they perform. All enrichment widgets list a term, a count and an associated p-value. The term can be something like a publication name or a GO term. The count is the number of times that term appears for objects in your list. The p-value is the probability that result occurs by chance, thus a lower p-value indicates greater enrichment without corrections. The p-value is calculated using the Hypergeometric distribution.
A data repository for deviations of twist angle for complete yeast genome may be found in [13]. Individual chromosomal flexibility peaks’ annotations in BED format, suitable for a visualisation through the Genome Browser [15] are part of online supplementary material. The algorithm StabFlex is available at http://home.gna.org/stabflex/.
|
10.1371/journal.pntd.0005150 | Epidemiologic Correlates of Mortality among Symptomatic Visceral Leishmaniasis Cases: Findings from Situation Assessment in High Endemic Foci in India | Visceral leishmaniasis (VL) is highly prevalent in the Indian state of Bihar and, without proper diagnosis and treatment, is associated with high fatality. However, lack of efficient reporting mechanism had been an impediment in estimating the burden of mortality and its antecedents among symptomatic VL cases. The objectives of the current study were to generate a reliable estimate of symptomatic VL caseload and mortality in Bihar, as well as to identify the epidemiologic and health infrastructure-related predictors of VL mortality.
Using an elaborate index case tracing method, we attempted to locate all symptomatic VL patients in eight districts of Bihar. Interviews and medical-record-reviews were conducted with cases (or next-of-kin for the dead) meeting the eligibility criteria. The information collected during the interviews included socio-demographic characteristics, onset of disease symptoms, place of diagnosis, pre- and post-diagnosis treatment history, type and duration of drugs received. In total, we analyzed data on 4925 VL patients—59% were male and 68% were less than 30 years old. There were 158 (3.2%) deaths and the incidence rate of mortality was 3.2/100 person-years. In the adjusted Cox-proportional-hazards analysis, treatment at public facility [Adjusted Hazard Ratio (AHR) = 0.61; 95% CI = 0.43–0.86], shorter (≤30 days) diagnostic delay [AHR = 0.62, 95% CI = 0.43–0.92], and treatment completion [AHR = 0.03, 95% CI = 0.02–0.05] emerged as significant negative predictors of mortality.
Mortality reduction efforts in Bihar should focus on improving access to early diagnosis, quality treatment and treatment-adherence measures, with special emphasis on marginalized communities.
| More than 70% cases of visceral leishmaniasis (VL), a parasitic disease transmitted by sand flies, in India are reported from Bihar, a resource-poor state. In absence of early diagnosis and treatment, the fatality is very high among symptomatic VL cases. However, community-based data on VL mortality is limited in Bihar. Also, little information on the factors associated with mortality among symptomatic VL patients is available from this Indian state. Evidences regarding mortality parameters can prove immensely helpful in designing specific interventions targeted at reduction of such mortality. In the current study, we created a registry of symptomatic VL cases in eight districts of Bihar to estimate the burden of VL mortality and the socio-demographic factors associated with it. The mortality rate among the symptomatic VL patients was found to be fairly high. Shorter interval between onset of symptoms and diagnosis, and completing course of treatment were associated with increased survival. We recommend improving access to early diagnosis and treatment services, especially among the marginalized communities, as a key measure to reduce mortality among symptomatic VL patients in Bihar.
| Globally, an estimated 500,000 new cases of visceral leishmaniasis (VL) or kala-azar, a ‘Neglected Tropical Disease’, occur annually [1]. More than 90% of the global burden of visceral leishmaniasis (VL) is contributed by six countries: Bangladesh, Brazil, Ethiopia, India, South Sudan and Sudan [2]. The parasite Leishmania donovani is the main causative agent of VL in India, Nepal, and Bangladesh, where it is transmitted by the sand fly vector Phlebotomus argentipes [3]. In Indian subcontinent, as per the prevailing knowledge of transmission dynamics, humans are considered to be the sole carrier of the disease (anthroponotic) and transmission pattern is believed to be peridomestic i.e. transmitted by the bite of female sand flies in and around the home [4, 5].
Based on recent estimates, in India, 70%–90% VL cases occur in a single state–Bihar [6, 7]. The disease was considered to be at the brink of elimination in Bihar during 1960s following widespread DDT use under National Malaria Eradication Program [8, 9]. However, in absence of a dedicated VL elimination strategy and surveillance system, the disease saw a resurgence a decade later [9, 10]. Since then VL has been widely acknowledged as a major public health concern in India. Still a significant proportion of VL caseload in Bihar was assumed to remain outside the public health detection system until recent past [11, 12]. Absence of a robust surveillance system as well as simple and inexpensive diagnostic tests were identified as potential reasons for the poor tracking of VL cases in Indian subcontinent [11, 13].
In order for an effective kala-azar elimination program in Bihar, identification of high endemic foci is essential. Health Management Information System (HMIS), the current source of population level information on VL, has been reported to suffer from various methodological and operational problems including underreporting and ill-defined population frame [11, 12]. For precise identification of blocks having VL incidence above the elimination level (i.e. annual incidence of 1/10000 population or more), a situation assessment project was undertaken in January’2013 [7]. It was also expected that this assessment would help in evaluating the effectiveness of various elimination strategies (e.g. treatment by amphotericin-B or miltefosine) and inform the kala-azar elimination program in India on the incidence of VL associated mortality and its epidemiologic correlates.
In absence of early diagnosis and treatment, fatality was detected to be very high among symptomatic VL cases [14, 15]. Even among treatment-recipients, the measured probability of dying within two years of the onset of symptoms were quite high [7, 16]. Further, among many diagnosed patients, outcome remained unknown due to lack of systematic follow up and cause of death remained obscured among many suspected cases in the community. Thus, alike incidence, it can be hypothesized that mortality due to VL also remain grossly under-reported. To the best of our knowledge, no large community-based study did yet report on either the incidence of mortality or its antecedents among symptomatic VL cases in India. The current exercise, therefore, was undertaken to address this knowledge gap. Our objectives were to provide a reliable estimate of symptomatic VL case load and mortality in selected districts of Bihar, and to identify the epidemiologic and treatment-related predictors of VL mortality.
In 2013, CARE India undertook a VL situation assessment project in eight (out of 38) districts (representing both high and low endemic foci) of Bihar. The principal objective of this project was to inform the kala-azar elimination program operations in this resource-poor state. An important component of this undertaking was to estimate the burden of symptomatic VL cases in the selected districts–through detection of such cases and by tracking them to their households. The reference period (period during which VL diagnosis took place) for the assessment was between January 2012 and June 2013. The following combination of methods was followed to meet this objective:
The following eligibility criteria were used to determine if a potential subject, identified through the above methodology, could be considered a case of VL:
Study investigators conducted face-to-face interviews with every VL cases who were identified by the above methodology and met the eligibility criteria. In case a patient died in the interim, interview was conducted with the next-of-kin. The information collected during the interviews included socio-demographic characteristics, onset of symptoms, place of diagnosis, pre- and post-diagnosis treatment history, type and duration of drugs received etc. If available, information on death was obtained from medical documents, else we relied on the information provided by interviewee. The date of onset of symptoms was based on recall. Local calendars and list of local festivals were used by the study investigators to facilitate participants’ recall and to estimate the approximate dates. Respondents who did not possess any treatment-related documents were asked about the treatment regimen. At the time of conduction of this study, three anti-leishmanial drugs were available in the study area–SSG, amphotericin B, and miltefosine. If intake of oral drugs was reported, then it was assumed that the patient had received miltefosine. In case of treatment with injectables, depending on the reported regimen, either SSG or amphotericin B was decided. The breakdown of frequencies regarding source of drug information is presented in supporting information S1 Table. Informed verbal consent was obtained from all adult respondents before collecting the information. If the respondent was a minor, then informed consent was obtained from the respective parent/caregiver.
CS Pro 5.0 was used for data entry, assessments of logical consistencies and detection of duplicate entries. Descriptive analyses were carried out to determine the distribution of socio-demographic and disease-related characteristics of the study population. Bivariate statistical tests (Pearson chi-square) were performed to check for differences in the characteristics of dead and surviving VL patients.
Kaplan-Meier survival curves were constructed and long-rank tests were performed to compare the survival probabilities (from the day of VL diagnosis) across the categories of various demographic and treatment related factors. For identification of the factors associated with deaths among VL cases, unadjusted and adjusted hazard ratios (HR) were obtained by employing Cox proportional-hazards (CPH) models. To account for subject-wise variabilities in the interval between the dates of VL diagnosis and interview, we used a censoring method similar to that employed for cohort studies (deaths were considered as failure events and surviving cases were right censored on the date of interview). In the unadjusted analyses, we assessed the association between death and type of treatment facility (public or private) place of diagnosis (public or private), interval between onset of symptoms and VL diagnosis (≤30 days and >30 days), type of drug used for treatment (amphotericin-B, miltefosine or SSG), and completion of treatment course. The hazard ratios for each of the five predictors, adjusted for patient’s age, gender, caste and socio-economic status, were assessed using five separate CPH models. According to caste, patients were categorized into–marginalized (scheduled castes and scheduled tribes), semi-marginalized (other backward castes) and general/other caste. Regarding socio-economic status, we summarized the information of three variables–type of house, ownership of cattle and belonging to ‘mahadalit’ caste (poorest of poor)–using principal component analysis. We used the derived principal components as surrogate measure for socio-economic status. The proportional hazards assumption was tested using Schoenfeld residuals method. Time dependent covariates were created from the independent variables used in CPH models. For each CPH model, it was assessed if these time dependent covariates had non-zero slopes in a generalized linear regression of the scaled Schoenfeld residuals on functions of time. SAS version 9.4 was used for statistical analyses.
The current study was approved by the Institutional Committee for Ethics and Review of Health Management Research Office of Indian Institute of Health Management Research (IIHMR), Jaipur, India (www.iihmr.org). Informed consent (including signature or left thumb impression of the respondent) was obtained from each agreeing participant before interview and measurements, after explaining the details of the study in a language that they could understand.
Through the index case tracing method, 5770 listed VL cases were identified from the public facilities. Among them, 4962 cases (or their next-of-kin for those who were dead) were successfully traced to their residence. Additionally, records of 1119 cases meeting the study eligibility criteria were obtained from private facilities and through snowballing technique. Following removal of duplicate listings, total 5432 eligible cases were thus recruited. Further, for the current analysis, we did not include i) cases with post-kala-azar dermal leishmaniasis, ii) cases with ambiguous mortality data, and iii) cases for which reported onset of symptoms occurred more than two years before date of diagnosis (because of reliability of recall). Thus, another 507 cases were excluded from the final dataset. Therefore, information on 4925 VL cases who were diagnosed between January 2012 and June 2013 was used for the current exercise. The distribution of cases identified through different sources and at various stages of the study is presented in supporting information S2 Table.
The mean age of the VL patients was 24.9 years (SD ±17.6 years) and about 59% of them were males. More than 80% belonged to either marginalized or semi-marginalized castes. About two-thirds (67%) of the eligible respondents lived in mud/thatched (Kaccha) houses. Majority of the cases not only got tested for VL at a public laboratory (63%) but also received treatment exclusively from public facilities (78%). About 5% patients did not complete their course of treatment, with proportion of treatment defaulter being higher among those prescribed SSG (8%) compared to miltefosine (4%) or amphotericin-B (4%). About 3% VL patients died during the interval between diagnosis and case tracing. The incidence rate of death among symptomatic VL patients was 3.2/100 person-years. The socio-demographic and disease-related characteristics of VL patients are depicted in Table 1.
Bivariate comparison of the characteristics between the dead and the alive (till the date of interview) VL patients revealed that there were significant differences in age (higher proportion of older patients among the dead) and caste (higher proportion of marginalized caste among the dead) distribution between the dead and the alive (Table 1). Moreover, patients who completed the course of their therapy, received earlier diagnosis of VL and those who received treatment from the public facilities were less likely to die. Regarding Kaplan-Meier estimates of the probability of survival in the period following diagnosis (Fig 1), there were statistically significant differences (log rank tests) in cumulative survival probabilities across the categories of age, caste, treatment facility, diagnosis interval and treatment completion.
Table 2 depicts the hazard ratios (and 95% confidence intervals) of death among VL patients associated with various predictors. Those who received treatment exclusively from public facilities were found to have lower risks of death compared to those treated (completely or partially) at private facilities [Unadjusted hazard ratio (HR) = 0.52; 95% confidence interval (CI) = 0.37, 0.72; Adjusted: HR = 0.61; 95% CI = 0.43, 0.86]. Getting an early diagnosis (within 30 days of onset of symptoms) was associated with lower hazard of death, compared to those diagnosed after 30 days (Unadjusted: HR = 0.67; 95% CI = 0.48–0.93; Adjusted: HR = 0.66; 95% CI = 0.47–0.92). There were no significant differences in terms of the drugs used for VL treatment. Irrespective of the drug, patients who completed treatment course had 97% lower risk of death (Unadjusted: HR = 0.03; 95% CI = 0.02, 0.04; Adjusted: HR = 0.03; 95% CI = 0.02, 0.05), compared to those who did not complete treatment. Tests for proportional hazards assumption (Schoenfeld residuals test), revealed that none of the CPH models had a significant p-value for the test of non-zero slope. Thus, the proportional hazards assumption could not be rejected for any of the CPH models.
In the current article, we present the findings of the VL situation assessment initiative aimed at estimating the burden and mortalities associated with symptomatic VL in the Indian state of Bihar. Previous estimates of VL incidence in Bihar, which has the highest burden of VL in India, were based either on the data from public facilities or surveys conducted in limited geographical areas with high endemic foci [11, 12]. Moreover, there existed very few studies on VL-associated mortalities in India. From the perspective of India’s kala-azar elimination program, which was principally focused on prevention, understanding the correlates of VL-associated deaths could help in designing interventions targeted at reduction of VL mortalities and prioritizing deliveries of essential health services to VL patients.
The case fatality of symptomatic VL was found to be around 3%, lower than the value reported by a large scale VL mortality study conducted in Bangladesh [13]. The comparatively higher proportion of death reported in the Bangladesh study could be attributed to the large tribal ethnic population included in that study, with about 83% VL deaths occurring among tribals [13]. Moreover, unlike Bihar, about one-third of the death cases in Bangladesh did not receive any treatment. Thus, the lower mortality seen in Bihar, compared to Bangladesh, might be a consequence of wider access to modern treatment for VL.
The proportion of males were higher among VL cases. This is similar to the findings from previous studies in the same geographical region and could be attributed to their sleeping habits (sleeping outside the house) and occupational exposures (e.g. farming) [6]. Also, families of about 61% symptomatic VL patients were owners of cattle. This was not surprising given that cattle-sheds are known breeding site for the vector in Bihar [17]. For majority of patients, the interval between the onset of symptoms and diagnosis was more than 30 days. An earlier study conducted in Bihar and neighboring Nepal also reported on the issue of diagnostic delay in Bihar [18]. However, unlike the said study, most of the VL patients included in the current research sought treatment from Government-run facilities. Minimizing the delay in diagnosis and treatment, besides being prognostically beneficial for individual patients, is crucial for restricting the disease transmission as the infected hosts serve as source of leishmania amastigotes for sandflies. About 8% patients were prescribed SSG, despite the widely reported resistance against it [19]. Moreover, the likelihood of non-completion of treatment was much higher among those prescribed SSG, possibly due to adverse effects or prolonged parenteral therapy or both [20].
Probability of death was higher among people from marginalized castes corroborating the observation in Bangladesh [13]. As marginalized people in India belong mostly to lower socio-economic strata, the observed higher mortality could be an outcome of their low awareness levels and/or poorer access to healthcare [17, 21]. The impact of economic status on mortality was also evident from the fact that ownership of cattle, which is a surrogate for better economic status in rural Bihar, was found to be associated with lower mortality risk. This is not surprising as economic status is a principal determinant of access to healthcare and, in turn, disease outcome. Moreover, ownership of cattle might have been associated with milk consumption and better nutritional status, which could have potentially lowered the risk of mortality [22, 23]. Further, treatment at public facilities was found to be associated with lower risk of death compared to getting treated by private healthcare providers. In rural Bihar’s scenario, unqualified practitioners are often the mainstay of health services. Despite the fact that health services provided at public health facilities are free and there are concerns about the quality of care provided by the rural medical practitioners, residents of hard-to-reach areas often have to rely on such practitioners. Poor access to quality healthcare among VL patients in India and other resource-limited settings have long been recognized as a major concern [24]. Therefore, future policies aimed at reducing VL mortalities in India should address issues related to proper healthcare access among residents of rural areas, especially those belonging to the marginalized communities.
Proportional hazards analysis revealed that diagnosis within 30 days of symptom onset and completion of treatment course to be significant predictors of survival among symptomatic VL cases. Earlier diagnosis is likely to be associated with earlier treatment as well–which is a widely recognized determinant of survival[25]. Treatment completion, irrespective of the prescribed drug, with 96%-97% reduced risk of mortality, emerged as a strong predictor of survival. Treatment adherence, besides being essential for cure, is also important for preventing emergence of drug resistance. Therefore, targeted interventions aimed at strengthening the adherence improvement measures, such as rigorous counselling and treatment follow-up—especially for potential defaulters, need to be emphasized upon.
There were some important limitations of the present undertaking. First, the current study included only the symptomatic VL cases. However, the burden of asymptomatic VL has been reported to be much larger compared to the symptomatic cases [26]. Nevertheless, as mortality due to VL is expected to occur mostly among symptomatic cases, the mortality estimate (and its predictors) is presumed to be somewhat accurate. Second, some of the dates used in the analysis (especially date of onset of illness), an essential component of survival analysis, were approximate and based on recall. Therefore, the possibility of some amount of information bias in the estimates should be borne in mind. Third, because of unavailability of past treatment documents, it was not possible for us to determine whether a VL case was new or relapse and if there were any associated comorbidities (e.g. HIV). Another potential limitation of our findings is that our analyses could not account for all possible factors that might have contributed to mortality among VL patients (nutritional status, comorbidities etc.), as the required information were not collected. Finally, it was not possible to ascertain the exact causes of death for all cases (where no medical documents were available). However, such missing information were likely to be non-differentially distributed across the categories of the variables incorporated in the statistical model(s). Because of the aforementioned limitations, caution should be exercised while interpreting the results of our analyses (and extrapolating the findings to the national level).
Notwithstanding the limitations, the findings of this first-ever attempt at identifying the epidemiologic and health-system related correlates of VL-associated mortalities in Bihar probably could generate important insights. It was revealed that the interventions aimed at reducing mortalities should specifically target the marginalized communities and focus on improving access to early diagnostic and quality treatment services along with strengthening the treatment adherence measures. We expect our findings to contribute towards the efforts of reducing VL mortalities and help in epidemiologic transition of VL to a non-fatal low-endemic disease in Bihar.
|
10.1371/journal.pgen.1001170 | Dual Functions of ASCIZ in the DNA Base Damage Response and Pulmonary Organogenesis | Zn2+-finger proteins comprise one of the largest protein superfamilies with diverse biological functions. The ATM substrate Chk2-interacting Zn2+-finger protein (ASCIZ; also known as ATMIN and ZNF822) was originally linked to functions in the DNA base damage response and has also been proposed to be an essential cofactor of the ATM kinase. Here we show that absence of ASCIZ leads to p53-independent late-embryonic lethality in mice. Asciz-deficient primary fibroblasts exhibit increased sensitivity to DNA base damaging agents MMS and H2O2, but Asciz deletion or knock-down does not affect ATM levels and activation in mouse, chicken, or human cells. Unexpectedly, Asciz-deficient embryos also exhibit severe respiratory tract defects with complete pulmonary agenesis and severe tracheal atresia. Nkx2.1-expressing respiratory precursors are still specified in the absence of ASCIZ, but fail to segregate properly within the ventral foregut, and as a consequence lung buds never form and separation of the trachea from the oesophagus stalls early. Comparison of phenotypes suggests that ASCIZ functions between Wnt2-2b/ß-catenin and FGF10/FGF-receptor 2b signaling pathways in the mesodermal/endodermal crosstalk regulating early respiratory development. We also find that ASCIZ can activate expression of reporter genes via its SQ/TQ-cluster domain in vitro, suggesting that it may exert its developmental functions as a transcription factor. Altogether, the data indicate that, in addition to its role in the DNA base damage response, ASCIZ has separate developmental functions as an essential regulator of respiratory organogenesis.
| ASCIZ is a DNA damage response protein that has been proposed to be a regulator and stabilizing co-factor of the ATM kinase, mutations of which lead to a syndrome involving neurological and immune dysfunctions, tumour predisposition, and X-ray hypersensitivity. To study Asciz function in vivo, we have generated a knockout mouse model lacking this gene. Here we show that ASCIZ has a specific role in mediating cell survival in response to DNA base damage, but it is not required for stabilization and regulation of ATM. Strikingly, Asciz knockout mice fail to survive to birth and have tissue-specific defects in embryonic development. In particular, Asciz null embryos fail to develop lungs and undergo an early arrest in tracheal development. The precursor cells that normally form the lung are present in our embryos, but they fail to segregate from the foregut. These observations indicate that ASCIZ plays an important and previously unrecognized developmental role that is most likely unrelated to its function in mediating responses to DNA damage. Our study delineates the function of ASCIZ in DNA damage survival and highlights an exciting new function of the protein in controlling the early stages of lung development.
| Pathways that maintain genome integrity by responding to spontaneous DNA damage are crucial for normal development and ageing, and act as tumor suppressors to prevent the onset of cancer [1]. While DNA damage signaling is rather generic in that structurally diverse lesions eventually lead to activation of one or both of the central checkpoint kinases ATM or ATR [2], DNA repair pathways are believed to be highly lesion-specific [1]. In addition to environmental DNA damage, eukaryotic cells incur a high level of spontaneous DNA damage as a consequence of normal metabolism, most notably abasic sites that are generated as repair intermediates of the base excision repair (BER) pathway with an estimated incidence of ∼10,000 per cell per day [3]. Abasic sites can emanate from various base modifications (e.g. oxidation, methylation), which for experimental purposes are most commonly generated by treatment with methylmethane sulfonate (MMS) or H2O2, and key BER enzymes for their repair include apyrimidinic/apurinic endonuclease (APE) and DNA polymerase beta (Polß) [4], [5].
The importance of the BER pathway is indicated by findings that absence of any of the BER genes acting downstream of abasic sites results in embryonic or perinatal lethality in mice [6], [7]. However, some of the key BER enzymes also seem to have DNA damage-independent functions; for example, APE1 has a separate role as a redox regulator of several transcription factors [8]. Similarly, increased apoptotic cell death during development of Polß-null mice can be suppressed by deletion of p53 (TRP53), indicating that this part of the phenotype is indeed due to defective base damage repair. On the other hand, the perinatal lethality of these mice that is associated with defective neuronal and lung development as a DNA damage-independent defect is not rescued by p53 deletion [9]–[11].
While the DNA damage processing enzymes of the BER pathway are clearly defined, new accessory factors that regulate the activity or stability of Polß and other BER enzymes keep emerging, including the non-histone DNA-binding protein HMGB1 [12], arginine methyl-transferases [13], and ubiquitin ligases [14]. We recently identified ASCIZ ( = ATM substrate Chk2-interacting Zn2+-finger) as a new Zn2+-finger (ZnF) protein with roles in the DNA base damage response. In human cells, ASCIZ forms DNA damage-induced nuclear foci specifically in response to DNA damaging agents that generate lesions repaired by the BER pathway (MMS and H2O2) in a manner that is enhanced by the BER inhibitor methoxyamine, and Asciz depletion by siRNA leads to increased MMS sensitivity [15]. Likewise, Asciz deletion in the chicken DT40 B lymphocyte line leads to increased sensitivity to MMS and H2O2, but not to ionizing radiation (IR), UV irradiation and other DNA lesions, as well as increased erroneous repair of enzyme-generated DNA base damage consistent with a role in the BER pathway [16]. Moreover, Asciz deletion suppresses the dramatic MMS hypersensitivity of Polß-deficient DT40 cells [16], reminiscent of the protective effect of simultaneous deletion of the relevant upstream methyl-purine-glycosylase (MPG) in Polß-deficient murine embryonic fibroblasts [17]. ASCIZ contains a large number of conserved ATM/ATR kinase phosphorylation sites in an SQ/TQ cluster domain [18], and consistent with its original classification as an ATM substrate, ASCIZ was subsequently re-isolated as an ATM-interacting protein (thus also called ATMIN)[19]. It was proposed that ASCIZ acts as an essential co-factor of ATM that was required for ATM stability (and vice versa ATM was required for ASCIZ stability) as well as for ATM activation by some stimuli, though surprisingly not by canonical DNA damaging ATM activators such as IR [19].
To better understand the role of ASCIZ in vivo, we have here generated a mouse line that lacks the vast majority of the Asciz protein-coding sequence in the germline. Our results confirm that Asciz-deficient cells are specifically hypersensitive to DNA lesions that are processed by the BER pathway, but challenge the proposed interdependence between ASCIZ and ATM levels. In contrast to Atm-deficient mice that overall develop normally [20], Asciz deletion results in late embryonic lethality with severe respiratory defects reminiscent of mouse mutants in Wnt2/2b and FGF10 signaling pathways. The data indicate that Asciz has an unexpected DNA damage-independent developmental function as an essential regulator of pulmonary organogenesis.
Human and mouse Asciz have a similar gene structure where exons A–C encode the N-terminal ZnF region of about 220 amino acid residues, and exon D encodes the bulk of the protein (601 of 823 or 818 residues) including the nuclear localization signal, core domain and SQ/TQ cluster domain (Figure 1A, 1B; NCBI Gene ID 23300). Because there is evidence for expression of alternative isoforms that differ in the number of N-terminal ZnFs (http://www.uniprot.org/uniprot/O43313), we integrated loxP sites flanking exon D into the murine Asciz locus to remove the majority of the protein-coding sequence (Figure 1B). Germline deletion of this exon after crossing with PGK-Cre knock-in mice, followed by outcrossing of PGK-Cre (all on a pure C57BL/6 background), was confirmed by Southern blot and PCR genotyping (Figure 1C). In over 600 offspring from Asciz+/− heterozygote intercrosses genotyped at weaning (∼3 weeks of age), we failed to detect any homozygous Asciz-deleted mice (Figure 1C and Table 1). However, homozygous Asciz-deleted embryos were readily detectable even at relatively late stages of gestation (Figure 1D; and more detail below).
Western blotting of head extracts confirmed the absence of ASCIZ protein in Asciz−/− embryos, and a ∼50% reduction of protein levels in heterozygotes compared to wildtype (WT) littermates (Figure 1E). Levels of other DNA damage response proteins (including ATM) appeared to be normal in Asciz-deficient embryos (Figure 1E and below). In Northern blots using a probe for the non-deleted exon C, the residual exon D-deleted Asciz transcript was present in homozygous targeted embryos at <15% of wildtype (WT) mRNA levels (Figure S1), indicating that the mutated mRNA is highly unstable. Using Asciz null embryo lysates as an antibody specificity control, we found that ASCIZ is ubiquitously expressed in adult mice, with overall similar levels relative to the loading control in all tissues except for somewhat higher levels in the brain, cerebellum and testes (Figure 1F).
The absence of homozygous Asciz−/− mice at weaning prompted us to investigate the development of ASCIZ-deficient embryos in more detail. Asciz−/− embryos were recovered at near-Mendelian ratios at all time points analysed (Table 1). Based on peripheral circulation scored during uterine dissections, Asciz−/− embryos appeared to lose viability around embryonic day 16.5 post conception (E16.5) (Table 1), at which point they were considerably growth-retarded compared to littermates (Figure 2A, 2B). Embryonic lethality due to DNA damage response gene deletions can often be suppressed by p53 deletion [6]. To test if p53 status affects the essential requirement for Asciz, we intercrossed compound Asciz+/−/p53+/− heterozygous mice. However, we could again not detect any viable Asciz−/− mice amongst >300 genotyped offspring at weaning (Table 2). Altogether, these data indicate that absence of Asciz leads to progressively impaired development during late gestation and becomes absolutely incompatible with life a few days before term.
To monitor DNA damage sensitivity of primary Asciz-deficient cells, we isolated murine embryonic fibroblasts (MEFs) from viable Asciz−/− embryos and matched WT littermate controls between E12.5–E14.5 (i.e., before growth retardation was apparent). Standardized proliferation assays using a 3T3 protocol under normoxic conditions (20% O2) revealed a modest premature senescence phenotype of Asciz-deficient MEFs compared to WT controls, with growth arrest after ∼20% fewer population doublings (Figure 3A). When normalized to the maximum population doublings within each litter, Asciz−/− MEFs always senesced earlier than the matched WT cultures (Figure 3B). As senescence of MEFs under these conditions is believed to involve an oxygen-induced DNA damage response [21], these results indicated a role of ASCIZ in the response to oxidative base damage in primary cells. To corroborate this, we treated early-passage MEFs (P2–P3, when proliferation differences between genotypes were minimal) with a panel of DNA damaging agents. In these assays, Asciz-deficient MEFs were significantly more sensitive to MMS and H2O2, which cause damage that is predominantly repaired by the BER pathway, compared to matched WT littermate controls (Figure 3C, 3D), but they were not hypersensitive to agents such as UV or hydroxyurea (HU) whose damage is repaired by other pathways (Figure 3E, 3F). MMS hypersensitivity of Asciz−/− MEFs was less pronounced than that of WT cells co-treated with methoxyamine (Figure S2), which blocks the single-nucleotide BER pathway through partial inhibition of APE1 and Polß [5]. In addition, MMS hypersensitivity of ASCIZ-deficient cells was further enhanced by methoxyamine (Figure S2), indicating that absence of ASCIZ only partially impairs BER. Altogether, these results are consistent with a role of Asciz as an accessory factor in the BER pathway in primary cells.
Because we had originally identified ASCIZ based on its interaction with Chk2 [15], and because ASCIZ was later proposed to be sometimes required for ATM activation [19], we tested if the MMS hypersensitivity of Asciz−/− MEFs could be due to defective ATM signaling. However, there was no reduction in MMS-dependent ATM activation (detected by an antibody against mouse pS1987-ATM; human pS1981-ATM) and phosphorylation of key ATM/ATR targets γH2AX and pS18-p53 in Asciz-deficient MEFs compared to WT littermate controls (Figure S3A, S3B). MMS-induced p53-S18 phosphorylation was completely abolished by the highly specific synthetic ATM kinase inhibitor KU55933 [22] (Figure S3C), indicating that it is a genuinely ATM-dependent process, whereas H2AX phosphorylation under these conditions was ATM-independent and thus likely ATR-mediated. Because our antibody for Chk2 phosphorylation on T68, widely considered to be one of the most ATM-specific phosphorylation sites, did not detect this site in MEF extracts (data not shown), we monitored Chk2-T68 phosphorylation in human U2OS cells following Asciz depletion by siRNA treatment. However, Chk2 was still efficiently phosphorylated in response to MMS (as well as IR) in Asciz-depleted cells (Figure S4A, S4B). Altogether, these data indicate that the increased MMS sensitivity of Asciz-deficient cells is not caused by impaired ATM signaling.
During analyses of DNA damage signaling in Asciz-depleted human cells (Figure S4B and data not shown) and initial protein blots of embryo extracts (Figure 1E), we could not detect any meaningful reduction of ATM levels in Asciz-deficient cells. Because this contradicted the recent report that ASCIZ and ATM levels were mutually dependent on each other [19], we explored this discrepancy first in additional embryos. Again, we saw no reduction of ATM protein levels in any of the Asciz−/− or heterozygous samples compared to WT littermate controls (Figure 4A, left panel, and data not shown). Likewise, we also did not see a reduction of ATM levels in human cells after almost complete depletion of ASCIZ by siRNA-treatment (Figure 4B, left panel, compare control siRNA treated to si-ASCIZ treated U2OS cells, lanes 1 and 2; and data not shown), or in two independently generated ASCIZ knockout clones [16] in the chicken DT40 B cell line (Figure 4C, left panel).
We also revisited the proposal that ATM was in turn required for ASCIZ stability [19]. In contrast to the severe reduction of ASCIZ levels in an ataxia telangiectsia (AT) fibroblast line reported by Kanu and Behrens (2007), we did not detect any loss of ASCIZ in another human AT patient-derived fibroblast cell line (Figure 4B, left panel, AT2221JE) that is considered to be bona fide ATM-deficient [23] compared to control fibroblasts (Figure 4B, left panel, GM847) or the isogenic AT cell line reconstituted with WT Atm (Figure 4B, left panel, AT2221JE+ATM [23]). We expanded this analysis to a panel of seven independent human AT-patient derived lymphoblastoid cell lines [24]. When adjusted for loading (actin), there was no reduction of ASCIZ levels in four of these lines that contained no or extremely low levels of residual ATM protein (AT1ABR, L3, AT4ABR, AT5ABR) compared to two independent healthy donor control cell lines (C3ABR, C35ABR)(Figure 4B, right panel); ASCIZ levels were also unaffected in two further AT lines that contained intermediate ATM protein levels (AT1ABR, AT33ABR), and there was only a modest reduction of ASCIZ levels in a third AT cell line with intermediate ATM levels (AT32ABR). Similarly, there was also no reduction of ASCIZ protein levels in brain lysates of Atm null mice [20] compared to WT littermates (Figure 4A, right panel), or in an Atm-deleted chicken DT40 clone [25] compared to the WT control (Figure 4C, right panel). Taken together, these data demonstrate that ASCIZ and ATM are not required for each other's stability in three different vertebrate species.
Because of the overall relatively mild DNA damage phenotypes of Asciz-deficient primary cells (Figure 3) and the absence of a p53-effect on viability (Table 2), we wondered whether the underlying cause for the late gestational lethality of Asciz−/− embryos could be DNA damage-independent, and performed histological analyses of litters between E12.5 and E18.5. The most striking defect at all time points was the complete absence of lungs in all Asciz-deficient embryos analyzed (n>30; Figure 5A–5C) and apparent lack of tracheal tissue in all but one of these (Figure 5C and data not shown); consistently, in all cases where absence of lungs was subsequently noticed during routine MEF or protein preparations, this phenotype was 100% predictive of the Asciz−/− genotype (not shown).
Interestingly, the absence of lungs seemed to lead to topological alterations in the position of the heart and its axis within the thoracic cavity, with an apparent drop of the atrium in Asciz null embryos into the space otherwise occupied by the lung in WT littermates (Figure 5B, 5C). In addition, the thymus appeared hypoplastic in all Asciz−/− embryos analyzed (Figure 5A), which could also be a secondary consequence of the defective respiratory system as the thymus descends into the mediastinum from its common origin with the parathyroid gland in close proximity of the upper trachea [26]. Besides these thoracic defects, ∼25% of Asciz−/− embryos exhibited already macroscopically obvious exencephaly (Figure 5A and Table 1), indicating that ASCIZ also contributes to neural tube development, but histologically other organs seemed to be developing normally.
The combined trachea and lung defects in Asciz−/− embryos were interesting because both organs originate at the same time but presumably independently of each other from the common respiratory endoderm in the ventral foregut [27], [28]. Shortly after specification of respiratory precursors that are characterized by expression of the Nkx2.1 transcription factor, bilateral lung buds and, just rostral of these, a central tracheal primordium emerge from the ventral foregut around E9.5 in the mouse. The junction of these primordia marks the bifurcation of the trachea into the two main bronchi, and the lung buds expand caudally into the surrounding mesoderm to form the bronchial tree and pulmonary epithelium by branching morphogenesis, whereas the trachea septates from the common foregut lumen in an upwards “unzipping” motion [27], [28]. Thus, in a simplified view, the origin of the respiratory system can be traced back to the projection of the tracheo-bronchial bifurcation onto the ventral oesophagus.
To more clearly assess the developing trachea and lungs in three dimensions, we performed optical projection tomography (OPT) on whole-mount E-cadherin stained embryos. WT embryos showed clear separation of oesophagus and trachea at the larynx, bifurcation of the trachea into two bronchi and advanced branching of the developing pulmonary epithelium (Figure 6A, 6C). As expected, all five Asciz−/− embryos analyzed again lacked developing pulmonary epithelium (Figure 6B, 6D, Figure S5, and data not shown). One Asciz null embryo contained a very short incompletely separated tracheal stump that ended bluntly where it would normally connect to the main bronchi (Figure 6B). Interestingly, the other Asciz null embryos contained single centrally located bud-like structures that emerged from the ventral oesophagus near the level where the trachea bifurcates into bronchi in the relevant WT littermates (Figure 6D, Figure S5); the central location suggested that this bud-like structure represented tracheal primordium.
Two of the Asciz−/− whole-mount embryos and littermate controls were sectioned at the level of the truncated trachea (Figure 7B, 7B′) or tracheal bud-like structure (Figure 7D, 7D′) for immunofluorescence staining with the respiratory marker Nkx2.1. The tracheal stump in the mutant stained homogenously with Nkx2.1 (Figure 7B, bottom panel), similar to the trachea in the WT littermate (Figure 7A), and the ventral part of the tracheal bud-like structure in the other Asciz−/− embryo was also enriched for Nkx2.1 (Figure 7D′) with staining intensity similar to the separated trachea in the matched WT littermate control (Figure 7C′). Interestingly, in stark contrast to the WT oesophagus, some ectopic Nkx2.1-positive cells remained in the ventral part of the oesophagus in the mutant where the trachea had partially separated (Figure 7B, top panel).
We also analysed these sections for expression of p63, a p53-like transcription factor that is normally highly expressed in the oesophagus, but also present in basal cells of the trachea [29]. Under our staining conditions at the developmental stages studied here, p63 seemed only to be present in the oesophagus but not in the trachea in WT embryos (Figure 7A′, 7C). However, p63-positive cells were readily detectable in the ventral part of the tracheal bud-like structure in the Asciz−/− embryo (Figure 7D), suggesting defective partitioning of specified cells between trachea and oesophagus. As ectopic p63 expression can result from increased Sox2 levels [29], [30], a transcription factor involved in foregut separation that is normally highly expressed in the oesophagus and dorsal part of the trachea but downregulated in the ventral part of the developing trachea, we also monitored Sox2 expression in these sections. WT tracheas (Figure 7A, 7C, bottom panels) and the partially separated Asciz−/− trachea (Figure 7B, bottom panel) exhibited the expected dorsally polarized Sox2 expression pattern; in contrast, Sox2 was still expressed at high levels throughout the ventral part of the bud-like structure in the Asciz−/− embryo (Figure 7D). Thus, while impaired local down-regulation of Sox2 could contribute to the Asciz−/− phenotype, it is interesting to note that most of the ectopic Sox2-positive cells in the tracheal bud-like structure were still able to downregulate p63. We also observed aberrantly high Sox2 levels in the ventral foregut in Asciz−/− embryos around E10.25, i.e. before oesophagus and trachea were separated in the matched littermate control with appropriately down-regulated Sox2 (Figure S6), indicating that impaired dorso-ventral patterning of Sox2 expression is not merely a secondary consequence of impaired foregut separation in our mutant.
Altogether, these analyses indicate that Asciz-deficient mice are able to initially specify the respiratory endoderm, based on Nkx2.1 expression, but then fail to remodel the endoderm in a manner required for initiation of lung budding and efficient separation of the trachea.
When ASCIZ was originally isolated in a yeast two-hybrid screen [15], we noticed during vector-swapping control experiments that ASCIZ could very strongly activate yeast two-hybrid reporter genes on its own once it was fused to the Gal4 DNA-binding domain (Gal4-DBD). As a large proportion of genes that regulate foregut development function as transcription factors (e.g., Sox2, p63, Nkx2.1 mentioned above), and because the modular domain composition of ASCIZ resembles some ZnF transcription factors (see below), we revisited the yeast reporter system to explore the potential of ASCIZ to function as a transcriptional regulator. Both the four-ZnF 823-residue and the two-ZnF 667-residue splice isoforms of human ASCIZ were able to activate the GAL1-HIS3 and GAL2-ADE2 reporter genes in these one-hybrid assays (Figure 8A). Importantly, similar dual luciferase reporter assays in human U2OS cells using the 667-residue isoform demonstrated that ASCIZ also has an intrinsic ability to activate gene expression in mammalian cells when tethered to promoters (Figure 8B). Interestingly, truncation analysis revealed that the SQ/TQ-cluster domain - but not the ZnF or core domains - of ASCIZ was sufficient for reporter gene activation (Figure 8A).
Here we have shown that Asciz is essential for pulmonary organogenesis during embryonic development in mice, and required for proper DNA base damage responses in primary cells. Although the lung defect is mechanistically most likely unrelated to defective DNA damage responses, the overall phenotype - MMS and H2O2 hypersensitivity and embryonic lethality - is consistent with a role of ASCIZ as an accessory BER factor downstream of glycosylases, as proposed by previous work in human and chicken cells [15], [16]. Although Asciz null embryos die a few days earlier and their lung defect is considerably more severe than in case of Polß-deficient embryos, the latter also seem to have a very comparable late gestational growth retardation [10], [11], and furthermore, the essential requirement for Polß is also not suppressed by deletion of p53 [9]. Likewise, embryos deficient in Yb-1, another protein recently linked to accessory functions in the BER pathway [31], [32], also share overall similar late embryonic growth retardation and lethality, frequent exencephaly and modestly increased cellular oxidative stress-induced senescence phenotypes [33].
In contrast to similarities with BER-related genes, the phenotype of Asciz-deficient mice differs fundamentally from the phenotype of Atm-deficient mice. For example, the key phenotype of Asciz-deficient mice - embryonic lethality with absence of lungs - is not shared by Atm-null mice [20], and the key phenotype of Atm-deficiency - dramatically increased ionizing radiation sensitivity - is not shared by Asciz-deficient cells [16], [19]. Consistent with normal ATM protein levels in human, mouse or chicken cells in the absence of ASCIZ, ATM signaling was also unaffected in our Asciz-deficient MEFs or Asciz-depleted human cell lines (Figures S3, S4, and data not shown), including in response to HU, hypotonic NaCl and chloroquine, that required ASCIZ for ATM activation according to Kanu and Behrens [19]. Thus, the completely different phenotypes and absence of ASCIZ effects on ATM stability and activation question the classification of ASCIZ as an “essential co-factor” and regulator of ATM [19]. It is not clear why the other group obtained different results, as our gene targeting strategy was identical to theirs. Kanu and Behrens did not provide genetic background information for their mice, but given that we consistently observed unimpaired ATM levels in Asciz-deficient human, chicken or mouse cells, it seems unlikely that the differing effects could be mouse strain-dependent. As we have confirmed normal ATM levels directly in freshly prepared tissue extracts, we can also exclude the possibility that we may have missed differences in protein levels as a result of variable cell culture conditions. Likewise, given that we did not see a meaningful correlation between ATM and ASCIZ levels in numerous independent AT cell lines, including isogenic AT cell controls reconstituted with WT Atm, as well as genuine mouse and chicken Atm gene deletions (Figure 4), we can only speculate that the previously reported dramatic loss of ASCIZ may be a peculiarity of that particular AT cell line, possibly due to increased genome instability of AT cells. Considering that the positions of 15 potential ATM phosphorylation sites are exactly conserved from chicken to human and mouse ASCIZ, we favour a model where DNA damage-related functions of ASCIZ may be modulated by its direct phosphorylation by ATM. Indeed, our preliminary data that ASCIZ can be directly phosphorylated by ATM in vitro and that its MMS-induced focus formation in vivo seems to be at least partially regulated by ATM (to be reported elsewhere in detail) are consistent with a functional interaction between the two proteins.
As early lung development is unlikely to be specifically affected by DNA damage signaling, the finding of complete pulmonary agenesis and severe tracheal atresia in Asciz null embryos was surprising, particularly as there are very few mouse mutants with comparable respiratory defects (reviewed in [27], [28], [34], [35]). Specification and early development of the respiratory tract is regulated by extensive signaling crosstalk between the foregut endoderm and surrounding mesoderm [27], [28], and mouse mutants have revealed major signaling pathways involved in these processes (Figure 9). Double-knockout mice lacking the Gli2 and Gli3 transcription factors of the hedgehog pathway also seem to lack lungs as well as the trachea; however, they also lack the oesophagus indicating a more severe foregut defect [36](NB, these defects are considerably less severe in sonic hedgehog (Shh) null embryos [37]). Foregut development appears overall normal in Wnt2/Wnt2b double-null embryos as well as Shh-Cre driven conditional ß-catenin (Ctnnb1) KO mice, but these never establish the Nkx2.1-positive respiratory endoderm and consequently exhibit complete lung and tracheal agenesis [38], [39]. Mice lacking FGF-10 [40], [41] or its cognate FGF-receptor 2b [42] also lack lungs, but seem to contain a grossly normal trachea (and are also characterized by a complete absence of limbs in contrast to the Asciz−/− phenotype). Conversely, FoxG1-Cre driven conditional Bmp4 deletion results in selective tracheal agenesis, where the main bronchi and primitive lungs emerge directly from the oesophagus [43].
Based on these comparisons (Figure 9), the Asciz−/− phenotype is less severe than the complete respiratory precursor defect with absence of Nkx2.1 expression and combined agenesis of lungs and trachea in Wnt2/2b and Shh-Cre/ß-catenin mutants, but more severe than the respiratory tract defect in Fgf10 or FGF-receptor 2b mutants with selective pulmonary agenesis yet preserved tracheal development. Genetically, these data thus suggest a crucial regulatory function for ASCIZ in the regulation of respiratory organogenesis at a level between endodermal ß-catenin and mesodermal FGF10 signaling pathways (Figure 9). As FGF10 has been proposed to regulate the downregulation of Sox2 expression in respiratory precursors [30], our finding of impaired dorso-ventral patterning of Sox2 expression in Asciz−/− embryos before foregut separation (Figure S6), and when tracheal separation stalls early (Figure 7D), are also consistent with a role of ASCIZ upstream of FGF10.
The signaling pathways discussed here ultimately regulate developmentally important gene expression programs during specification, morphogenesis and differentiation of the respiratory system. ASCIZ is a predominantly nuclear protein [15], [19], its ZnF structure is generally reminiscent of transcriptional regulators [44], and we have shown here that ASCIZ has the propensity to function as a transcriptional activator via its SQ/TQ cluster domain (Figure 8). In some regards, ASCIZ can be considered as a mirror image of the Sp1 transcription factor. Whereas ASCIZ contains a ZnF domain at the N-terminus and an extended SQ/TQ cluster towards the C-terminus, Sp1 that also is essential for murine development [45] contains an SQ/TQ rich N-terminal transcription activation domain and a triple-ZnF domain at the C-terminus. While Sp1 has been extensively studied as a transcription factor, it is now becoming apparent that it also has transcription-independent roles as an ATM substrate that relocates into DNA damage-induced foci [46], [47], somewhat similar to our original interest in ASCIZ. Altogether, these analogies make it tempting to speculate that ASCIZ may regulate pulmonary development as a transcription factor.
In conclusion, we have shown here that ASCIZ has dual functions with a role in the response to DNA lesions that are repaired by the BER pathway, as well as pleiotropic functions during murine embryonic development, most notably as a member of a very select group of essential regulators of respiratory organogenesis. Nkx2.1-positive respiratory precursors seem to still be specified in the absence of Asciz, but then fail to properly segregate within the foregut. Impaired foregut separation in Asciz−/− embryos seems to correlate with an inability to downregulate Sox2 expression in the ventral foregut, but the exact mechanism responsible for this defect remains to be determined. Asciz null embryos die a few days before birth rather than perinatally from an acute inability to breathe, indicating that additional developmental defects beyond the respiratory system may contribute to the lethality. Our study provides a basis to further investigate how exactly ASCIZ regulates respiratory organogenesis and possibly other developmental processes by expanding the analysis to tissue-specific or temporally regulated conditional knockout systems.
All mouse procedures were approved by the St. Vincent's Hospital Animal Research Ethics Committee.
The mouse Asciz gene contains four exons spread over ∼15 kbp on chromosome 8 and was targeted in C57BL/6 ES cells by integrating loxP sites on both sides of exon D (Figure 1B) using standard homologous recombination, ES cell and blastocyst manipulation techniques as a contracted service by Ozgene Pty Ltd, Perth. A diagnostic ScaI restriction site was integrated just 3′-terminal of the downstream loxP site. Gene targeting was confirmed by Southern blotting using 5′- and 3′-probes located outside the targeting vector. The 5′-probe can be used for Southern blot analysis of ScaI-digested DNA to distinguish between the WT (22.5 kbp), targeted (11.5 kbp), and Asciz-KO (7.6 kbp) alleles (Figure 1B). The germline Asciz KO allele was generated by crossing the targeted line with C57BL/6 mice containing a PGK-Cre knockin in the ROSA locus, followed by two C57BL/6 backcrosses to remove PGK-Cre and for embryo transfer into the St. Vincent's Hospital Biological Resources Centre. Thus, the Asciz KO line is on a pure C57BL/6 genetic background. Animals were housed in SPF microisolators. Genotyping can also be performed by PCR using primers F1 (5′-CATGGAATTGTTAAAAGCTC-3′), F2 (5′-CCGACTGGGGATGTAGTCAG-3′) and R1 (5′-AAAAGATAGAATAGCTACAC-3′), which result in bands of 170 bp for the WT and 220 bp for the KO allele.
Asciz+/− mice were crossed with germline p53-targeted mice (deletion of exon 2–10 [48], [49]) to generate compound heterozyotes, and offspring were genotyped at weaning using primers above and p53 primers Trp53-1F (5′-CACAAAAAACAGGTTAAACCCAG-3′), Trp53-1R (5′-AGCACATAGGAGGCAGAGAC-3′) and Trp53-10R (5′-GAAGACAGAAAAGGGAGGG-3′), which result in bands of 290 bp for the WT and 612 bp for the KO allele. Recovery of p53−/− mice at weaning was approximately half of the expected Mendelian ratios, a known phenomenon for p53 null homozygosity on inbred backgrounds [50].
The time of pregnancies was defined as E0.5 on the morning vaginal plugs were observed in Asciz+/− intercrosses. Embryos were dissected from the uterus in cold PBS, weighed after blotting off excess fluid and immediately fixed for histology, or processed for protein extraction or MEF isolation, and genoyped by PCR using yolk sac or tail DNA. For histology, whole embryos were fixed in Bouin's solution or paraformaldehyde, processed to paraffin and sagittal sections were stained using haematoxylin-eosin and scanned using a Zeiss Mirax Digital Slide Scanner by the Australian Phenomics Network Histopathology and Organ Pathology Service, University of Melbourne, or manually processed and photographed as described [51].
Human and chicken cell lines were cultured as described [15], [16], [24], [52]. MEFs were prepared by dissecting embryos in cold PBS, heads and internal organs were removed, and remaining corpses were sliced into smaller pieces and trypsinized cells were cultured in Dulbecco's Modified Eagles medium containing 10% fetal calf serum for 2 days, trypsinized, Coulter-counted, and re-seeded at 106 cells per 10 cm dish. This passage, defined as P1, was incubated for 3 days, trypsinized, counted and re-seeded at 106 cells per 10 cm dish (P2), and this process was repeated for 8 passages. For DNA damage sensitivity assays, 5×104 MEFs (P2–P3) were seeded per 35 mm well, grown in Dulbecco's modified Eagle's medium and treated as indicated in the figure legend, and after 18 hours cell viability was determined by propidium iodide exclusion using flow cytometry. Each set of DNA damage sensitivity experiments was performed in parallel with MEFs from at least three independent embryos per genotype.
Staged embryos were stained for OPT [53] with an antibody to E-cadherin (ECCD2, Invitrogen, 1/200 dilution) as described [54], with 48 hour primary and secondary antibody incubations interspersed with extensive 12 hour washes to remove unbound antibody. Samples were imaged on a Bioptonics 3001 OPT machine (Bioptonics, UK) and datasets reconstructed by NRecon (Skyscan, Belgium) and visualized using Drishti (http://anusf.anu.edu.au/Vizlab/drishti/). Embryos were rescued from agarose after imaging, processed to paraffin and sectioned, or directly prepared for cryo-sectioning. After antigen retrieval in citrate buffer sections were stained with antibodies to Nkx2.1 (anti-TTF1, Zymed, 1/200), p63 (Abcam, 1/200), and Sox2 (Chemicon) to examine differentiation.
Southern, Northern and Western blots were performed as described [15], [51]. Antibodies against ASCIZ ([15], available from Millipore) and chicken ATM [52] were described before. Other antibodies: Actin (MAB1501, Millipore), ATM (5C2, Abcam), ATR (sc-1887, Santa Cruz Biotechnology), human p53 (sc-126, Santa Cruz Biotechnology), mouse p53 (1C12, Cell Signaling Technology, PML (sc-5621, Santa Cruz), XRCC1 (sc-11429, Santa Cruz), γH2AX (05-636, Millipore), pS1981(mouse: pS1987)-ATM (200-301-400, Rockland; or 10H11.E12, Cell Signaling Technology), pS15(mouse: pS18)-p53 (9284, Cell Signaling Technology), pT68-Chk2 (2661 or 2197, Cell Signaling Technology).
For yeast assays, ASCIZ constructs were cloned in pAS2.1 and transformed into PJ69-4A, except the isolated SQ/TQ cluster domain that was cloned into the low-level expression vector pGBT9 because its high level expression was toxic in yeast. One-hybrid reporter assays were performed essentially as described previously for two-hybrid assays in our laboratory [55], [56] except that plates were supplemented with leucine. For mammalian dual luciferase reporter assays, the 667-residue ASCIZ isoform was cloned into pCDNA3-Gal4DBD for transient transfection of U2OS cells with equal amounts of the reporter vectors pFR-Luc and pRL-CMV for use with the Dual-Luciferase Reporter Assay kit (Promega) according to the manufacturer's instructions and measurement of luminescence using a Polarstar Optima (BMG Labtechnologies).
|
10.1371/journal.pgen.1003852 | Genome Wide Analysis Reveals Zic3 Interaction with Distal Regulatory Elements of Stage Specific Developmental Genes in Zebrafish | Zic3 regulates early embryonic patterning in vertebrates. Loss of Zic3 function is known to disrupt gastrulation, left-right patterning, and neurogenesis. However, molecular events downstream of this transcription factor are poorly characterized. Here we use the zebrafish as a model to study the developmental role of Zic3 in vivo, by applying a combination of two powerful genomics approaches – ChIP-seq and microarray. Besides confirming direct regulation of previously implicated Zic3 targets of the Nodal and canonical Wnt pathways, analysis of gastrula stage embryos uncovered a number of novel candidate target genes, among which were members of the non-canonical Wnt pathway and the neural pre-pattern genes. A similar analysis in zic3-expressing cells obtained by FACS at segmentation stage revealed a dramatic shift in Zic3 binding site locations and identified an entirely distinct set of target genes associated with later developmental functions such as neural development. We demonstrate cis-regulation of several of these target genes by Zic3 using in vivo enhancer assay. Analysis of Zic3 binding sites revealed a distribution biased towards distal intergenic regions, indicative of a long distance regulatory mechanism; some of these binding sites are highly conserved during evolution and act as functional enhancers. This demonstrated that Zic3 regulation of developmental genes is achieved predominantly through long distance regulatory mechanism and revealed that developmental transitions could be accompanied by dramatic changes in regulatory landscape.
| The Zic3 transcription factor regulates early embryonic patterning, and the loss of its function leads to defects in left-right body asymmetry. Previous studies have only identified a small number of Zic3 targets, which renders the molecular mechanism underlying its activity insufficiently understood. Utilizing two genomics technologies, next generation sequencing and microarray, we profile the genome-wide binding sites of Zic3 and identified its target genes in the developing zebrafish embryo. Our results show that Zic3 regulates its target genes predominantly through regulatory elements located far from promoters. Among the targets of Zic3 are the Nodal and Wnt pathways known to regulate gastrulation and left-right body asymmetry, as well as neural pre-pattern genes regulating proliferation of neural progenitors. Using enhancer activity assay, we further show that genomic regions bound by Zic3 function as enhancers. Our study provides a genome-wide view of the regulatory landscape of Zic3 and its changes during vertebrate development.
| Early embryonic patterning is achieved through a process involving the determination of body axes and defining which cell types develop at each coordinate. The Zic family of transcription factors (TFs) is involved in such process [1]–[4]. Zic genes are the vertebrate homologues of the odd-paired gene, which is involved in the generation of segmental body plan in the Drosophila embryo [5], [6]. Although functions of Zic proteins partially overlap, their loss-of-function cause distinct phenotypes, suggesting unique roles in development [7], [8].
Of particular interest is ZIC3, which is linked to the heritable defects of the left-right internal organs placement (situs inversus) in humans [9]. Studies in animal models reveal the involvement of Zic3 the establishment of left-right (L-R) asymmetry [1], [10]–[12]. In Xenopus, Zic3 established left-sided expression of Xnr1 and Pitx2 [12], two determinants of internal organs asymmetry [13]–[15]. However, zic3 is expressed symmetrically along the L-R axis in the Xenopus embryo and its loss-of-function (LOF) affects structures in which its expression was not detected [1], [12]. Results from several studies provided clues to the mechanism of L-R patterning by Zic3. First, Zic3 acts in organizer formation by inhibiting the canonical Wnt signaling pathway [16]. Second, Zic3 regulates gastrulation in mouse [1], [17]. Furthermore, studies in zebrafish revealed a correlation between convergence-extension (C-E) and L-R patterning defects in Zic3 LOF [10]. These suggest that Zic3 may regulate L-R patterning through its role in an earlier developmental event such as C-E.
Zic3 is one of the earliest TFs expressed in the neuroectoderm [3], [18]. Its expression is regulated by determinants of the early neural fate specification and dorsal-ventral (D-V) axis formation, including BMP, FGF, and Nodal signaling [3], [17], [19], [20]. The role of Zic3 in establishing neural cell fate was demonstrated through experiments in Xenopus, where its overexpression resulted in the expansion of the neuroectoderm and induction of neural and neural crest markers [18]. This led to the assumption that Zic3 activates the expression of proneural genes such as Achaete-scute homologs, Neurogenin, and NeuroD [2]. However, Zic3 lacks the ability to induce ectopic neuronal differentiation in the epidermis [18], which suggested the complex interaction between Zic3 and the proneural genes.
Increasing evidence has established the presence of long-distance interactions between TFs and their target genes [21]–[24]. This feature is especially true for TFs regulating specific functions outside of the core transcription machinery [25]–[27]. Therefore, an unbiased evaluation of binding sites throughout the whole genome would be a more comprehensive and biologically relevant method in the context of a developing organism. However, genomic approaches to study TFs in vivo are often limited by the quantity of available tissue sample. Furthermore, in mammalian systems, this problem is exacerbated by the short supply of embryos at early developmental stages. The zebrafish, with its unlimited supply of embryos and external development, substitutes for the inconveniences of a mammalian system. Its genome annotation is also the most complete among non-mammalian vertebrates and the expression of many genes are well-defined. This makes the zebrafish a robust model system for functional studies of vertebrate development.
To understand the developmental role of Zic3, we applied a genomic approach to identify genes directly regulated by Zic3. To capture genome-wide binding sites of Zic3, chromatin fragments bound by Zic3 were immunoprecipitated from gastrulating embryos at 8 hpf and zic3 expressing cells were sorted from transgenics [21], [28] at 24 hpf and sequenced in-depth using ChIP-seq methodology. This provided unbiased coverage of Zic3 binding events during the period of gastrulation and segmentation. We used microarray expression profiling to characterize changes at the transcription level as a result of Zic3 LOF during gastrulation. In addition, we compared gene expression profiles of zic3-positive and -negative cells at 24 hpf to identify genes co-expressed with zic3. Combining binding site analysis and expression data, we demonstrated that Nodal and Wnt pathways are the main downstream targets of Zic3 during gastrulation, and show distinct pathways regulated by Zic3 in the dorsal neural tube at the end of segmentation. Finally, in vivo enhancer assay validated selected binding sites as developmental enhancers. Our results provide novel insights into the molecular mechanism underlying Zic3 regulation of developmental events during gastrulation and neural development, which ultimately results in the L-R patterning and neural fate specification and patterning.
The earliest zic3 transcript was detected at 3 hpf (Fig. 1A,B), coinciding with the initiation of zygotic transcription during mid-blastula transition [29]. At 4 hpf zic3 expression is restricted to dorsal blastoderm (Fig. 1C,C′), and is subsequently found in the dorsal neuroectoderm and marginal blastomeres (Fig. 1D, D′). To capture genome-wide Zic3 binding profile during zebrafish gastrulation, we performed ChIP-seq analysis at 8 hpf, a time coinciding with the beginning of neurogenesis [30]. At this time zic3 is expressed largely in the dorsal neuroectoderm (prospective neural plate) and blastoderm margin (presumptive mesendoderm; Fig. 1E,E′; [3]). Hence, the interaction of Zic3 with its targets could be considered within a context of neural induction and mesendodermal development. Although neuroectoderm does not show any obvious morphological organization at this time, its anteroposterior patterning at the molecular level was shown by fate mapping studies [31] and in vitro explant assays [32], [33]. At 24 hpf zic3 is expressed in the brain and dorsal spinal cord (Fig. 1F,F′). To identify Zic3 binding sites specifically in zic3-expressing cells, we performed ChIP-seq using sorted cells from transgenic line SqET33 [28], [34] at this stage. Since gfp expression in this line faithfully recapitulates zic3 expression (Fig. 1G–H″), we considered GFP-positive cells as zic3-expressing cells and GFP-negative cells as non- zic3-expressing cells. However, it is worth to note that in SqET33 line at least one zic3-positive domain (presomitic mesoderm) does not express GFP. This suggests that a small fraction of non-neuronal zic3-expressing cells may be present in the GFP-negative pool of cells.
Sequencing of the 8 hpf ChIP sample generated 23,945,552 reads (11,037,221 or 46% were mapped to the zebrafish genome); the 24 hpf ChIP sample generated 23,083,504 reads (11,797,011 or 51% were mapped). We identified 3209 and 2088 Zic3 binding sites (hereafter referred to as peaks) with high significance value at 8 hpf (Table S13) and 24 hpf (Table S14), respectively. Interestingly, both datasets showed that only a small fraction (8.6% at 8 hpf and 4% at 24 hpf) of the peaks mapped to promoter regions (within 5 kb of transcription start site, TSS), while the rest were aligned to intragenic (26.8% at 8 hpf and 29% at 24 hpf) and intergenic (64.6% at 8 hpf and 67% at 24 hpf) regions (Fig. 2A). This suggested that Zic3 mainly acts via distal regulatory elements. To validate the ChIP-seq performance, we carried out quantitative PCR (qPCR) on randomly selected peaks from the 8 hpf dataset, five within promoter region and sixteen at regions outside of gene promoters. Taking a fold-change of 2 as a cutoff for positive enrichment, the qPCR analysis validated all but one peak tested (Table S1).
To determine the biological relevance of our data, we used the gene association rule ‘basal plus 100 kb extension’ according to GREAT algorithm [35] (Fig. 2B). Using this criterion, the number of peaks associated with either none, one, or two genes were evenly distributed in both 8 hpf and 24 hpf datasets (Fig. 2C). Distribution of the peaks relative to the TSS of genes associated with them showed strong bias towards regions beyond 5 kb of the TSS (Fig. 2D). In agreement with known Zic3 functions at 8 hpf [10], [16], [18], [36] functional categories enriched were embryonic morphogenesis, gastrulation, and dorsal/ventral pattern formation (2835 genes, Fig. 2F; Table S2). Enrichment was also observed for neural tissue-specific genes, predominantly expressed in the neuroectoderm at 8 hpf (Fig. 2G). In contrast, at 24 hpf, different categories were enriched (neural crest development and migration, nervous system development; Fig. 2H,I) in agreement with these events of neurodevelopment taking place at this stage [18], [37].
To identify the common regions bound by Zic3 as well as those unique to either developmental stage, we overlapped the 8 hpf and 24 hpf peaks (Fig. 2E). Taking the combined list of peaks from 8 hpf and 24 hpf, we performed clustering using ChIP-seq signals around the peaks. We found 937 regions bound by Zic3 at both stages (class I), 2729 regions bound only at 8 hpf (class II), and 1630 regions only at 24 hpf (class III). A clear distinction of functional categories was observed among genes associated with each individual class (Fig. S2), which reflect the shift of Zic3 function from regulating gastrulation at 8 hpf, to directing neurodevelopment at 24 hpf.
To identify the consensus motif in Zic3-binding sites, we performed de novo motif search using sequences within 50 bp (total length 100 bp) of the top 1000 peaks summit. The highest scoring motif in both datasets consisted of a CAGCAG core (Fig. 3A) and was similar to that previously identified in mouse ES cells using ChIP-chip [38] (Fig. S3A) and Zic3 motif in UniPROBE database [39]. This motif occurred in 48.5% (1556/3209) of 8 hpf peaks and 54.3% (1134/2088) of 24 hpf peaks (Fig. 3B). This consensus motif was bound in a dose-dependent manner by a recombinant protein encompassing the Zic3 DNA binding domain (Zic3_ZF2-5; Fig. 3C). This binding was reduced upon introducing three-point mutations to the motif, confirming binding specificity. The mouse Zic3 recombinant protein mZic3-DBD-HisMBP [38] also recognized the consensus motif derived from the zebrafish genome (Fig. S3B), demonstrating cross-species conservation of Zic3 consensus motif. On the other hand, two other motifs enriched in the dataset to a lesser extent were not specifically recognized by Zic3_ZF2-5 recombinant protein (Fig. S3C). Enrichment of these motifs among the identified peaks might signify an indirect binding of Zic3 to these sequences through interaction with other TFs. Interestingly, Gli motif was found in both 8 hpf and 24 hpf datasets (273 peaks, 8.5% in 8 hpf; 203 peaks, 9.7% in 24 hpf; Fig. 3B). More than half of peaks containing Gli motifs also had an adjacent consensus Zic3 motif at both developmental stages, in support of interactions between Gli and Zic3 [40], [41].
To identify Zic3 target genes during gastrulation and early neural development, we profiled the transcriptome of 8 hpf embryos after Zic3 morpholino (MO)-mediated knockdown. Embryos injected with the same MO dosage as in Cast et al. [10] exhibited similar gastrulation and convergent extension (C-E) defects (data not shown). However, to minimize the detection of non-direct targets in microarray, we injected the embryos with a lower dose of MO (1.7 ng in our experiments versus 7.5 ng in [10]) which did not cause visible morphological defects during gastrulation (refer to Methods section), but affected heart laterality and caused curvature of the A-P axis at later stages (Fig. 4A). These phenotypes were rescued by co-injection with Zic3 mRNA which, when injected alone, had little effect (Fig. 4B). This confirmed the specificity of the phenotypes caused by Zic3 MO injection.
We identified 1316 genes differentially expressed in MO injected embryos (morphants, fold change >1.2; p≤0.05; Table S3). GO analysis revealed prominent enrichment in functions related to embryonic morphogenesis (Table S4). When the same or higher dose of MO (3.4 ng) was injected, the expression of several representative genes showed similar trend when measured by qPCR. This validated a possibility of their regulation by Zic3 (Fig. 4C; Table S7). We then determined the presence of Zic3 binding peaks within 100 kb of the TSS of these differentially expressed genes, which we defined as a selection criterion for Zic3 target gene. Based on this selection, 454 genes out of the total 1316 were identified as putative targets of Zic3 (Table S5 and Table S6). This set contains genes of the Nodal signaling pathway such as oep, lft1 and pitx2 (Fig. 5). While the presence of Zic3 binding in association with oep suggests direct regulation of Nodal pathway, the association of Zic3 peaks with lft1 and pitx2 suggests that Zic3 could also regulate the pathway through its modulators [42], [43]. These three genes, along with other members of this pathway not associated with Zic3 peaks (foxh1, bon, and gsc), were concurrently upregulated in Zic3 morphants (Fig. 4C; Table S3) suggesting negative regulation of the Nodal pathway by Zic3. Inhibition of Nodal signaling indicates suppression of endodermal fate [15], [44]–[46]. This correlated with broader expression of endodermal marker sox17a in 8 hpf Zic3 morphants (Fig. S4A). The inhibition of endodermal development by Zic3 is in line with previous observation in murine ES cells [38]. Similarly, peaks were associated with three genes of the canonical Wnt signaling pathway: axin1, jun, and vent (Table S5). In support of this association, microarray analysis revealed that the negative regulator of canonical Wnt pathway axin1 was downregulated in Zic3 morphants, while the downstream components jun and vent were upregulated (Fig. 5; Table S3). The expression of some other members of this pathway (axin2 and nlk1) without association with peaks has changed in Zic3 morphants based on microarray data. This implied that such genes could be the indirect targets of Zic3. Such observation provided further support for Zic3 regulation of the canonical Wnt pathway. The inhibition of canonical Wnt signaling by Zic3 was previously reported in frogs as a mechanism for organizer development [16]. Interestingly, Zic3 LOF only affected downstream components of these signaling pathways, and not the ligands, suggesting that at 8 hpf Zic3 is more likely to modulate the response to Wnt signaling in the target cells rather than initiation of signaling.
Apart from genes previously implicated as targets of Zic3, the combined ChIP-seq and microarray screen also identified novel candidates. Zic3 peaks were found in association with genes known to regulate cell proliferation in the neural plate, dlx4b and msxe [47], [48]. These genes perform a function [49], [50] similar to that of msxc, irx1a, and irx7, which do not have associated peaks but were nevertheless downregulated in Zic3 morphant (Table S3; S7). This observation suggests the role of Zic3 in promoting proliferation of neural progenitors at 8 hpf. Since these genes are known to inhibit neural differentiation, we assayed the expression of proneural gene neurog1 [51] in Zic3 morphants at 10 hpf. As expected, neurog1 was upregulated, in concert with the downregulation of her9 (Fig. 4C; Table S7), which provided further support for Zic3 role as a promoter of proliferation of neural progenitors and repressor of neural differentiation.
More interestingly, the novel candidate targets include members of the non-canonical Wnt signaling pathway (dvl2, rock2b and invs). These genes were co-expressed with zic3 during gastrulation (Fig. S5A, B) and were downregulated in the microarray (Table S3; Fig. 4C). One of the non-canonical Wnt pathways, the planar cell polarity (PCP), regulates convergence-extension (C-E) [52] and controls the positioning of the motile cilia [53]. The changes in expression of sox17, ntl, pax3a and sox19a mark correspondingly, endoderm, mesoderm, neural crest and neural plate. The broadening of their expression domains suggested that in Zic3 morphants C-E is affected (Fig. S4B–D, [10]). On the other hand, the disorganized expression of foxj1a and sox17a in the dorsal forerunner cells at an earlier stage indicated abnormalities of their migration in Zic3 morphants (Fig. S6), which may lead to abnormalities in L-R patterning. A correlation between C-E defects and L-R defects in Zic3 morphant was reported [14], suggesting Zic3 regulation of these events through the non-canonical Wnt pathway.
Several genes implicated in cell migration and polarity were among the targets. These include npy [54], ptenb [55], sepn1, srsf1a [56], and sparc [57], [58], all of which were downregulated in microarray and associated with peaks. WISH analysis showed that their expression overlap that of zic3 (Fig. S5C; ZFIN; University of Oregon, Eugene, OR 97403-5274; URL: http://zfin.org/; 21 June 2013). In addition, other genes with similar function, such as ccdc88a (probe generated from BC057440 which correspond to the annotated ccdc88a sequence) [59], [60] and tsg101 [58], were also downregulated in the microarray despite not having associated peaks. Hence the direct and indirect regulation of these genes by Zic3 could be the mechanism behind cell movements during gastrulation.
To identify potential zic3 targets during late neurogenesis, we performed microarray expression analysis on 24 hpf GFP-positive zic3 expressing cells that were FACS-sorted (Table S8). Comparing expression levels to a control dataset derived from GFP-negative cells (cells negative for zic3 expression), we identified genes enriched in GFP-positive cells (zic3-expressing cells). A total of 689 genes (p-value<0.05; fold change ≥1.5) were enriched in zic3-expressing cells (zic3-coexpressed genes). Among these genes were six members of the Zic family and other genes expressed in the dorsal neural tube. This confirmed the identity of the sorted cells as dorsal neural cells. Among the zic3-coexpressed genes, 167 had at least one peak within 100 kb of their TSS, rendering them putative Zic3 targets (Table S10). Similar to the 8 hpf stage, members of the Wnt pathway were also among the targets. However, Zic3 seems to regulate a different set of Wnt components, including wnt11r and lef1 (Fig. 6, Table S8). qRT-PCR revealed that wnt11r, were down-regulated in Zic3 morphants at 24 hpf (Fig. 4C; Table S7), confirming their positive regulation by Zic3. Two other genes encoding Wnt ligands, wnt10a and wnt10b, were co-expressed with zic3, and regulated upon Zic3 knockdown (Table S7; Fig. 4C) although they were not associated with peaks in ChIP-seq, suggesting that they may be indirect targets of Zic3. A striking difference between 8 hpf and 24 hpf regulatory landscape is apparent from the distinct functions associated with Zic3 target genes at each stage. For example, many genes regulating cell migration and polarity were identified as Zic3 targets at 8 hpf, whereas at 24 hpf neural crest determinants were found. The latter included foxd3, and pax3a which were further confirmed to be responsive to Zic3 knockdown (Fig. 4C, Table S7, S11).
On the other hand, in zic3-negative cells, 835 genes were enriched by at least 2-fold (non zic3-coexpressed genes enriched for endoderm and mesoderm-specific expression terms, Table S9). Among these, 195 had peaks within 100 kb of their TSS, suggesting repression of these genes in cells expressing zic3 (Table S10). Several proneural genes (neurod, neurod4, ascl1a) were found under this category, which may reflect that the zic3-expressing cells in the dorsal neural tube are not differentiating. Interestingly, the presence of a Zic3 peak in association with oep suggests that a similar inhibition of Nodal by Zic3 occurs at both 8 hpf and 24 hpf (Fig. 6).
Taken together, an entirely different set of candidate Zic3 target genes were found at 24 hpf compared to 8 hpf (Fig. 6). Although similar signaling pathways, such as the Wnt and Nodal pathways, were regulated by Zic3 at both developmental stages, different members of these pathways were targeted by this regulation at each stage. Furthermore, the global shift in Zic3 binding sites from 8 hpf to 24 hpf suggested the presence of complex regulatory changes accompanying developmental transitions.
The large number of Zic3 binding sites in the distant intergenic regions suggested that Zic3 may direct the expression of target genes by binding to the distal regulatory elements. In support of this idea, relevant biological categories could be observed among genes associated with peaks located outside of their basal regions of −5 kb to +1 kb of TSS (2716 genes; Table S2; Fig. S7A) or at a distance more than 50 kb (989 genes; Table S2; Fig. S7B). In contrast, no particular enrichment of GO categories could be observed for 119 genes associated with peaks in their basal region (Table S2). Of these, 77 had expression data in ZFIN (University of Oregon, Eugene, OR 97403-5274; URL: http://zfin.org/; 21 June 2013), but none of these were co-expressed with zic3 at 8 hpf, while only 6 (lppr3a, p2rx3b, lingo1b, myo15aa, robo4, gng3) had expression overlapping with zic3 at 24 hpf (not shown).
To test whether peaks associated with distal genes function as regulatory elements, we used the enhancer activity reporter assay [61]. We chose five distal peaks associated with genes from Nodal and Wnt signaling pathways, including oep (fragment 10-02, 94.7 kb downstream from TSS), axin1 (fragment 3-43, 71.53 kb downstream), lft1 (fragment 20-35, 29.77 kb downstream), dvl2 (fragment 7-214, 55.92 kb downstream), and invs (fragment 16-297, 78.08 kb downstream). A canonical Zic3 motif was present within 100 bp of each peak summit except for fragment 10-02. Only fragment 16-297, associated with invs, showed enhancer activity (Fig. 7B,C,G; Table S12). When the association region was extended to 500 kb, we found more peaks associated with dvl2 (fragment 7-211, 236.6 kb upstream), axin2 (fragment 3-56, 147.9 kb upstream), and pitx2 (fragment 14-37, 180.32 kb upstream). These peaks had at least one canonical Zic3 motif and exhibited positive enhancer activity (Fig. 7, Table S5, S12). Intriguingly, some of the expression patterns driven by the tested enhancers only partially matched that of the associated genes (fragments 14-37 and 3-56; Fig. 7D,E), which could be due to functional dependence on interaction of multiple regulatory elements [62], [63]. Nevertheless, the presence of Zic3-binding sites with an enhancer activity near genes responding to Zic3 LOF suggested that these genes were direct targets of Zic3.
To validate the activation of the enhancer fragments by Zic3, we co-injected fragment 7-211, which drove the strongest reporter gene expression at 8 hpf and 24 hpf (Fig. 7C), and Zic3 MO into the zebrafish embryo. When assayed by qRT-PCR at 8 hpf, a significant decrease in reporter expression in a MO dose-dependent matter was observed (Fig. S8). No reduction in reporter expression was observed when control MO was used. A similar result was obtained when two other fragments, 4-16 and 17-24 which coincided with CNEs (Tables 1, S12), were tested (Fig. S8), demonstrating Zic3-dependent induction of reporter expression through these fragments.
To study whether Zic3 binding sites were evolutionarily conserved, we overlapped the 8 hpf dataset with a list of known conserved non-coding elements (CNEs; ANCORA database) [64].We identified 228 peaks as CNEs conserved between zebrafish and Tetraodon, and 56 as CNEs conserved between zebrafish and humans (Fig. 8A), with 31 in common between the two groups. Similar to the distribution profile of the full set of peaks, these CNE peaks were mostly located outside of the basal promoter region (Fig. 8B). Genes associated with these CNEs were enriched for developmental functions and neural tissue-specific expression (Fig. 8C,D; Table S2).
Of 15 CNE peaks tested for enhancer activity, 11 (73%) drove gfp expression at either 8 hpf or 24 hpf, or both (Table 1). Of these eleven, eight drove higher gfp expression compared to the reporter vector alone at 8 hpf (fold change at least 1.5 compared to enhancer-less vector). Of these eight, four continuously drove reproducible tissue-specific gfp expression in various regions of the CNS up to 24 hpf (Fig. 8E–H), which overlapped with known expression domains of zic3 (Fig. 1F). Another three CNE peaks drove reporter expression only at 24 hpf. The CNE peaks with enhancer activity included the fragments 4-16 and 20-4, which drove expression in the brain, eye and trunk. In the hindbrain, both drove similar expression in neuroepithelial cells with radial morphology. In the trunk, activity of 4-16 was detected in muscle cells, whereas that of 20-4 was largely confined to the neural tube (Fig. 8E,F). The gfp expression pattern driven by 4-16 partially recapitulated that of a nearby gene, sox5. On the other hand, 20-4 was located in a gene desert region, suggesting long distance regulation. Fragment 15-26 drove gfp expression largely in cells along the neural tube (Fig. 8G), which partially recapitulated the expression of tbx2b nearby. Fragment 1-22 drove gfp expression mainly in the hindbrain region (Fig. 8H), which partially recapitulated that of the nearby mab21l2.
On the other hand, out of 12 non-CNE peaks tested only two (17%) drove higher gfp expression than the reporter vector alone at 8 hpf (Table 1). Together with the fragments corresponding to peaks associated with microarray-identified genes, out of 35 fragments tested for activity as enhancers, 17 (49%) were positive. Two thirds of the active peaks were previously identified as CNEs. Whereas this indicated somewhat better chance of finding enhancers amongst CNEs, it also suggested that a significant number of enhancers are not conserved in evolution.
The majority of Zic3 binding sites were found outside promoter regions. While this could be partially attributed to the incomplete annotation of promoter regions in the zebrafish genome, the predominantly distal distribution of Zic3-binding sites revealed that Zic3 regulates transcription largely via distal regulatory elements. Such distribution of binding sites was previously observed in other genome-wide analyses of several TFs in cell culture or mammalian tissues [21], [22], [25], [65]. Our findings therefore establish that a similar distal regulatory mechanism is in effect within the context of Zic3 function during development in vivo.
Some of the Zic3 binding sites overlapped with CNEs, most of which drove expression in neural tissues. CNEs are known to regulate developmental genes [66]–[69]. However, in our dataset CNEs represented only 5% of the total Zic3 binding sites identified, while the majority was under weak evolutionary constraint. Tissue-specific enhancers have been shown to differ in the extent of evolutionary conservation of their sequence [70], [71]. Having only 5% overlap with CNEs, the set of Zic3-binding sites showed a similar trend. The lack of sequence conservation could be explained by the relaxation of selection pressure towards regulatory elements [72] owing to the genome duplication event in teleosts [73]–[75]. Given that at least for now the data available in zebrafish and mammals suggest that only a minority of sites are conserved in both classes of animals, other explanations should be considered. Detailed characterizations of other TFs in the zebrafish would provide a better understanding of the extent of conservation in regulatory regions in teleosts.
Cell culture studies have demonstrated interactions between multiple enhancer elements in regulating the transcription of a target gene [24], [62], [63], [76], [77], as well as interactions between a TF and different binding partners which can result in alternative transcriptional outputs [26], [78], [79]. Our results provide an insight of such complexity of transcriptional regulation by Zic3 in developmental context in vivo. For instance, the concurrent upregulation and downregulation of different subsets of direct target genes by Zic3 suggest that Zic3 binding can result in either activation or repression of target genes, and implies that additional mechanisms determine these two outcomes. Another facet of the data revealed distinct Zic3 binding profiles at 8 hpf and 24 hpf. The genes associated with binding events at these two stages showed relevant functional enrichments. This shift in binding was not dictated by a change in DNA recognition motif as almost identical dominant motifs were identified in both stages.
The combinatorial analysis of ChIP-seq and microarray datasets revealed an entirely distinct set of candidate Zic3 target genes at 8 hpf and 24 hpf. Whereas not totally unexpected, this analysis revealed some surprises. First, a developmental switch towards regulation of different members within the same signaling pathway was detected. In the context of Wnt signaling this shifted Zic3 impact from the intracellular part of Wnt signaling towards extracellular ligands in this pathway. Second, that cells expressing Zic3 show a reduced level of transcription of proneural genes placed an impact of Zic3 on cells that are in a state either before or after neural differentiation. Zic3 has been linked with pluripotency of stem cells in mammals [80]. Whereas it is less likely that Zic3 positively regulates the proneural genes at 24 hpf, at the same time this does not exclude a possibility that it could be involved in this process (as suggested [2]) during earlier stages. Taken together, these observations suggest that functional relationship between Zic3 and its target gene could not be deduced from a simple one-to-one interaction model. Factors, such as the presence of different subsets of interacting partners or accessibility of certain binding sites as dictated by chromatin states, in different spatiotemporal contexts may affect transcriptional output.
One implication of an interactive regulatory landscape is that genes targeted by a particular TF may not be determined by simply observing binding of the TF near its genomic locus. Additional proof, such as responsiveness of the particular target gene to LOF of the TF, would be necessary. In our data, there is a surplus of Zic3 binding events compared to those associated with responsive target genes. Widespread binding of TFs exceeding their known target genes have been reported in cell culture and in Drosophila [81]–[87] and is suggestive of non-functional binding. This may happen due to interaction of TFs with randomly occurring target sequences in the genome [78], [88]. The availability of expression data helps to identify candidate target genes within the vicinity of a TF binding event by providing additional functional cues. Nevertheless, given that TF-target genes interactions could occur over long distances [22], [89], [90], it is still possible that seemingly isolated Zic3 binding events with no responsive genes within a set distance criteria might actually be regulating a target located further away. Until a more detailed understanding of the architecture of genome-wide interactions have been achieved, this possibility could not be ruled out.
The highly interconnected TF regulatory network also necessitates a careful interpretation of enhancer function by reporter assays: while such assays can be useful to identify independently acting regulatory elements, evidence exists for regulatory elements acting in tandem, resulting in higher transcriptional output [24], [62], [63], [76], [77]. While other possibilities such as non-functional occupancy and repressive interactions could not be ruled out, the TF interaction model could account for the inactivity of several of the tested enhancers inferred from the reporter assay. The occurrence of Zic3 consensus motifs in close proximity to 50% of peaks containing Gli consensus motif supports this idea. Interestingly, the presence of Gli motifs does not seem to be specific to a particular developmental stage, as both 8 hpf and 24 hpf data show similar proportions of Zic3 peaks containing Gli motifs nearby. As in vitro data have demonstrated physical and functional interactions between Zic and Gli proteins [40], [41], such interaction, as well as interactions with other binding partners, may also occur in vivo in regulating transcription of target genes.
Our identification of novel target genes of Zic3 has improved an understanding of the mechanism by which Zic3 regulates development. These results demonstrated that Zic3 inhibits Nodal signaling (either directly or indirectly) which is implicated in mesendodermal specification [15], [44]–[46]. Similarly, Lim and colleagues [38] observed that murine ES cells acquired endodermal fate upon Zic3 knockdown, which supported an idea that Zic3 acts as an inhibitor of endodermal fate. Coincidentally, Nodal and Wnt signaling is known to regulate gastrulation [91]–[94]. Their regulation by Zic3 therefore may account for the gastrulation defect observed in Zic3 morphants. On the other hand, proper midline development during gastrulation is essential for proper L-R patterning [15], [95], [96]. Therefore, an involvement of Zic3 in regulating gastrulation through Nodal and canonical Wnt per se could have been sufficient to ensure a proper L-R asymmetry. However, our results suggested that Zic3 may also regulate the non-canonical Wnt (PCP) signaling pathway which is implicated in ciliogenesis. Interaction of these signaling pathways culminates in the establishment of a proper embryonic L-R axis [97]–[102]. Therefore, we could not rule out the possibility of direct involvement of Zic3 in later events specific to L-R patterning. In this context, it is noteworthy that mkks was also found as one of the Zic3 targets (Table S5) which is implicated in both L-R patterning and C-E movements during gastrulation through interaction with vangl2 [103]–[106]. Therefore, the regulation of non-canonical Wnt signaling by Zic3 could be at a core of developmental events linking C-E movement and L-R patterning [10].
Our finding that Zic3 regulates genes implicated in proliferation of neural progenitors agrees with the idea that Zic3 has properties of a stem cell factor [38], [80]. A mode of Zic3 regulation of genes responsible for the proliferation of neural progenitors reconciles the role of Zic3 in both early neuroectodermal specification and later events of neurogenesis. In essence, it establishes a particular role of Zic3 (and possibly other Zic family members) as an important regulator of proliferation of neural progenitors [7]. This model challenges previous assumptions that Zic3 induces the expression of proneural genes shown in overexpression studies [18], and suggests that an activation of proneural genes could be a downstream consequence of Zic3 regulation of proliferation of neural progenitor at an earlier stage of neurodevelopment. Given that neurog1 expression was upregulated upon Zic3 knockdown, and Zic3 binding sites were found near neurog1, as well as other proneural genes such as neurod4 and ncam1a, Zic3 may have an additional direct role in neural differentiation as its inhibitor. This possibility is also supported by the downregulation of her9. This places Zic3 within a regulatory landscape of Notch signaling in support of an early hypothesis based on functional analysis of Zic1 [107].
Zebrafish of wild type (AB strain) and transgenic line SqET33 [28], [34] were maintained according to established protocols [108] following all the ethical practice recommended for fish maintenance. Embryos were staged according to standard morphological criteria [109].
Dechorionated 24 hpf transgenic embryos were deyolked in PBS by pipetting through the 1 ml pipette tip. Cells were dissociated with trypsin solution (0.05% trypsin and 0.2 mM EDTA) in PBS for 15 min at room temperature. To facilitate dissociation of cells, embryos were pipetted through the 200 µl pipette tip. Trypsin was inhibited with complete protease inhibitor cocktail (Roche) and cell suspension was filtered through a nylon mesh (40 µm Cell Strainer, BD Falcon). Immediately, an equal volume of 4% paraformaldehyde (PFA) in PBS was added to cell suspension and cells were fixed for 10 min at room temperature. Reaction was stopped by an equal volume of ice-cold 0.25 M glycine in PBS, cells were washed three times with 0.125 M glycine-PBS and resuspended in the same buffer. Cell sorting was carried out with FACSAriaII Cell Sorter (BD Bioscience). To set autofluorescence level, cell sorter was calibrated with PFA-fixed GFP-negative cells before cell separation. GFP-positive and GFP-negative cells were collected in 0.125 M glycine-PBS, frozen in liquid nitrogen and kept at −80°C until use. For microarray analysis, PFA fixation step was omitted and cells were sorted into complete L-15 Leibovitz medium (Gibco) containing 20% fetal bovine serum.
Chromatin Immunoprecipitation (ChIP) was performed according to an established protocol (Wardle et al., 2006) with an addition of a deyolking step according to Link and colleagues (2006), with modifications (see Text S1). ChIP DNA was sequenced on the Illumina Genome Analyzer II (Illumina, USA). Detailed ChIP-seq methods are described in Supplementary information. Sequencing reads were mapped to the zebrafish Refseq genome assembly (Zv9), following which peak finding was performed using the QuEST algorithm [110] using the following parameters: bandwidth = 30 bp, region size = 600 bp, and FDR q-value<0.01. Peaks mapped to unassembled chromosomal contigs, centromeric regions, telomeric regions, segmental duplications and peaks consisting of >70% repeat sequence were removed. The ChIP-seq data have been deposited in the Gene Expression Omnibus database under the accession number GSE41458. To validate the ChIP-seq performance, we carried out quantitative PCR (qPCR) on randomly selected peaks, 5 within promoter region and 16 at regions outside of gene promoters. Taking a fold-change of 2 as a cutoff for positive enrichment, the qPCR analysis validated all but one peak tested (Table S1).The Database for Annotation, Visualization, and Integrated Discovery (DAVID) [111], [112] and Genomic Regions Enrichment of Annotations Tool (GREAT) [35] was used to find gene ontology-enriched terms. Overlapping of 8 hpf and 24 hpf ChIP-seq signals around peaks was performed within a region of +/−2 kb from each peak summit. Notice that some peak regions in 8 hpf dataset were not detected as peaks in 24 hpf dataset but they could be having sufficient amount of ChIP-seq tags at 24 hpf because of true binding by Zic3. Similarly there were regions detected as peaks in 24hpf samples and not detected in 8hpf but they may be bound by Zic3 in both samples and be having enriched ChIP-seq tag count in both. Hence ChIP-seq signal based clustering further clarified the status of detected peaks. Motif search was performed with MEME de novo motif finder [113]. From the top 1000 peaks by statistical significance, we extracted sequences comprising +/−50 bp from the summit of each peak. After finding the similarity of de novo motif from MEME with other published Zic3 motifs [39], [80], the quantification of occurrence of these motifs was done on all ChIP-seq peaks. For this the sequences within 400 bp from the peak summit were matched with PWM of motifs and the best matching score were calculated. After having the best matching score a threshold was used to determine the presence of motif. The PWM-matching threshold value for each motif was calculated using simulation such that when 10000 sequences were randomly designed to have probability similar to corresponding nucleotides in its PWM then 85% of those sequences could be detected. CNE peaks were identified by comparing the 8 hpf ChIP-seq dataset against a list of known CNEs in ANCORA database [64]. We performed the comparison to both human and Tetraodon CNE database to take into consideration the genome duplication event during teleosts evolution, which relaxed selection pressure on the conservation of important developmental enhancers [68], [72].The genomic coordinates of each peak summit were extended by 500 bp on each side and compared against the genomic coordinates of CNEs identified through comparison with either human hg19 or Tetraodon tetNig2 assemblies. A threshold of at least 70% sequence conservation within every 50 bp was used to define CNEs in each species.
Two recombinant constructs of the zebrafish Zic3 protein were produced, the full-length protein (Zic3_ORF) and the DNA-binding domain encompassing Zn-fingers 2 to 5 (Zic3_ZF2-5, amino acid residues 273–391). DNA sequences corresponding to each domains were PCR-amplified using the following primers: Zic3_ORF: 5′-GGG GAC AAG TTT GTA CAA AAA AGC AGG CTT CGA AAA CCT GTA TTT TCA GGG CAG CTT ACG TGA AAT TGC G CTC-3′ and 5′-GGG GAC CAC TTT GTA CAA GAA AGC TGG GTT TAC TCC ACC TGA AAA CGG ACT TG-3′; Zic3_ZF2-5: 5′-GGG GAC AAG TTT GTA CAA AAA AGC AGG CTT CGA AAA CCT GTA TTT TCA GGG CGC CTT CTT CAG ATA CAT GCG-3′ and 5′-GGG GAC CAC TTT GTA CAA GAA AGC TGG GTT TAT GAT TCG TGT ACC TTC ATA TG-3′. Each forward and reverse primer contained an attB recombination site overhang, with an additional Tobacco Etch Virus (TEV) protease cleavage site in the forward primer preceding the N-terminal Zic3 coding sequence. Protein expression and purification was performed as previously described (Lim et al., 2010). Electrophoretic mobility shift assay (EMSA) was performed as previously described [38]. Briefly, Cy5-labeled oligonucleotide pairs (1st BASE, Singapore) were annealed by heating to 95°C for 5 minutes in annealing buffer (500 mM MgCl2; 500 mM KCl; 200 mM Tris-HCl, pH 8.0) and left in room temperature to cool down overnight. These were subsequently incubated with the recombinant Zic3 in EMSA buffer (10 mM Tris, pH 8.0; 0.1 mg/ml BSA; 50 µM ZnCl2; 100 mM KCL; 0.5 mM MgCl2; 10% glycerol, 0.1% SDS; 2 mM β-mercaptoethanol) for 1 hour at 4°C. The reaction was subsequently run on 5% native Tris-Glycine polyacrylamide gel electrophoresis. Gel was scanned in Typhoon Scanner (GE Healthcare, USA). The affinity of protein to DNA was determined by titrating 0–250 nM of protein against 1 nM of annealed probes.
Zic3 knockdown was performed using a translation-blocking antisense morpholino oligonucleotide (MO) purchased from Gene Tools, LLC (USA). The MO sequence was 5′-AGG TTA GTG GAG TGA ACG GGT ACC G-3′. A standard control antisense MO was also obtained from Gene Tools, LLC with the following sequence 5′-CCT CTT ACC TCA GTT ACA ATT TAT A-3′. For microarray, 1.7 ng Zic3 MO was injected into 1-cell stage embryos. Rescue was performed using 20 pg of zic3 mRNA without morpholino-binding site. Capped zic3 mRNA was synthesized using mMessage mMachine Kit (Ambion, USA). Results were obtained from at least three different experiments on embryos from random pairs.
For gene expression profiling, custom made zebrafish oligonucleotide microarray (Agilent Technologies; GIS V2 with some modifications) containing 44,000 oligonucleotide probes (60 mer long; including positive and negative controls designed by Agilent and beta-actin controls) was used. The microarray was performed according to Agilent's One-Color Microarray Based Gene Expression Analysis (Quick Amp Labeling) protocol (Version 5.7, March 2008) and RNA Spike-In-One Color. Arrays were probed using cDNAs reverse transcribed in the presence of Cy3-dUTP using 600 ng of total RNA from either wild-type control or Zic3 knockdown embryos (8 hpf), or from either non zic3-expressing cells or zic3-expressing cells (24 hpf). Labeled cDNA was denatured and hybridized at 42°C for 16 h in a hybridization oven (Agilent Technologies, USA). After hybridization, the slides were washed and scanned for fluorescence detection on Agilent DNA Microarray Scanner. Scanned images were analyzed using Agilent Feature Extraction Software (v10.5.1.1). Feature extracted data were analyzed in Genespring software (Agilent Technologies, USA). Statistically significant gene expression was identified using Significance Analysis of Microarrays (SAM 3.05) for each successive time point [114]. Threshold values were set as follows: q-value<0.8, predicted false discovery rate (FDR)<0.05%. Genes were annotated using the “Unigene & Gene Ontology Annotation Tool” available at GIS site (http://123.136.65.67/). Genes were subjected to pathway assembly using Ingenuity Pathway Analysis (IPA; http://www.ingenuity.com). Selected genes (Fig. 4C; Table S7) were validated using real time RT-PCR (qRT-PCR) by assessing their expression level changes in embryos injected with higher dose of morpholino (3.4 ng) to show similar trend with microarray regulation.
Tested genomic regions encompassing the peaks with ∼200 bp flanking sequence at each side were amplified using PCR (primer list in Additional file 5) and cloned into SalI and BamHI sites of the pTol2-GFP reporter vector containing a minimal promoter from the mouse cFos gene [115]. Transposase mRNA was synthesized using mMESSAGE mMACHINE T3 Kit (Ambion, USA) and purified using RNeasy Mini Kit (QIAGEN, Germany). A total of 20 pg of the circular reporter plasmid and 50 pg of transposase mRNA were co-injected into 1–2-cell stage embryos. For each construct, two batches of at least 100 embryos were injected and assayed for egfp expression at 24 hpf. A consistent egfp expression pattern observed in at least 20% of injected embryos was considered as positive. The reporter vector alone showed expression in muscles and blood cells in G0 embryos (data not shown). Embryos positive for egfp expression were subsequently processed for whole mount immunohistochemistry (IHC) with anti-GFP antibody. qPCR was used to determine egfp expression level at 8 hpf since morphological identification of tissue specificity at this stage was difficult.
|
10.1371/journal.pgen.1005226 | Burden Analysis of Rare Microdeletions Suggests a Strong Impact of Neurodevelopmental Genes in Genetic Generalised Epilepsies | Genetic generalised epilepsy (GGE) is the most common form of genetic epilepsy, accounting for 20% of all epilepsies. Genomic copy number variations (CNVs) constitute important genetic risk factors of common GGE syndromes. In our present genome-wide burden analysis, large (≥ 400 kb) and rare (< 1%) autosomal microdeletions with high calling confidence (≥ 200 markers) were assessed by the Affymetrix SNP 6.0 array in European case-control cohorts of 1,366 GGE patients and 5,234 ancestry-matched controls. We aimed to: 1) assess the microdeletion burden in common GGE syndromes, 2) estimate the relative contribution of recurrent microdeletions at genomic rearrangement hotspots and non-recurrent microdeletions, and 3) identify potential candidate genes for GGE. We found a significant excess of microdeletions in 7.3% of GGE patients compared to 4.0% in controls (P = 1.8 x 10-7; OR = 1.9). Recurrent microdeletions at seven known genomic hotspots accounted for 36.9% of all microdeletions identified in the GGE cohort and showed a 7.5-fold increased burden (P = 2.6 x 10-17) relative to controls. Microdeletions affecting either a gene previously implicated in neurodevelopmental disorders (P = 8.0 x 10-18, OR = 4.6) or an evolutionarily conserved brain-expressed gene related to autism spectrum disorder (P = 1.3 x 10-12, OR = 4.1) were significantly enriched in the GGE patients. Microdeletions found only in GGE patients harboured a high proportion of genes previously associated with epilepsy and neuropsychiatric disorders (NRXN1, RBFOX1, PCDH7, KCNA2, EPM2A, RORB, PLCB1). Our results demonstrate that the significantly increased burden of large and rare microdeletions in GGE patients is largely confined to recurrent hotspot microdeletions and microdeletions affecting neurodevelopmental genes, suggesting a strong impact of fundamental neurodevelopmental processes in the pathogenesis of common GGE syndromes.
| Epilepsy affects about 4% of the general population during lifetime. The genetic generalised epilepsies (GGEs) represent the most common group of epilepsies with predominant genetic aetiology, accounting for 20% of all epilepsies. Despite their strong heritability, the genetic basis of the majority of patients with GGE remains elusive. Genomic microdeletions constitute a significant source of genetic risk factors for epilepsies. The present genome-wide burden analysis in 1,366 European patients with GGE and 5,234 ancestry-matched controls explored the role of large and rare microdeletions (size ≥ 400 kb, frequency < 1%) in the complex genetic architecture of common GGE syndromes. Our results revealed a 2-fold excess of microdeletions in GGE patients relative to the population controls, 2) a 7-fold increased burden for known hotspot microdeletions (15q11.2, 15q13.3, 16p13.11, 22q11.2) previously associated with a wide range of neurodevelopmental disorders, and 3) a more than 4-fold enrichment of microdeletions carrying a gene implicated in neurodevelopmental disorders. Our findings reinforce emerging evidence that genes affected by microdeletions in GGE patients have a strong impact in fundamental neurodevelopmental processes and dissect novel candidate genes involved in epileptogenesis.
| The epilepsies comprise a clinically heterogeneous group of neurological disorders defined by recurrent spontaneous seizures due to paroxysmal excessive and synchronous neuronal activity in the brain [1]. Epilepsy affects about 4% of the general population during their lifetime [2] and about 40% of all epilepsies are thought to have a strong genetic contribution. The genetic generalised epilepsies (GGEs) represent the most common group of epilepsies with predominant genetic aetiology, accounting for 20% of all epilepsies [3]. Their clinical features are characterised by unprovoked generalised seizures with age-related onset, generalised spike and wave discharges on the electroencephalogram and no evidence for an acquired cause [4,5]. Despite their strong familial aggregation and heritability [6–9], the genetic architecture of common GGE syndromes is likely to display a biological spectrum, in which a small fraction (1–2%) follows monogenic inheritance, whereas the majority of GGE patients presumably display an oligo-/polygenic predisposition with extensive genetic heterogeneity [10]. Although causative mutations for rare GGE with monogenic inheritance have been identified in genes primarily affecting neuronal excitability, synaptic transmission, and neurodevelopmental processes [11,12], the genetic basis of the majority of patients with GGE remains largely unsolved.
Genomic copy number variations (CNVs) constitute a significant source of genetic risk factors for common focal and generalised epilepsies [13–20]. By targeted screening of rearrangements at genomic hotspots associated with neurodevelopmental disorders [21], we previously identified recurrent microdeletions at 15q11.2, 15q13.3 and 16p13.11 as important genetic risk factors of common GGE syndromes [14,16,17,22–24]. The microdeletions at 15q13.3 and 16p13.11 represent the most prevalent genetic determinants of GGE identified so far [14,16]. In addition, we were able to show that non-hotspot exonic microdeletions in three brain-expressed genes encoding gephyrin (GPHN) [25], neurexin 1 (NRXN1) [26] and the RNA-binding protein FOX1 (RBFOX1) [27] confer susceptibility of GGE. Although the GGE-associated microdeletions identified to date are individually rare (<1%), they cumulatively account for a significant fraction of the genetic burden in more than 3% of patients with common GGE syndromes [14–16,22].
In the present genome-wide burden analysis, we used the Affymetrix SNP 6.0 array to screen large (≥ 400 kb) and rare (< 1%) autosomal microdeletions with high calling confidence (≥ 200 markers) in European case-control cohorts of 1,366 GGE patients and 5,234 population controls. We aimed to: 1) assess the genetic burden of large and rare microdeletions in common GGE syndromes, 2) evaluate the contribution of recurrent hotspot and unique microdeletions to the genetic burden of GGE, and 3) identify novel candidate genes for GGE. Specifically, we tested the hypothesis whether microdeletions affecting genes involved in neurodevelopmental processes account for a significant fraction of the genetic risk of GGE syndromes.
We identified 103 microdeletions in 100 out of 1,366 GGE patients compared to 214 microdeletions in 208 out of 5,234 controls (S1 Table). Overall, 7.3% of patients with GGE carried at least one microdeletion compared to 4.0% in controls (P = 1.77 x 10–7; OR = 1.91, 95%-CI: 1.48–2.46) (Table 1). We observed a marginal increase in microdeletion frequency in the GGE patients when we considered only microdeletions affecting either at least one protein-coding RefSeq gene (n = 18,299; P = 5.86 x 10–7; OR = 1.95, 95%-CI: 1.48–2.57) or at least one brain-expressed gene (n = 8,878; P = 1.38 x 10–7; OR = 2.19, 95%-CI: 1.61–2.98) (Table 1). Likewise, the median size of microdeletions was larger in the GGE patients (713 kb; interquartile range (IQR) = 523 kb—1,537 kb) compared to controls (589 kb; IQR = 488–930 kb; P = 3.99 x 10–3; Wilcoxon-Mann-Whitney-Test). The number of individuals carrying at least two microdeletions did not differ significantly between the GGE patients (n = 3) and controls (n = 6; P = 0.40, Fisher´s exact test). The microdeletion burden was similar for males (7.2%) and females (7.4%) affected by GGE (P = 0.91; OR = 0.97, 95%-CI: 0.63–1.52). The distribution of GGE subsyndromes did not differ between 100 GGE patients carrying a microdeletion (33 JME, 50 CAE/JAE, 17 EGTCS/EGMA) and the group of 1,266 GGE patients without a large and rare microdeletion (507 JME, 548 CAE/JAE, 211 EGTCS/EGMA; P > 0.15).
The spectrum of 103 microdeletions identified in 100 GGE patients comprised: 1) 38 (36.9%) recurrent microdeletions at seven known genomic rearrangement hotspots previously associated with a wide range of neurodevelopmental disorders [29], 2) 27 (26.2%) genic microdeletions that were detected only in the GGE patients, 3) 16 (15.5%) microdeletions without a protein-coding RefSeq gene and that were not present in the controls, and 4) 22 (21.4%) non-hotspot microdeletions which overlap with the microdeletions identified in the controls (S1 Table). Most prominent was the 7.5-fold excess of recurrent hotspot microdeletions in the GGE patients compared to the controls (P = 2.61 x 10–17; OR = 7.46, 95%-CI: 4.20–13.33; χ2-test, df = 1) (Table 2). Overall, 2.8% (n = 38) of 1,366 GGE patients carried one of the known recurrent microdeletions at 1q21.1 (n = 1), 15q11.2 (n = 13), 15q13.3 (n = 11), 16p11.2 (n = 1), 16p12 (n = 3), 16p13.11 (n = 6) and 22q11.2 (n = 3), whereas these hotspot microdeletions were observed only in 0.4% (n = 20) of 5,234 population controls (S1 and S2 Figs). Significant associations with GGE patients were found for single hotspot microdeletions at 15q11.2 (P = 4.18 x 10–4; OR = 3.58; 95%-CI: 1.58–8.09, χ2-test, df = 1), 15q13.3 (P = 2.89 x 10–8, Fisher´s exact test), 16p13.11 (P = 1.48 x 10–3; OR = 11.48, 95%-CI: 2.05–116.5, Fisher´s exact test), and 22q11.2 (P = 8.85 x 10–3, Fisher´s exact test). All hotspot microdeletions in GGE patients identified by SNP arrays were validated by TaqMan qPCR. Altogether, the present findings highlight the cumulative impact of the recurrent microdeletions at 15q11.2, 15q13.3, 16p13.11 and 22q11.2 on the genetic risk of common GGE syndromes.
Besides the recurrent hotspot microdeletions, we identified 27 GGE patients carrying a genic microdeletion that was not observed in the controls (Table 3 and S1 Table and S3 Fig). These microdeletions affected 158 protein-coding RefSeq genes and exhibited an enrichment of genes previously associated with epilepsy (NRXN1, RBFOX1, PCDH7, KCNA2, EPM2A, RORB, PLCB1) and neuropsychiatric disorders (DPYD, CADM2, PARK2, GRM8, TSNARE1, TPH2, MACROD2) (Table 3). Microdeletions involving NRXN1 exons 1–2 were observed in two GGE patients with genetic absence epilepsies [26]. In addition, two partially overlapping microdeletions were identified in the chromosomal region 8q24.3 encompassing the genes encoding the t-SNARE domain containing 1 protein (TSNARE1; chr8: 143,293,441–143,484,601) and the brain-specific angiogenesis inhibitor 1 (BAI1; chr8: 143,545,376–143,626,368). All other unique microdeletions occurred only once. The microdeletions affecting the neuronal genes, NRXN1 and RBFOX1, have been reported in two previous publications [26,27].
To explore the hypothesis whether neurodevelopmental genes affected by the microdeletions have an impact on the genetic risk of common GGE syndromes, we performed enrichment analyses of the deleted genes, using two previously published sets of genes implicated in neurodevelopmental disorders (ND): 1) ND-related genes (n = 1,547) compiled by literature and database queries [30], and 2) genes implicated in autism spectrum disorder (ASD-related genes) comprising 1,669 brain-expressed genes with an enrichment of deleterious exonic de novo mutations in ASD [31]. Microdeletions carrying at least one ND-related gene were 4.6-fold enriched in the GGE patients as compared to the controls (P = 8.02 x 10–18; OR = 4.58, 95%-CI: 3.09–6.82) (Table 1). Likewise, microdeletions encompassing at least one ASD-related gene showed a 4.1-fold excess in the GGE patients relative to the controls (P = 1.29 x 10–12; OR = 4.11, 95%-CI: 2.64–6.40) (Table 1). To explore the impact of neurodevelopmental genes that are not covered by the recurrent hotspot microdeletions, we combined the ND- and ASD-related gene lists [30,31] and removed all genes affected by observed recurrent hotspot microdeletions. Non-recurrent microdeletions carrying at least one of the 2,495 selected ND/ASD-related genes showed a 2.3-fold excess in GGE patients (n = 1,328) compared to control subjects (n = 5,214), when individuals with recurrent hotspot microdeletions were excluded (P = 4.56 x 10–4; OR = 2.48, 95%-CI: 1.42–4.30). To rule out an artificial enrichment of microdeletions in the GGE patients, we compiled two control gene assemblies comprising: 1) 3,256 randomly selected autosomal protein-coding RefSeq genes, and 2) 3,837 autosomal protein-coding RefSeq genes not expressed in the brain [28]. Both control gene assemblies did not show evidence for an increase of the microdeletion burden in GGE patients compared to controls (P > 0.40) (Table 1).
The Disease Association Protein-Protein Link Evaluator (DAPPLE v2.0) tool [85] was applied to identify significant physical connectivity among proteins encoded by genes affected by microdeletions. Therefore, we separately tested the gene assemblies for the GGE patients and the control subjects. Based on an initial regional query we extracted 191 seed genes from 103 microdeletions found in the GGE patients and 221 seed genes from 214 microdeletions observed in controls. There was an overlap of 61 genes between the two assemblies. DAPPLE network analyses revealed a significant enrichment for direct connections between the seed genes (P = 0.01) in the GGE microdeletion carriers, while the control gene network did not show evidence for an enrichment (P = 0.40). Finally, in GGE we found eleven genes with significant connectivity: PLCB1 (P = 0.002), GRM1 (P = 0.002), ARC (P = 0.002), CNTN6 (P = 0.015), CHL1 (P = 0.033), BAI1 (P = 0.033), CYFIP1 (P = 0.040), TRIP13 (P = 0.042), MAPK3 (P = 0.044), GJ8 (P = 0.048), and KCNA2 (P = 0.050) (S4 Fig).
Utilising the Enrichr tool [86], functional enrichment analysis of the gene assembly affected by the microdeletions in the GGE patients revealed a significant enrichment of the MGI Mammalian Phenotype term "abnormal emotion/affect behaviour" (MP:0002572; Padj = 1.30 x 10–3) and the GO biological process term “cognition” (GO:0050890; Padj = 0.012) (Table 4). Enrichr network analysis identified one significant PPI Hub in the GGE patients based on an enrichment of nine deleted genes (ARC, TJP1, MAPK3, MYH11, EXOC3, NRXN1, PARK2, PLCB1, GRM1) among 219 network genes (Padj = 0.018), for which GRIN2B encodes the shared interacting protein.
The present burden analysis applied a screening strategy that focused on both large (≥ 400 kb, ≥ 200 markers) and rare (< 1%) autosomal microdeletions to ensure a high calling accuracy [87] and to enrich pathogenic microdeletions among confounding benign copy number polymorphisms [88–90]. We found a significant 1.9-fold excess of microdeletions in the patients with GGE compared to the controls (Table 1). Overall, 7.3% of the 1,366 GGE patients carried at least one microdeletion compared to 4.0% in 5,234 controls. These findings highlight the important impact of microdeletions on the genetic susceptibility of common GGE syndromes with an attributable risk of about 3.3%.
The spectrum of 103 microdeletions identified in the GGE patients contained a high proportion (36.9%) of recurrent microdeletions at genomic rearrangement hotspots, known to play a pathogenic role in a wide range of neuropsychiatric disorders including epilepsy [13,91,92]. In total, 2.8% of the GGE patients carried one of the known pathogenic hotspot microdeletions at 1q21.1, 15q11.2, 15q13.3, 16p11.2, 16p12, 16p13.11 and 22q11.2 (Table 2), whereas these hotspot microdeletions were found only in 0.4% of the population controls (S1 and S2 Figs). Although these hotspot microdeletions are individually rare (< 1%), they collectively result in a 7.5-fold increased burden in the GGE patients and a population-attributable risk of about 2.4%. A previous genome-wide CNV search in epilepsies observed a similar cumulative prevalence of recurrent hotspot microdeletions in 3.5% out of 399 GGE patients [16]. Likewise, a targeted screening of the microdeletions at 15q11.2, 15q13.3 and 16p13.11 showed a cumulative frequency of 3.1% in 359 GGE patients and an even higher frequency of 10% in 60 GGE patients with intellectual disability [17]. Several other CNV studies targeting these genomic rearrangement hotspots also emphasised a substantial impact of recurrent microdeletions at 15q11.2, 15q13.3 and 16p13.11 in the pathogenesis of GGE and other epilepsies [14–20,22–24,93,94]. To our knowledge, this is the first study demonstrating a significant association of the recurrent microdeletion at 22q11.2 with GGE. Re-evaluation of the clinical records of three GGE patients carrying a 22q11.2 microdeletion revealed additional congenital and developmental features fitting to known conditions of the 22q11.2 deletion syndrome (OMIN 188400/192430). GGE patient (EC-EGMA094) had a moderate psychomotoric retardation, patient (EC-EGTCS145) was affected by a cleft palate and an atrial septal defect, and patient (EC-EGTCS044) had a mild impairment of his motoric coordination during childhood, moderate learning disabilities and hypocalcaemia, highlighting the 22q11.2 deletion syndrome as a multisystem disorder with high penetrance and variable phenotypic spectrum [95]. According to our ascertainment scheme [96], the present GGE patients with recurrent microdeletions did not exhibit severe intellectual disability or severe psychiatric comorbidities at the age of exploration but may evolve psychiatric disorders at later age. Considering the published CNV studies of epilepsies [14–20,24], meta-analyses may demonstrate an association of the less frequent recurrent hotspot microdeletions at 16p11.2 and 16p12 with GGE. Haploinsufficiency of CYFIP1 at 15q11.2 [97], CHRNA7 at 15q13.3 [98], NDE1 at 16p13.11 [99] and PRRT2 at 16p11.2 [100] has been implicated as risk-conferring mechanism for epilepsy and other neurodevelopmental phenotypes [88,89,91].
Functional-enrichment, pathway and network analyses showed significant connectivity of genes affected by microdeletions in GGE patients (S4 Fig) and a significant enrichment for the MGI Molecular Function category "abnormal emotion/affect behaviour" (MP:0002572) as well as the GO biological process term “cognition” (GO:0050890). The protein-protein interaction analyses highlight several genes that have been implicated in epileptogenesis (CYFIP1, GRIN2B, KCNA2, NRXN1, PLCB1) [14,16,26,39,74,75,97] and neurodevelopmental processes (ARC, GRM1, PARK2) [51,52,55,57–59].
In line with our neurodevelopmental hypothesis, we found a significant 4.6-fold excess of microdeletions carrying at least one ND-related gene [30] and a 4.1-fold enrichment of microdeletions affecting at least one ASD-related gene [31] in the GGE patients compared to the control subjects. In contrast, the two control gene assemblies did not show an increase of the microdeletion burden in GGE patients compared to controls (P > 0.40). Accordingly, the intriguing enrichment of ND- and ASD-related genes demonstrates that genes involved in neurodevelopmental processes play an important role in the epileptogenesis of common GGE syndromes. Notably, the moderate overlap of the previously published assemblies of ND- and ASD-related genes implicates a large number of neurodevelopmental genes contributing to the risk of common GGE syndromes and extensive genetic heterogeneity. The emerging overlap of gene-disrupting microdeletions and the rapidly evolving landscape of loss-of-function gene mutations in rare and common epilepsy syndromes will facilitate the prioritisation of causal epilepsy genes and the elucidation of the leading molecular pathways of epileptogenesis [101,102].
We identified 27 gene-covering microdeletions in non-hotspot genomic regions that were present only in GGE patients (Table 3 and S3 Fig). These autosomal microdeletions involved several genes previously implicated in epilepsy and neurodevelopmental disorders. Although it remains challenging to distinguish benign and pathogenic microdeletions, several of these contain plausible candidate genes for epilepsy. Of particular interest were seven genes at seven microdeletion loci that have been associated with epilepsy.
Three of the epilepsy-associated microdeletions have been reported in two previous publications demonstrating an association of microdeletions affecting the 5´-terminal exons of the neuronal genes encoding the adhesion molecule neurexin 1 (NRXN1; 2p16.3, chr2: 50,145,642–51,259,673, hg19) and the splicing regulator RNA-binding protein fox-1 homolog (RBFOX1; 16p13.3, chr16: 5,289,468–7,763,341, hg19) [26,27]. The microdeletions involving NRXN1 exons 1–2 were observed in two female GGE patients with genetic absence epilepsies [26]. The 5´-terminal untranslated RBFOX1 exons 1–2 were deleted in a female patient with childhood absence epilepsy [27]. Deleterious mutations and microdeletions of the genes, NRXN1 and RBFOX1, have been reported in a large number of patients with a broad range of neuropsychiatric disorders, who were frequently also affected by epilepsy [40,41,54,72,81]. A recent study demonstrated that the splicing regulator Rbfox1 controls neuronal excitation in the mammalian brain and the Rbfox1 knockout in mice results in an increased susceptibility to spontaneous and kainic acid-induced seizures [71]. Furthermore, molecular, cellular, and clinical evidence supports a pivotal role of RBFOX1 in human neurodevelopmental disorders [73,103].
A 3.45 Mb microdeletion harbouring the protocadherin PCDH7 gene (chromosomal location: 4p15.1, chr4: 30,721,950–31,148,422, hg19) was found in a female GGE subject with juvenile myoclonic epilepsy. An international GWAS meta-analysis including 8,696 epilepsy patients and 26,157 controls highlights PCDH7 as susceptibility gene for epilepsy in general and GGE syndromes in particular [45]. The PCHD7 gene encodes a calcium-dependent adhesion protein that is expressed in neurons of thalamocortical circuits and the hippocampus [46]. PCDH7 has been implicated as neuronal target gene of MECP2 [47], the gene for Rett syndrome (OMIM #312750), which manifests as a progressive neurodevelopmental disorder with recurrent seizures. Moreover, mutations in the X-chromosomal protocadherin gene PCDH19 cause epilepsy and intellectual disability in females [48]. These lines of evidence suggest an involvement of PCDH7 in epileptogenesis.
A 788 kb microdeletion involving the Shaker-like voltage-gated potassium channel gene KCNA2 (1p13, chr1: 111,136,002–111,174,096, hg19) was identified in a male GGE patient with generalised tonic-clonic seizures starting at the age of 14. The Kv1 subfamily plays an essential role in the initiation and shaping of action potentials, influencing action potential firing patterns and controlling neuronal excitability as well as seizure susceptibility [36,38,39]. De novo loss- or gain-of-function mutations in KCNA2 have been identified to cause human epileptic encephalopathy [39]. Furthermore, the Kcna2 knockout mice exhibit spontaneous seizures and have a reduced life span [35,37].
One female GGE patient with childhood absence epilepsy carried a 2.4 Mb microdeletion in the chromosomal region 6q24.6 encompassing two neuronally expressed genes encoding the metabotropic glutamate receptor type 1 (GRM1; chr6: 146,348,917–146,758,734, NM_001278065, hg19) and laforin (EPM2A; chr6: 145,946,439–146,056,991, NM_005670, hg19). Deleterious mutations in the GRM1 gene have been found in patients with schizophrenia [52]. Also, familial segregation analysis of deleterious non-synonymous sequence variants revealed a co-segregation with multiple neuropsychiatric conditions including epilepsy in some families. Recessive mutations/microdeletions of EPM2A cause progressive myoclonic epilepsy type 2A (Lafora disease, OMIM #254780) [53].
A 582 kb microdeletion encompassing exon 1 of the gene encoding the RAR-related orphan receptor B (RORB; 9q21.13, chr9: 77,112,251–77,303,533, NM_006914, hg19) was found in a male patient with childhood absence epilepsy, overlapping with the critical region of a novel microdeletion syndrome at 9q21.13 characterised by intellectual disability, speech delay, facial dysmorphisms and epilepsy [63]. The RORB gene is a strong candidate for the neurological phenotype because RORB was deleted in all affected individuals [63], it is expressed in the cerebral cortex and thalamus, and genetic associations of RORB with bipolar disorder [64] and verbal intelligence [65] have been reported.
The gene encoding the enzyme phospholipase C-beta 1 (PLCB1; 20p12.3, chr20: 8,112,911–8,865,546, hg19) was partially deleted (exons 1–3, NM_015192, hg19) in a male GGE patient with childhood absence epilepsy. PLCB1 catalyses the generation of inositol 1,4,5-trisphoshate and diacylglycerol from phosphatidylinositol 4,5-bisphosphate, a key step in the intracellular transduction of many extracellular signals. Homozygous microdeletions of chromosome 20p12.3, disrupting the promoter region and first three coding exons of PLCB1, have previously been reported in two consanguineous families with early infantile epileptic encephalopathy [74]. Mutation analysis of a family with severe intractable epilepsy and neurodevelopmental delay revealed compound heterozygous mutations in PLCB1 composed of a 476 kb microdeletion encompassing PLCB1 and a deleterious PLCB1 splice site mutation [75]. Girirajan et al. [54] found an enrichment of microdeletions and duplications involving the PLCB1 gene in individuals with autism. Together, these findings implicate that the PLCB1 gene contributes to the genetic risk of neurodevelopmental disorders including epilepsy.
In addition to the epilepsy-associated microdeletions, nine deleted genes have been previously implicated as genetic risk factors in a broad range of neuropsychiatric disorders. Unique hemizygous microdeletions in GGE patients involved DPYD/1p13.3 [32–34], CADM2/3p12.1 [43], BCHE/3q26.1 [44], PARK2/6q24 [54,55,57,58], GRM8/7q31.33 [59–61]. TSNARE1/8q24.3 [62], MPP7-ARMC4-MKX/10p12.1 [66], TPH2/12q21.1 [67–69], MACROD2/20p12.1 [78–81], and ADARB1/21q22.3 [83,84]. Notably, overlapping microdeletions encompassing TSNARE1 at chromosome 8q24.3 in two GGE patients indicate its potential role in epileptogenesis. A recent GWAS meta-analysis of psychiatric disorders identified TSNARE1 as susceptibility gene for schizophrenia, schizoaffective and bipolar disorders [62]. While the function of TSNARE1 remains elusive, bioinformatic predictions suggest a vertebrate-specific function in synaptic vesicle exocytosis [104]. Further studies will be necessary to disentangle the pathogenic genes and to elucidate their molecular pathways in neurodevelopmental disorders and epileptogenesis.
Our burden analysis of large and rare autosomal microdeletions (size ≥ 400 kb, frequency < 1%) revealed: 1) a nearly 2-fold excess of microdeletions in GGE patients relative to the population controls, 2) a 7-fold increased burden for known hotspot microdeletions previously associated with neurodevelopmental disorders, and 3) a more than 4-fold enrichment of microdeletions carrying a gene implicated in neurodevelopmental disorders. Recurrent microdeletions at seven genomic rearrangement hotspots accounted for 37% of all microdeletions identified in the GGE patients and predominantly contributed to the excess of microdeletions in GGE patients. Comorbidity of GGE with other neurodevelopmental disorders, such as intellectual disability, ASD and schizophrenia, may result in even higher prevalence of recurrent hotspot microdeletions [17] and emphasises a valuable diagnostic contribution to the clinical management of these severely affected comorbid patients with GGE. The remarkable phenotypic variability observed for the recurrent hotspot microdeletions suggests a shared susceptibility of a wide range of neuropsychiatric disorders and GGE [105]. Several genes affected by microdeletions that were found only in GGE patients highlight novel candidate genes for GGE. Altogether, the present findings reinforce converging lines of evidence that genes affected by microdeletions in GGE patients reside in fundamental neurodevelopmental processes.
The study protocol was approved by the local institutional review boards of the contributing clinical centres. All study participants provided written informed consent. Genomic DNA samples of all study participants were processed by the Affymetrix SNP 6.0 array. For the genome-wide CNV burden analysis, we did not include individuals with excessive CNV counts (> 50 autosomal deletions per individual for deletions spanning > 40 kb in size and covering > 20 markers). In addition, we excluded all Affymetrix SNP 6.0 array data derived from lymphoblastoid cell lines because of the clonal source of the DNA which is prone to CNV artefacts compared to genomic DNA samples derived from blood cells [21]. All study participants were of self-reported North-Western European origin.
Unrelated GGE patients of European descent were ascertained through the primary diagnosis of a common GGE syndrome according to the classification of the International League Against Epilepsy [1,4]. The standardised protocols for phenotyping of GGE syndromes as well as inclusion and exclusion criteria are available online at: http://portal.ccg.uni-koeln.de/ccg/research/epilepsy-genetics/sampling-procedure/ [96]. GGE patients with a history of severe major psychiatric disorders (autism spectrum disorder, schizophrenia, affective disorder: recurrent episodes requiring pharmacotherapy or treatment in a hospital), or severe intellectual disability (no basic education, permanently requiring professional support in their daily life) were excluded. The GGE cohort comprised 1,366 patients (853 females, 513 males) with the following age-related GGE syndromes: childhood absence epilepsy (CAE, n = 398), juvenile absence epilepsy (JAE, n = 191), unspecified genetic absence epilepsy (GAE, n = 9), juvenile myoclonic epilepsy (JME, n = 540), epilepsies with generalised tonic-clonic seizures (GTCS) alone predominantly on awakening (EGMA, n = 94), and epilepsies with recurrent unprovoked GTCS alone starting before the age of 26 (EGTCS, n = 134). These 1,366 GGE patients were collected from Austria (n = 142), Belgium (n = 39), Denmark (n = 97), Germany (n = 801) and the Netherlands (n = 287). Notably, 1,052 of the GGE patients and 3,022 population controls investigated in the present study were part of a previous study that investigated six target microdeletions at genomic rearrangement hotspots [14].
Affymetrix SNP 6.0 data from 5,234 German population controls (2,559 females, 2,675 males) were obtained from three epidemiologically based cohorts: 1) KORA cohort from South Germany (n = 1,507) [106], 2) PopGen cohort from North Germany (n = 1,143) [107], and 3) SHIP cohort from East Germany (n = 2,584) [108]. The population controls were unscreened for epilepsy or major neuropsychiatric disorders. EIGENSTRAT principal component analysis [109] was applied to remove ancestry outliers and to match for European ancestry of the case-control cohorts [96].
Genomic DNA samples were investigated by the Affymetrix Genome-Wide Human SNP Array 6.0 (Affymetrix, Santa Clara, CA, USA). CNV analysis was performed as previously described [14,22], using the Birdsuit algorithm implemented in the Affymetrix Genotyping Console version 4.1.1. All annotations refer to the genome build GRCh37/hg19. The present genome-wide burden analysis focused on rare and large autosomal microdeletions to ensure a high reliability of the microdeletion calls [87] and to enrich pathogenic microdeletions [88–90]. Therefore, we filtered out autosomal microdeletions with high calling confidence according to the following criteria: a) size ≥ 400 kb, b) coverage of ≥ 200 probe sets, and c) microdeletion frequency < 1% in the entire study sample. The microdeletion size of at least 400 kb was selected because all known pathogenic hotspot microdeletions identified in neurodevelopmental disorders exceed this size in CNV scans with the Affymetrix SNP Array 6.0 [29,88–90]. We did not include microduplications in the present burden analysis because the accuracy of CNV detection is lower for microduplications compared to microdeletions [110]. In particular, genomic DNA samples with substantial degradation are prone to spurious microduplication calls. Moreover, microduplications seem to exert pathogenic effects less frequently compared to microdeletions [88]. We excluded microdeletions with an overlap of > 10% with 12 chromosomal regions prone to artificial CNV calls according to a recently published "artefact list" [111]. For all QC-filtered microdeletions identified by SNP array screening, the segmental log2 ratios of the signal intensities and the SNP heterozygosity state were visually inspected by the Chromosome Analysis Suite v1.2.2 (Affymetrix, Santa Clara, CA, USA) to exclude spurious microdeletion calls. Validation of all 38 recurrent hotspot microdeletions and four GGE-associated microdeletions identified by SNP arrays in the GGE patients was carried out by real-time quantitative PCR (qPCR) according to the manufacturer´s instructions (Life Technologies, Carlsbad, CA, USA).
Overall burden analyses were carried out for three assemblies of autosomal microdeletions: 1) any microdeletion, 2) genic microdeletions encompassing at least one protein-coding RefSeq gene, defined by the largest NM gene transcript (n = 18,299, hg19), and 3) microdeletions affecting a brain-expressed gene (n = 8,878), specified by a log(RPKM) > 3.32 of the BrainSpan RNA-Seq transcriptome dataset (http://www.brainspan.org/) [28].
Specifically, we tested the hypothesis whether microdeletions affecting genes involved in neurodevelopmental processes account for a significant fraction of genetic risk of GGE syndromes. Therefore, we investigated two recently published assemblies of genes associated with neurodevelopmental disorders (ND): 1) ND-related genes compiling 1,547 genes that were associated with neuropsychiatric disorders, autism candidate genes and genes of known genomic disorders based on literature and database queries [30], and 2) ASD-related genes comprising 1,669 brain-expressed genes that were selectively enriched for deleterious exonic de novo mutations in ASD individuals relative to their healthy siblings [31]. To evaluate a spurious enrichment of microdeletions in the GGE patients relative to the population controls, we tested two control gene assemblies comprising: 1) 3,256 randomly selected autosomal genes, and 2) 3,837 autosomal genes not expressed in the brain [28], defined by the BrainSpan RNA-Seq transcriptome dataset. ND- and ASD-related genes, genes located in genomic rearrangement hotspots, or the artefact list were removed from the compiled control gene assemblies.
Functional-enrichment tests, pathway and network analyses were performed with the Disease Association Protein–Protein Link Evaluator version 2.0 program (DAPPLE v2.0; http://www.broadinstitute.org/mpg/dapple/dappleTMP.php; [85]) and the gene-set enrichment tool Enrichr (http://amp.pharm.mssm.edu/Enrichr/index.html; [86]). Therefore, we compiled two lists of genes affected by microdeletions in either the GGE patients (number of genes; n = 329; n = 191 regional seed genes) or the controls (n = 428 genes; n = 221 regional seed genes). There was an overlap of 103 genes (n = 61 seed genes) in both gene lists. To explore potential physical interactions among proteins encoded by deleted genes, DAPPLE uses experimentally validated, protein-protein interaction (PPI) databases to identify network and protein connectivity. Empirically, 1,000 random networks were generated by permutation to determine whether the connectivity of each seed protein with the PPI reference network was greater than that expected by chance.
The gene-set enrichment tool Enrichr was applied separately to explore patient and control lists of genes affected by microdeletions for an overlap with pathway gene-set libraries, specifically the database PPI Hub Proteins [112], and gene-set libraries created from Gene Ontology [113] as well as MGI Mammalian Phenotype terms [114]. A pathway or ontology term was considered as significantly enriched if the false discovery rate (FDR, Benjamini-Hochberg) was lower than 5% for an assembly of more than two genes and occurred only in the GGE patients but not in the controls.
Burden analysis was performed by comparisons of the frequency of autosomal microdeletions in GGE patients and controls. The P-values and corresponding odds ratios (ORs) with the 95%-confidence intervals were calculated with a two-sided χ2-test or Fisher´s exact test if appropriate. The Wilcoxon-Mann-Whitney-Test was applied to compare differences in the genomic size of microdeletions. In addition, the individual burden of microdeletions was assessed for comparisons of microdeletion size. Nominal two-sided P-values < 0.05 were considered significant.
|
10.1371/journal.ppat.1004512 | Crystal Structure of Cytomegalovirus IE1 Protein Reveals Targeting of TRIM Family Member PML via Coiled-Coil Interactions | PML nuclear bodies (PML-NBs) are enigmatic structures of the cell nucleus that act as key mediators of intrinsic immunity against viral pathogens. PML itself is a member of the E3-ligase TRIM family of proteins that regulates a variety of innate immune signaling pathways. Consequently, viruses have evolved effector proteins to modify PML-NBs; however, little is known concerning structure-function relationships of viral antagonists. The herpesvirus human cytomegalovirus (HCMV) expresses the abundant immediate-early protein IE1 that colocalizes with PML-NBs and induces their dispersal, which correlates with the antagonization of NB-mediated intrinsic immunity. Here, we delineate the molecular basis for this antagonization by presenting the first crystal structure for the evolutionary conserved primate cytomegalovirus IE1 proteins. We show that IE1 consists of a globular core (IE1CORE) flanked by intrinsically disordered regions. The 2.3 Å crystal structure of IE1CORE displays an all α-helical, femur-shaped fold, which lacks overall fold similarity with known protein structures, but shares secondary structure features recently observed in the coiled-coil domain of TRIM proteins. Yeast two-hybrid and coimmunoprecipitation experiments demonstrate that IE1CORE binds efficiently to the TRIM family member PML, and is able to induce PML deSUMOylation. Intriguingly, this results in the release of NB-associated proteins into the nucleoplasm, but not of PML itself. Importantly, we show that PML deSUMOylation by IE1CORE is sufficient to antagonize PML-NB-instituted intrinsic immunity. Moreover, co-immunoprecipitation experiments demonstrate that IE1CORE binds via the coiled-coil domain to PML and also interacts with TRIM5α We propose that IE1CORE sequesters PML and possibly other TRIM family members via structural mimicry using an extended binding surface formed by the coiled-coil region. This mode of interaction might render the antagonizing activity less susceptible to mutational escape.
| Research of the last few years has revealed that microbial infections are not only controlled by innate and adaptive immune mechanisms, but also by cellular restriction factors, which give cells the capacity to resist pathogens. PML nuclear bodies (PML-NBs) are dot-like nuclear structures representing multiprotein complexes that consist of the PML protein, a member of the TRIM family of proteins, as well as a multitude of additional regulatory factors. PML-NB components act as a barrier against many viral infections; however, viral antagonistic proteins have evolved to modify PML-NBs, thus abrogating this cellular defense. Here, we delineate the molecular basis for antagonization by the immediate-early protein IE1 of the herpesvirus human cytomegalovirus. We present the first crystal structure for the evolutionary conserved core domain (IE1CORE) of primate cytomegalovirus IE1, which exhibits a novel, unusual fold. IE1CORE modifies PML-NBs by releasing other PML-NB proteins into the nucleoplasm which is sufficient to antagonize intrinsic immunity. Importantly, IE1CORE shares secondary structure features with the coiled-coil domain (CC) of TRIM factors, and we demonstrate strong binding of IE1 to the PML-CC. We propose that IE1CORE sequesters PML and possibly other TRIM family members via an extended binding surface formed by the coiled-coil domain.
| Promyelocytic leukemia protein PML is the organizer of small nuclear matrix structures termed nuclear bodies (NBs) or nuclear domain 10 (ND10) [1]. PML, also named TRIM19, is a member of the tripartite motif (TRIM) family of proteins, which are characterized by the presence of RING, B-box and coiled-coil domains [2]. Recent studies showed that an unprecedented large number of TRIMs positively regulate innate immune signaling pathways by acting as E3-Ub ligases [3], [4]. Additionally, a subgroup of TRIMs, including PML, was demonstrated to exhibit small ubiquitin related modifier (SUMO) E3 activity and PML itself is covalently conjugated to SUMO on three lysine residues [5], [6]. This modification, which affects PML localization, stability and interaction with other partners, is critical for NB functions [7]. In response to stimuli, PML-NBs recruit a number of proteins implicated in different cellular processes such as DNA damage response, apoptosis, senescence and protein degradation [8], [9].
Accumulating evidence implicates this subnuclear structure as an important component of intrinsic immunity against viruses from different families including herpes-, adeno-, polyoma, rhabdo- and retroviruses [10]–[12]. Unlike the innate and adaptive immunity, the intrinsic immune response is mediated by cellular restriction factors that are constitutively expressed and permanently active, even before a pathogen enters the cell. Other characteristics of intrinsic immune mechanisms are that they are saturable and subject to viral countermeasures [13]. Besides PML, a number of NB components, such as Sp100, hDaxx and ATRX, function as cellular restriction factors. Recent evidence suggests that NB proteins independently contribute to the repression of herpesvirus replication, raising the concept that individual NB components, rather than the PML-NB structure as a whole, restrict viral infections [14]–[20]. Consequently, various viruses have been shown to antagonize the intrinsic cellular defense via the modification of NB proteins. For instance, the herpes simplex virus type I immediate-early protein ICP0 has been described as a viral ubiquitin ligase with preferential substrate specificity for SUMO-modified isoforms of PML thus promoting the degradation of PML [21]. However, no structural information on this and many other NB-antagonistic proteins is available, yet.
Human cytomegalovirus (HCMV), a ubiquitous beta-herpesvirus causing serious disease in immunocompromised individuals, encodes an abundant immediate-early protein termed IE1 that modulates innate immune mechanisms as well as other cellular processes (reviewed in [22]). Although IE1 is a major player in the initiation of lytic HCMV infection and has been subject to extensive studies over the last decades, structural data on this protein are still limited. Four distinct regions have been identified within the 491 amino acid IE1 protein: a short N-terminal region that is required for nuclear import, a large core domain, an acidic region near the C-terminus that harbors a SUMOylation site and a 16 amino acid chromatin-tethering domain (CTD) at the extreme C-terminus [23]–[26]. Recent results have suggested that the acidic C-terminal region of IE1 is characterized by a lack of well-defined three-dimensional structure, but contains a binding motif for signal transducer and activator of transcription (STAT) proteins. This interaction site enables IE1 to compromise STAT-mediated interferon signaling, thereby interfering with a crucial branch of the innate immune system and promoting viral replication [27]–[29]. In addition to its effects on the innate immune system, IE1 is required to overcome the PML-NB-mediated intrinsic immunity that targets HCMV immediately upon infection. IE1 transiently co-localizes with PML-NBs during the first 2–4 hours after infection but subsequently induces disruption of these structures [30]–[32]. NB dispersal correlates with the functional activities of IE1 during infection and a PML knock-down efficiently compensates for IE1 in promoting replication of an IE1-deficient virus, establishing IE1 as an important antagonist of PML-mediated cellular repression of viral replication [15], [16], [33]. Studies on the mechanism of NB dispersal have demonstrated that IE1 induces the loss of the SUMOylated forms of PML, and also influences the SUMOylation state of Sp100 [34], [35]. However, in contrast to ICP0, this neither requires proteasomal activity nor does IE1 affect the abundance of unmodified PML [24], [35]. In further studies, a physical interaction between IE1 and PML, which requires the N-terminal TRIM region of PML, has been detected as prerequisite for the transient co-localization and subsequent disruption of PML-NB integrity. The interaction site for PML has been mapped to the large core region of IE1, since deletions or mutations affecting this domain abrogate PML binding and NB disruption [23], [35], [36]. However, it was noted in several reports that mutations in the core region often result in unstable IE1 proteins, so that the molecular basis for the IE1-PML interaction remains uncharacterized [37], [38].
Here we report the crystal structure of the evolutionary conserved globular core domain of primate cytomegalovirus IE1 proteins, determined to 2.3 Å resolution. Unexpectedly, the overall structure does not resemble any known protein fold, but exhibits an unusual all α-helical, femur-like shape which shares secondary structure features recently observed in the coiled-coil domain of TRIM proteins. We show that this IE1CORE domain binds with high affinity to PML via the coiled-coil domain. This induces PML de-SUMOylation thus releasing the PML-associated factors hDaxx, Sp100 and ATRX, while PML accumulations itself are not dispersed. Since IE1CORE efficiently complements lytic replication of an IE1-deleted HCMV, we conclude that sequestration of PML via IE1CORE is sufficient for antagonization of NB-mediated intrinsic immunity. Thus, cytomegaloviruses may have evolved a distinct structural fold to effectively bind and neutralize an important cellular hub protein that exerts critical roles during the regulation of innate immune responses as well as the control of programmed cell death [8]–[10], [12].
In order to further clarify the mechanism of IE1-mediated PML antagonization, we investigated the molecular architecture of the IE1 proteins from human, chimpanzee and rhesus cytomegalovirus (h-, c- and rhIE1) (Figure S1). As previously proposed by Krauss et al. [28], in silico predictions using the web server IUPred [39] suggested that the N- and C-terminal regions of all IE1 proteins display consistently high intrinsic disorder propensities (Figure 1A). Based on these predictions as well as on sequence conservation and the characterization of protease-resistant IE1 subdomains, we generated truncated IE1 constructs covering the folded core (Figure 1B). Limited proteolysis of recombinant full-length hIE1 as well as of C- or N/C-terminally truncated hIE1 proteins confirmed the in silico predictions (Figure 1C). These studies revealed the existence of a stably folded IE1CORE domain of about 360 residues (Figure 1C, hIE1 20-382) that is flanked at the N- and C-termini by intrinsically disordered regions (IE1N-IDR and IE1C-IDR). Circular dichroism (CD) spectroscopy [40], [41] was applied to investigate the secondary structure composition of the IE1 variants. All hIE1 proteins produced typical α-helical spectra with negative ellipticity above 200 nm and two distinct minima at 208 nm and 222 nm (Figure 1D and S2). The spectra of full-length hIE1 and hIE1CORE differed in the region below 210 nm. Calculation of the difference spectrum revealed strong random coil characteristics thereby confirming the predicted predominantly disordered nature of the terminal regions (Figure S2).
Sequence identities between 24 and 73% between h-, c- and rhIE1CORE domains suggest that the core domains share identical folds. Indeed, the CD spectra of the recombinant proteins hIE1CORE, cIE1CORE and rhIE1CORE match extremely well and indicate that all core domains consist mainly of α-helical segments (Figure 1D).
Crystallization trials with full-length hIE1 remained unsuccessful and, in case of the IE1CORE variants, yielded suitable crystals only for rhIE1CORE after chemical methylation of surface exposed lysine residues. The structure of rhIE1CORE was solved using experimental phases and refined to 2.3 Å resolution (Rwork = 19.73%, Rfree = 24.96%) (Figures 2 and S3; Methods and Table S1). The main chain of the model was traced between amino acids 41 and 393. Since no well-defined electron density was visible for residues preceding residue 41 or following residue 393, we conclude that the core domain spans at least amino acids 42 to 392 of rhIE1 corresponding to residues 27 to 379 of hIE1. RhIE1CORE consists of a total of 11 α-helices (Figure 2 and S1). Helices H3 and H9 are unusually long and contain as many as 16 and 17 helical turns, respectively. RhIE1CORE adopts an elongated, femur-like shape with dimensions of 130×25×25 Å3.
The structure can be divided into three distinct regions, namely an N-terminal head region (rhIE1: residues 62–118, 236–283; hIE1 residues 46–103, 221–267) and a C-terminal head region (rhIE1 residues 151–207, 315–393; hIE1 residues 136–192, 300–380) interconnected by a stalk region (rhIE1 residues 41–61, 119–150, 208–235, 284–314; hIE1 residues 27–45, 104–135, 193–220, 268–299) (Figure 2). The stalk consists of an uncommon right-handed three-helix coiled-coil (α-helices H3, H6 and H9) with the N-terminal helix H1 added to one side of the three-helix bundle (Figure 2). The right-handed pairing of the helices goes in hand with the presence of hendecad repeats in the sequences of these helices [42]. In these repeats of 11 residues (numbered alphabetically abcdefghijk) the hydrophobic amino acids at positions a, d, and h are interspaced by 2 (bc), 3(efg) and 3(ijk) amino acids of predominantly polar nature (Figure S1). In contrast, the patterning of hydrophobic residues within the head regions of rhIE1CORE frequently resembles that observed in heptad repeats (residue labeling abcdefg). Here the hydrophobic amino acids at positions a and d are interspaced by 2 (bc) and 3(efg) polar residues, and the interdigitation of the a and d residues from neighboring helices gives rise to more commonly observed left handed coiled-coil supersecondary structure elements [42]. Hence, both head regions display left-handed coiled-coils. Whereas the N-terminal head region consists of a three-helix bundle (α-helices H3, H7 and H8) with an additional helix H2 added onto one side of this bundle, the C-terminal head-region comprises 5-helical segments (H4, H5, H9, H10 and H11) in total. These can be grouped into two pairs of antiparallel left-handed coiled coils (H4–H5 and H9–H10) and an additional C-terminal helix (H11). The coiled-coil helix pairs pack against each other with crossing angles of approximately 50°, and thus the interactions between these coiled-coils resemble the ridges into grooves side chain packing observed in globins [43].
RhIE1CORE also displays an extended loop region between helices H1 and H2 (residues 62 to 82) devoid of secondary structure elements. The conformation of this loop region is stabilized through extensive crystal packing contacts and differs between the two monomers of the dimeric unit as observed within the crystal (see below). Overall, the structure of rhIE1CORE is in full agreement with the observed CD spectrum. Since, with the exception of the very terminal helices (H1, H2, H10 and H11), all intervening helical segments span the entire length of the molecule, IE1CORE is described best as consisting of a single contiguous domain (Figure 2).
A search for structurally similar proteins revealed only partial hits that cover less than 50% of the total rhIE1CORE length. The top-scoring hits belong to a considerable variety of domain folds which either contain α-helical orthogonal bundles or up-down bundles that partly resemble the IE1 head or stalk region, respectively (Table S2). This indicates that IE1 cannot readily be assigned to any known topology and suggests that the overall fold of IE1 is so far unique.
An extended search for local structural similarities, which also considered multimeric proteins, revealed a similarity between IE1 and the recently described coiled-coil region of homodimeric TRIM25 (Figure 3) [44]. The coiled-coil region of TRIM25 is composed of three helices in which the long helices H1 and H1' (from the second monomer) align in an antiparallel fashion to form the TRIM25 dimer (Figure 3B). The H1/H1' helix pair can be superimposed onto rhIE1CORE such that the helices superimpose with the H5/H6 and H8/H9 helices of IE1. The crystal structure of TRIM25 is also highly similar to that of two further TRIM family members, namely TRIM69 [45] and TRIM5α [46]. These structures suggest that the coiled-coil topology may be conserved among the entire family of TRIM proteins, possibly extending to PML (TRIM19), the target protein of IE1. This is further corroborated by sequence comparisons, demonstrating that TRIM family members, including PML, exhibit a distinct pattern of heptad and hendecad repeats for helix H1 (Figure S4) [44].
Interestingly, the topological arrangement of helices H1 to H3 of rhIE1 closely resemble the topology of helices H1 to H3 in TRIM25 when allowing for an inversion of the sequential order of the helices (Figures 3C and D). This also extends to the joint presence of heptad and hendecad repeats in H1 in TRIM25 and H3 in rhIE1. Whereas in TRIM25, the hendecad repeats occur in the central segment of helix H1 and are flanked on both sides by heptad repeats, H3 in rhIE1 displays a number of heptad repeats towards its N-terminus and switches to a segment of hendecad repeats that covers the second half of helix H3 (Figures S1 and S4). Taken together, sequence comparisons and three available coiled-coil structures demonstrate that the pattern of heptad and hendecad repeats is highly conserved across the TRIM protein family, and is also present in the viral IE1 protein. These observations suggest a common architecture of TRIM coiled-coils and provide evidence for structural similarities between IE1 and TRIM proteins.
RhIE1CORE forms a dimer with C2 point group symmetry in the crystal (Figure 4A), and oligomerization of rhIE1 and hIE1 was confirmed both by gel filtration experiments and co-immunoprecipitation analyses (hIE1) (Figure 4B and C). RhIE1CORE dimerizes with both stalk regions juxtaposed in an antiparallel fashion (Figure 4A). The main-chain conformation differs in the two monomers (Cα-RMSD for all helical segments = 2.13 Å, Figure S5). This deviation originates from a pronounced kink that is observed in one of the two monomers and that causes a repositioning of the loop that interconnects helices H8 to H9 with a concomitant displacement of the N-terminal half of helix H9 (Figure S6). This kink solely occurs in the tetragonal crystal form whereas in the monoclinic space group all four monomers in the asymmetric unit are highly similar. Since this space group transition is triggered by the experimental dehydration of the crystals, we propose that this in situ molecular shaping reflects an inherent flexibility of the IE1CORE fold that allows for small readjustments in the packing of the helices.
Intermolecular contacts are formed along the entire rhIE1 length resulting in an extraordinary large interface area (Ø 3173 Å2 per molecule). In the stalk region, these contacts are predominantly hydrophilic, whereas several sparse hydrophobic patches are formed between head regions. Because of its predominantly hydrophilic nature, the dimer interface does not resemble interfaces typically observed in permanent oligomers, suggesting that the dimer could become disrupted upon interaction with binding partners.
This idea is further corroborated by the analysis of IE1CORE surface conservation indicating that the dimer interface is not higher conserved than the solvent exposed regions (Figure 4D). Evolutionary conserved surface patches are distributed almost over the entire surface of IE1CORE (Figure 4D, blue), whereas non-conserved patches are mainly restricted to loop regions (Figure 4D, red). This indicates that the overall biophysical properties are conserved within the IE1 family of proteins despite the rather low degree of sequence identity.
This is in line with the results of a molecular model generated for hIE1 based on the rhIE1 crystal structure (Figure 5A). This model exhibits a good global and local quality further indicating that hIE1CORE and rhIE1CORE adopt highly similar folds (Figure S7). Due to the observed structural conservation we asked whether rhIE1 is likewise capable of disrupting human PML-NBs. Infection of primary human fibroblasts with rhesus macaque cytomegalovirus (RhCMV) revealed an initial accumulation of rhIE1 at PML-NBs followed by a dispersal of NBs (Figure 5B). Furthermore, RhCMV infection resulted in a depletion of polySUMOylated PML species (Figure 5C) and the isolated rhIE1 expression was sufficient to redistribute PML (Figure 5D, lower panel) indicating that hIE1 and rhIE1 do not only share biophysical properties but also functional activities.
In order to study whether the function of hIE1CORE differs from that of full-length hIE1, we investigated the subcellular localization of the hIE1CORE (Figure 6A). For this purpose, primary human fibroblast cells (HFFs) were transfected with eukaryotic expression plasmids encoding full-length hIE1 or truncated hIE1CORE proteins. Surprisingly, while full-length hIE1 exhibited a dispersed nuclear localization and induced a loss of PML-aggregates, all truncated proteins showed a punctate staining pattern colocalizing with PML foci. Based on these data, we conclude that a region within the C-terminal hIE1IDR is necessary for PML-dispersal. In contrast, hIE1CORE alone was sufficient to induce deSUMOylation of PML in transient expression experiments using 293T cells (Figure 6B). Consistent results were obtained with a whole cell population of HFF cells stably expressing the hIE1CORE variant 1–382 (Figure 6C). Given that SUMO modification of PML is a prerequisite for the recruitment of other NB components like Sp100, hDaxx and ATRX, it was important to explore the subcellular localization of these factors after expression of hIE1CORE. Interestingly, while PML was detected in a dot-like pattern, Sp100, hDaxx and ATRX were released from NBs in the presence of hIE1CORE (Figure 6C). Taken together, these data demonstrate that hIE1CORE is sufficient to sequester and deSUMOylate PML resulting in the dissociation of other NB components.
Due to the accumulation of hIE1CORE at PML foci, it was attractive to speculate that the two proteins might strongly interact with each other, which was investigated by co-immunoprecipitation (Figure 7A). Intriguingly, while only a trace amount of PML was associated with full-length hIE1, PML was efficiently coprecipitated with hIE1CORE variants. An increased affinity of hIE1CORE for PML was also confirmed by yeast two-hybrid experiments (Figure 7B), which is in line with previous results by Lee et al. (2004) that show an enhanced interaction of PML with an IE1 variant lacking the acidic C-terminus (IE1 1–420) [35]. Having observed a structural similarity between IE1 and coiled-coil regions of TRIM proteins, we asked whether this domain of PML is required for binding of IE1. In a yeast two-hybrid analysis utilizing a series of C-terminal PML deletion mutants we observed that a truncation of the coiled-coil domain abrogated the interaction with IE1 (Figure 7C). To further confirm this finding, coimmunoprecipitation analyses were performed with additional N- and/or C-terminal PML deletions. Importantly, this experiment revealed that the coiled-coil domain of PML was sufficient to mediate an interaction with IE1 (Figure 7D, lower panel, lane 4). Furthermore, we observed that IE1 was also able to bind to TRIM5αsuggesting that IE1 targets additional TRIM factors via coiled-coil interactions (Figure 7E).
Having shown that hIE1CORE binds with high affinity to and deSUMOylates PML, but fails to disrupt PML accumulations, it was important to investigate whether this is sufficient to antagonize PML-NB mediated repression of viral infection. We constructed a recombinant HCMV expressing hIE1 lacking the C-terminal IE1IDR (Figure 8A) and could observe that this virus exhibited a severe defect to disperse PML after infection of HFFs (Figure 8B). Consistent with our results obtained after isolated expression of hIE1CORE, deSUMOylation of both PML and Sp100 was fully preserved (Figure 8C). Most importantly, however, the hIE1CORE-expressing virus replicated nearly as efficient as wild-type virus while an hIE1-deleted virus exhibited a severe growth defect (Figure 8D). The approximately 10fold growth reduction observed for the hIE1CORE-expressing virus at 4 and 6 dpi is in line with previously published results on viruses lacking the C-terminal acidic domain which binds STAT2 thus antagonizing the interferon response [28], [29]. This was also confirmed in a complementation experiment after infection of either hIE1- or hIE1CORE-expressing HFFs with an hIE1-deleted HCMV, finally demonstrating that hIE1CORE can efficiently substitute for full-length hIE1 during lytic HCMV infection (Figure 8E).
The immediate-early protein IE1 of human cytomegalovirus that directly binds to PML is known as an important herpesviral antagonist of PML-NB-mediated intrinsic immunity [15], [22], [33], [36]. However, the structural basis for its function has remained elusive due to the paucity of high-resolution structural information. Here, we present the first crystal structure for the evolutionary conserved primate cytomegalovirus IE1 proteins and demonstrate that a structurally conserved IE1CORE domain is sufficient to antagonize PML-mediated intrinsic immunity. The structure of IE1CORE consists of a femur-shaped bundle of helices, which surprisingly does not share any overall fold similarity with known protein structures. IE1CORE binds with high affinity to PML and efficiently abrogates PML SUMOylation, but fails to disrupt PML accumulations itself. Only upon inclusion of the C-terminal, intrinsically disordered region (IE1C-IDR) PML dispersal is observed. Thus, our study demonstrates that PML deSUMOylation can be discriminated from PML dispersal. Whereas the first activity is achieved by a distinctly folded IE1CORE domain, the second activity requires inclusion of C-terminal sequences of the IE1C-IDR region that is highly susceptible to proteolytic degradation and for which we did not observe any stable secondary structure formation. As it has been observed for many intrinsically disordered proteins, folding of the natively disordered IE1C-IDR region may occur upon binding to a specific interaction partner. First evidence for this comes from a recent study predicting that the chromatin-tethering domain (CTD) at the extreme C-terminus of IE1 forms a β-hairpin when bound to histone proteins [47]. Importantly, IE1CORE is able to release other NB-components like Sp100, hDaxx and ATRX into the nucleoplasm and this correlates with antagonization of NB-mediated repression. This shows that PML dispersal as observed during infection with herpes simplex virus type I and HCMV is not a prerequisite to antagonize the repressive effects of this cellular multiprotein complex on viral gene expression [24], [48]. As also suggested by recent findings on the γ-herpesviruses Herpesvirus saimiri and Kaposi sarcoma herpesvirus as well as on the polyomavirus BKV more subtle modifications like the release or degradation of individual NB-components appear to be sufficient [49]–[51].
IE1CORE resembles the action of the SUMO-specific protease SENP-1 which also abrogates the SUMOylation of PML but leaves most PML aggregated [52]. It was previously speculated that IE1 could harbor an intrinsic SUMO protease activity itself or recruit SUMO-specific proteases to the NBs [53]. However, our structural analysis of IE1CORE provides no evidence for the presence of a potential active site with hydrolase activity. Furthermore, in earlier studies no interaction of IE1 with SENPs could be detected, but it was reported that full-length hIE1 could still disassemble foci formed by a PML protein with all SUMOylation sites mutated [53]. Based on this observation, the authors raised the idea that SUMO-independent interference with PML oligomerization followed by exposure of SUMOylated PML to cellular SUMO proteases may account for NB disruption. However, the results of our study argue against such a scenario, since abrogation of PML SUMOylation by IE1CORE was detected while PML aggregates were still present. Thus, IE1 affects its targets via direct, SUMO-independent substrate interaction and this suggests that IE1 does not directly or indirectly act as a hydrolase that specifically targets SUMOylated PML.
Importantly, our study revealed structural similarities between IE1CORE and the crystal structure of the tripartite motif coiled-coil that appears to act as a critical scaffold organizing the biochemical activities of TRIM proteins [44], [45]. Moreover, we were able to confirm that the coiled-coil of PML is sufficient for strong binding to IE1. Increasing evidence suggests that TRIM proteins function as E3-ubiquitin ligases in agreement with the family-wide presence of several conserved domains, namely a RING domain followed by two B-boxes and a coiled-coil region [3]. Based on the recently solved crystal structure of the TRIM25 coiled-coil it was shown that TRIM proteins dimerize by forming interdigitating antiparallel helical hairpins that position the N-terminal catalytic RING domain at opposite ends of the dimer and the C-terminal substrate-binding domains at the center [44]. For some of the TRIM members, and among these PML, E3-SUMO instead of E3-ubiquitin ligase activity has been reported [5]. Thus, we would like to propose that IE1CORE, via its strong interaction with the PML coiled-coil, may inhibit an E3-SUMO ligase activity of PML that is required for auto-SUMOylation. Alternatively, IE1 binding to the coiled-coil might block the accessibility of PML for other components of the cellular SUMOylation machinery. Thus, the results of our study favor a model whereby IE1 primarily affects the on-rate of SUMO modification which is also supported by the slow kinetics of IE1-mediated loss of PML SUMOylation [54]. This is different from the ICP0 protein of herpes simplex virus type I which induces the rapid degradation of SUMO-conjugated proteins by acting as a SUMO-targeted ubiquitin ligase (STUbL) [21]. Similar to IE1, the adenoviral E4-ORF3 protein which has been shown to form a multivalent matrix via extensive self-interactions, appears to inactivate PML via tight binding [55]. This specific assembly of E4-ORF3 creates avidity-driven interactions that capture PML as well as other tumor suppressors thus disrupting PML bodies. However, in contrast to IE1, the recently solved crystal structure of E4-ORF3 revealed the molecular mechanism of multimerization, but not the exact mode of PML recognition [55].
In this context it should be noted that the nonstructural NS1 protein of influenza A virus has also been shown to target a TRIM protein, TRIM25, via interaction with the coiled-coil domain to inhibit its E3 ligase function [56]. Since TRIM25 catalyzes a critical ubiquitination of the viral RNA sensor RIG-I this constitutes a mechanism by which influenza virus inhibits the host IFN response. Interestingly, we detected that IE1CORE not only binds to PML but also to TRIM5α and a recent publication reported an interaction with TRIM33 [57]. Thus, the unique structure of IE1core may have developed during evolution to target an extended spectrum of members of the TRIM family via the conserved coiled-coil domain of these factors [44], [45]. This is also supported by our analysis of evolutionary conserved surface patches of IE1CORE. When assuming that sites of protein-protein interaction are reflected by conserved surface patches, our observation that conserved residues are distributed evenly over the entire IE1CORE protein surface suggests that rather large parts of the IE1 surface are involved in recognition of the PML coiled-coil. Consequently, the helical structure of IE1CORE might have evolved as a decoy that, by means of extensive helix-helix interactions might either pair up with the coiled-coil region of PML or substitute for one of the PML monomers within the PML dimer interface. In this respect, the similarity between the topological arrangement of helices H1 to H3 of IE1 and of predicted helices H3 to H1 of PML in combination with the joint occurrence of regions with extended hendecad repeats might facilitate the formation of heteromeric assemblies. The formation of extended coiled-coil interactions would also readily offer an explanation for the finding that single mutations within the conserved surface patches of IE1 only moderately affect its interaction properties with PML. In contrast, mutations affecting the overall tertiary structure (e.g. IE1 L174P) abrogate the functionality.
Furthermore, it agrees with our observation that rhIE1 can substitute for hIE1 during infection of human cells despite low overall sequence identity. Thus, the size and unique elongated fold of the IE1CORE could have developed during evolution to accommodate efficient binding of PML and possibly other TRIM factors via an extended surface involving coiled-coil interactions. This feature might render the interaction less amenable to mutational escape.
All variants of h-, c- and rhIE1 were recombinantly produced in E.coli strain BL21(DE3) (Novagen) as GST-tagged fusion proteins for in vitro experiments and crystallization. LB media (Carl Roth GmbH + Co. KG, Karlsruhe, Germany) were inoculated with freshly transformed E. coli colonies, and cell cultures grown at either 30° or 37°C. Seleno-methionine labeling of rhIE1(residues 36–395) was achieved by incubation of the cells with non-inducing PAG medium (pre-culture) and auto-inducing PASM-5052 medium (main culture). Cell pellets were resuspended in phosphate buffer and lysed by sonication. Protein purification was achieved by the following steps: a first affinity chromatography (Glutathione sepharose, GE Healthcare, Freiburg), proteolytic cleavage with PreScission protease, a second affinity chromatography and a final size exclusion chromatography (Superdex 200 prepgrade, GE Healthcare). The gel filtration column was pre-equilibrated in 25 mM Tris, 150 mM NaCl, 10 mM DTT, pH 7.5. The samples were separated with an isocratic gradient of 1.2 column volumes (CV) of the same buffer at a flow rate of 1.5 mL/min. The column was calibrated utilizing the elution peaks of thyroglobulin (670 kDa), bovine γ-globulin (158 kDa), chicken ovalbumin (44 kDa) and equine myoglobin (17 kDa) of the Bio-Rad gel filtration standard (Bio-Rad Laboratories, Munich, Germany). The molecular weight of the samples was determined by linear regression. The Kav coefficients of the standard proteins were plotted vs the logarithm of their molecular weights to obtain the calibration curve, with Kav = (Ve-V0)/(Vc-V0), where V0 is the column void volume, Ve is the elution volume and Vc is the geometric column volume. All purification steps were performed in the presence of 10 mM DTT. For the crystallization of variant rhIE1(36–395) the protein was chemically modified by lysine methylation prior to the final size exclusion chromatography step. A 1 mg/mL IE1 protein solution was incubated on ice with 20 µL of 1 M dimethylamine borane (DMAB) and 40 µL of 1 M formalin per mL of IE1 solution. After two hours the addition of DMAB and formalin was repeated and, following an additional two-hour incubation, 10 µL of 1 M DMAB per mL of IE1 solution were added, and the solution was incubated at 4°C overnight. The reaction was quenched by adding 125 µL 1 M Tris/HCl, pH 7.5 per mL of IE1 solution, and the protein was stabilized by addition of 10 mM DTT [58]. Following the final chromatography step, the protein samples were concentrated to 20 mg/ml and stored at −20°C in 25 mM Tris/HCl, 1.5 mM NaCl, 15 mM DTT, 1 mM EDTA, pH 7.4 before further usage.
Limited proteolysis was performed in order to probe the conformational architecture of the protein [59]. The assay was conducted at 21°C with protein concentrations between 0.2 and 0.5 mg/mL and 0.014 mU subtilisin (Sigma-Aldrich) per µg IE1 protein. Aliquots of 10 µL were taken at different time points, for example at 1 min, 10 min, 30 min, 60 min, 120 min, 180 min, 240 min and 300 min, mixed with 3.3 µL 4× SDS loading buffer and boiled at 95°C for 5 min to stop the cleavage reaction.
Circular dichroism spectra were recorded between 185 and 260 nm from protein samples containing 1.5 µM or 2 µM protein for full-length or truncated IE1, respectively. The measurements were performed in 10 mM potassium phosphate buffer, pH 7.5 with a Jasco J-815 spectropolarimeter (Jasco, Tokyo, Japan) at 20°C with standard sensitivity. The cuvette had a path length of 0.1 cm, the band width was 1.0 nm, the scan speed 20 nm*sec−1, data integration time 1 sec and the data pitch 0.1 nm. All measurements were accumulated ten times and corrected for the sample buffer. Conversion of the data to concentration-independent mean residual weight (MRW) ellipticities [θ]MRW was done as described previously [40].
Initial crystallization conditions were identified with a sparse matrix screening approach (Index Screen, Hampton Research, Aliso Viejo, USA) and a Phoenix protein crystallization robot (Art Robbins Instruments, Sunnyvale, USA) [60]. The crystallization conditions were optimized in a hanging drop vapor diffusion setup and involved microseeding. Crystals of diffraction quality were obtained by mixing 1 µL of protein solution with 1 µL of reservoir solution and equilibrating the droplet of 2 µL against 700 µL reservoir solution [0.4 M magnesium formate, 15% (w/v) PEG 3350]. The crystals were soaked in cryo-solution [0.4 M magnesium formate, 15% (w/v) PEG 3350, 15% (v/v) ethylenglycol or 20% (v/v) dimethyl sulfoxide (DMSO)] prior to flash-cooling in liquid nitrogen.
The crystal structure of rhIE1(36–395) was initially solved in space group P21 using the following diffraction datasets collected at 100 K at beamline BL14.1 at BESSY II synchrotron (Helmholtz Zentrum Berlin): a native dataset 1 extending to 2.85 Å resolution, a MAD dataset (peak, inflection point, remote) recorded from a gold-soaked crystal diffracting to 3.5 Å and a peak dataset from a seleno-methionine derivatized protein crystal diffracting to 3.1 Å resolution (Table S1). The gold-soaked crystal was prepared by incubating crystals for three days in cryo-solution containing DMSO and 2.5 mM KAu(CN)2. Before flash cooling, the crystals were back-soaked in cryo-solution for several minutes to remove unspecifically bound heavy atoms. All monoclinic datasets are highly isomorphous. The Matthews coefficient is 2.96 (58.42% solvent) when assuming the presence of four rhIE1(36–395) molecules in the asymmetric unit [61]. The final refinement of rhIE1(36–395) was performed against a dataset with space group symmetry P43 extending to 2.3 Å (Table S1). The increase in resolution and concomitant space group change were obtained upon controlled dehydration of the monoclinic crystals from above with the HC1c crystal humidifier device at beamline BL14.3 at BESSY II (Helmholtz Zentrum Berlin). Crystals were first equilibrated against 98% relative humidity before decreasing the humidity in steps of 4% and 10 min equilibration time to a final value of 86% humidity. Upon observation of an increase in diffraction power, the crystals were flash cooled and transferred to beamline BL14.1 for the recording of a complete dataset extending to 2.3 Å resolution (Table S1). All diffraction datasets were processed with XDS and scaled with XSCALE [62].
Initial protein phases were derived for the monoclinic crystal form using the MAD dataset collected from a gold-soaked crystal (Table S1). The positions of the gold atoms could be readily located with program SHELXD [63]. The non-crystallographic symmetry (NCS) relationship between the 4 monomers, i.e. the presence of two IE1 dimers with C2 point group symmetry, became apparent upon visualization of the gold positions in program COOT and the inspection of the initial electron density maps calculated with program SHELXE [64], [65]. The NCS relationship was corroborated by the self-rotation function, calculated with program POLARRFN from the CCP4 program suite [66]. The quality of these initial electron density maps could be significantly improved upon phase calculation with program SHARP/autoSHARP [67], [68] and density averaging with program DM [69]. The improved phases also allowed for the identification of the selenium positions in the peak dataset of the seleno-methionine-labeled protein crystal and the inclusion of this dataset in the calculation of the experimental protein phases. An initial atomic model covering a single monomer was then manually built starting from protein fragments derived with program autoSHARP and following the lead of electron density maps calculated with either program autoSHARP or MLPHARE/DM [66], [68]. The registration of the protein sequence was obtained from the shape of the local electron density and the positions of the selenium atoms as visualized by an anomalous difference map. These considerations also showed that one gold cation is bound via a free cysteine side chain in each IE1 monomer chain. The model was then stepwise completed, extended to four molecules in the asymmetric unit and refined with program PHENIX [70]. Convergence of the refinement at 2.8 Å in space group P21 was facilitated upon inclusion of NCS weights and secondary structure restraints in program PHENIX [70]. The final model of rhIE1(36–395) was obtained after transferring the monoclinic model into the tetragonal unit cell with program PHASER and upon refinement against the 2.3 Å dataset in space group P43 (Table S1) [71]. Refinement converged at crystallographic R-factors of 19.73% (Rwork) and 24.96% (Rfree). The space group change that took place upon dehydration can be easily explained by small readjustments in the packing of the IE1 dimers in the crystals.
Coordinates and structure factors for the IE1CORE structure have been deposited in the Protein Data Bank under accession code 4WID.
The size of the dimerization interface between the two rhIE1CORE monomers was calculated with the program PISA [72]. The reported value is the average of the buried surface of both chains. The rhIE1CORE monomers shown in Figure S5 were superposed with the program LSQKAB [66]. Only the α-helical segments of the protein as defined in Figure S1 were superimposed. Comparative modeling of hIE1 was performed with MODELLER 9.9 [73] and the resulting model was validated using ProSA [74], [75]. Searches for structurally similar proteins were performed with PDBeFold [76]. Since standard parameters did not result in any hits, the following search options were set. (i) The threshold of 70% for the lowest acceptable match in query and target was reduced to 30% and 60%, respectively. (ii) The search was extended to proteins with a different connectivity of their secondary structure elements. Searches were performed independently for chain A and chain B of rhIE1, and the list of hits was merged. For reasons of clarity, duplicate hits and hits related closely in sequence (>90% identity) were removed from the list. The normRMSD was calculated according to the following equation [77]: normRMSD = [RMSD • max(L1,L2)]/Naln. Where RMSD is the root mean square deviation of the superposition of query and target, max(L1,L2) is the number of amino acids of the largest chain in the superposition, and Naln is defined by the number of structurally equivalent residue pairs. Sequence conservation was calculated based on a Blosum30 matrix using the MultiSeq Plugin [78] of VMD [79]. The IE1 sequences from the following viruses served as input: human CMV (strain AD169), Rhesus-CMV, Baboon-CMV, Simian-CMV, and Panine-HV2/Chimpanzee CMV (Uniprot-accession-numbers P13202, Q2FAE9, D0UZW7, Q98682, Q8QRY6). Prediction of intrinsically disordered regions in hIE1, cIE1 and rhIE1 was performed with IUPred [39] using the prediction type “short disorder”. The disorder tendencies in the three IE1 homologs were plotted in one diagram using the rhIE1 amino acid numbering. Multiple sequence alignment was performed with TCoffee (http://www.ebi.ac.uk/Tools/services/web/toolform.ebi?tool=tcoffee).
The oligonucleotide primers used for this study were purchased from Biomers GmbH (Ulm, Germany) and are listed in Table S3. All prokaryotic expression plasmids were generated by PCR amplification of the respective codon-optimized IE1 sequences and subsequent cloning into pGEX-6P-1 (GE Healthcare Bio-Sciences AB, Uppsala, Sweden). The synthetic, codon-optimized hIE1 cDNA (strain AD169) was obtained from Mr. Gene GmbH (Regensburg, Germany). The codon-optimized cDNAs of cIE1 (NP_612746.1) and rhIE1 (Q2FAE9) were synthesized by GENEART gene synthesis service (Regensburg, Germany). The eukaryotic expression plasmids encoding full length or truncated hIE1 were generated via PCR amplification of the respective fragments using pHM494 [80] as template, followed by insertion into pHM971 (pcDNA3.1-FLAG) [80], pHM1580 (pcDNA3.1-Myc) [80], or into the yeast expression vectors pGBT9 and pGAD424 (Clontech, Mountain View, CA). The synthetic gene coding for rhIE1 (Q2FAE9) was obtained from GENEART gene synthesis service (Regensburg, Germany). The rhIE1 coding sequence was subcloned into pHM971 (pcDNA3.1-FLAG) [80] using BamHI and XhoI. Full length PML, isoform VI, and truncated PML variants were amplified from pAS-PML (a gift from G.G. Maul, Philadelphia, USA) and inserted into pHM1580 (pcDNA3.1-Myc) [81], pHM971 (pcDNA3.1-FLAG) [80], pHM972 (pcDNA3.1-FLAG-NLS), or yeast expression vectors pGBT9 and pGAD424 (Clontech, Mountain View, CA). The eukaryotic expression plasmid encoding rhesus TRIM5α was a gift from T. Gramberg (Erlangen, Germany). For transduction experiments, hIE1 variants were amplified utilizing pHM494 [80] as template and inserted into a pLKO-based lentiviral vector (a gift of R. Everett, Glasgow, UK).
HEK293T cells and primary human foreskin fibroblast (HFF) cells (obtained from Life Technologies) or telomerase-immortalized HFFs (HFFi) were cultured as described previously [54], [82]. HFFs were infected with either the HCMV laboratory strain AD169, a recombinant HCMV expressing hIE1 1–382 (AD169/hIE1 1–382), an IE1-deficient virus (AD169ΔhIE1), or rhesus macaque CMV (RhCMV) at specified multiplicities of infection (MOI). Titers of wild-type (wt) AD169 and the AD169 recombinants were determined by UL112/113 fluorescence. For this purpose, HFFs were infected with various dilutions of virus stocks. After 72 h of incubation, cells were fixed and stained with a monoclonal antibody directed against UL112/113. Subsequently, the number of UL112/113-positive cells was determined and was used to calculate viral titers. RhCMV was titrated via rhIE1 fluorescence, which was analyzed 24 h postinfection.
The AD169-based HCMV bacterial artificial chromosome (BAC) HB15 was used for recombination-based genetic engineering of AD169/hIE1 1–382 and AD169ΔhIE1. AD169/hIE1 1–382 was constructed by introducing a stop codon into the hIE1 gene replacing residue 383. For this purpose, the two-step red-mediated recombination technique was utilized [83], which uses the kanamycin gene as a first selection marker. The linear recombination fragment was generated by PCR using primers 5′BAC_short and 3'BAC_hIE1_382 (Table S3), and pEPkan-S (kindly provided by K. Osterrieder, Berlin) as template DNA. The PCR product was treated with DpnI, gel purified, and subjected to a second round of PCR amplification using primers 5′BAC_hIE1_382 and 3′BAC_hIE1_382_short (Table S3). For homologous recombination, the PCR fragment was transformed into Escherichia coli strain GS1783 (a gift of M. Mach, Erlangen) already harboring HB15, and bacteriophage λ red-mediated recombination was conducted as described elsewhere [83]. To identify positive transformants, the bacteria were plated on agar plates containing 30 µg/mL kanamycin (first recombination) or 30 µg/mL chloramphenicol and 1% arabinose (second recombination) and incubated at 32°C for 2 days. BAC DNA was purified from bacterial colonies growing on these plates and was further analyzed by PCR, restriction enzyme digestion and direct sequencing. For construction of the AD169ΔhIE1 BAC by homologous recombination, a linear recombination fragment, comprising a kanamycin resistance marker along with 5′ and 3′ genomic sequences, was generated by PCR amplification using pKD13 as template and primers 5′Intron3/pKD13 and 3′Exon 4/pkd13 (Table S3). This fragment was used for electroporation of competent Escherichia coli strain DH10B harboring HB15 and recombination was performed as described previously in order to delete exon 4 of the IE1 gene [84]. The integrity of the resulting recombinant BAC was confirmed by PCR, restriction enzyme digestion and direct sequencing.
For reconstitution of recombinant AD169, HFFs seeded in six-well dishes (3×105 cells/well) were cotransfected with 1 µg of purified BAC DNA, 0.5 µg of the pp71 expression plasmid pCB6-pp71, and 0.5 µg of a vector encoding the Cre recombinase using FuGENE6 transfection reagent (Promega, Mannheim, Germany). Transfected HFFs were propagated until viral plaques appeared, and the supernatants from these cultures were used for further virus propagation.
For the generation of HFF cells stably expressing full length hIE1 or hIE1 1–382, replication-deficient lentiviruses were generated using pLKO-based expression vectors. For this purpose, HEK293T cells seeded in 10 cm dishes (5×106 cells/dish) were cotransfected with a pLKO vector encoding either full length hIE1 or hIE1 1–382 together with packaging plasmids pLP1, pLP2, and pLP/VSV-G using the Lipofectamine 2000 reagent (Invitrogen, Karlsruhe, Germany). Viral supernatants were harvested 48 h after transfection, cleared by centrifugation, filtered, and stored at −80°C. Primary HFFs or telomerase-immortalized HFFs were incubated for 24 h with lentivirus supernatants in the presence of 7.5 µg/mL polybrene (Sigma-Aldrich, Deisenhofen, Germany). Stably transduced cell populations were selected by adding 500 µg/mL geneticin to the cell culture medium.
HFF cells were transfected with the DNA transfection reagent FuGENE6 (Promega, Mannheim, Germany). One day before transfection, 3×105 cells were seeded into six-well dishes. DNA content and transfection procedure were according to the instructions of the manufacturer. 48 hours after transfection, cells were harvested for further analyses. HEK293T cells were transfected by applying the standard calcium phosphate precipitation method. For this, 5×105 to 5×106 HEK293T cells were seeded into six-well dishes or 10 cm dishes one day before transfection. For Western blot analyses and coimmunoprecipitations, 1 to 10 µg of plasmid DNA were used for each transfection reaction. At about 16 hours later, the cells were washed two times with PBSo and provided with fresh medium. 48 hours after transfection, cells were harvested for further analyses.
Monoclonal antibodies used for immunofluorescence and Western blot analyses were: α-IE1 CH443 (Santa Cruz Biotechnology, Santa Cruz, CA, USA), α-UL112/113 M23, α-UL44 BS510 (kindly provided by B. Plachter, Mainz, Germany), α-UL69 69–66, α-FLAG M2 (Sigma-Aldrich, Deisenhofen, Germany), α-Myc 9E10, α-β-actin AC-15 (Sigma-Aldrich), α-PML PG-M3 (Santa Cruz). Polyclonal antibodies used for immunofluorescence and Western blot analyses were: α-rhesus IE1 (a kind gift from M. Mach, Erlangen, Germany), α-PML #4 (a kind gift from P. Hemmerich, Jena, Germany), α-PML H238 (Santa Cruz), α-PML A301–167A (Bethyl Laboratories, Montgomery, TX, USA), α-PML A301–168A (Bethyl Laboratories), α-Sp100 #2 (a kind gift from P. Hemmerich, Jena, Germany), α-Sp100 GH3 (kindly provided by H. Will, Hamburg, Germany), α-hDaxx C-20 (Santa Cruz), α-ATRX H-300 (Santa Cruz). Secondary antibodies used for immunofluorescence and Western blot analyses were: Alexa Fluor 488-/555-/647-conjugated secondary antibodies for indirect immunofluorescence experiments were purchased from Molecular Probes (Karlsruhe, Germany), horseradish peroxidase-conjugated anti-mouse/-rabbit secondary antibodies for Western blot analyses were obtained from Dianova (Hamburg, Germany).
HFF cells grown on coverslips in six-well dishes (3×105 cells/well) were washed twice with PBSo at 48 hours after transfection or at various times after virus infection. Cells were fixed with a 4% paraformaldehyde solution for 10 min at room temperature (RT) and then washed for two times. Permeabilization of the cells was achieved by incubation with 0.2% Triton X-100 in PBSo on ice for 20 min. Cells were washed again with PBSo over a time period of 5 min and incubated with the appropriate primary antibody diluted in PBSo-1% FCS for 30 min at 37°C. Excessive antibodies were removed by washing four times with PBSo, followed by incubation with the corresponding fluorescence-coupled secondary antibody diluted in PBSo-1% FCS for 30 min at 37°C. The cells were mounted using the DAPI-containing Vectashield mounting medium (Alexis, Grünberg, Germany) and analyzed using a Leica TCS SP5 confocal microscope, with 488 nm, 543 nm, and 633 nm laser lines, scanning each channel separately under image capture conditions that eliminated channel overlap. The images were exported, processed with Adobe Photoshop CS5 and assembled using CorelDraw ×5. In order to quantify PML-NB disruption in infected HFFs, 150 cells were analyzed for the presence of PML dots.
Lysates from transfected or infected cells were prepared in a sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) loading buffer, separated on sodium dodecyl sulfate-containing 8 to 15% polyacrylamide gels, and transferred to nitrocellulose membranes. Chemiluminescence was detected according to the manufacturer's protocol (ECL Western blot detection kit; Amersham Pharmacia Biotech).
Transfected HEK293T cells (1×106 or 5×106 were lysed for 20 to 40 min at 4°C in 800 µL of CoIP buffer (50 mM Tris-HCl [pH 8.0], 150 mM NaCl, 5 mM EDTA, 0.5% NP-40, 1 mM PMSF, 2 µg/mL of aprotinin, 2 µg/mL of leupeptin, and 2 µg/mL of pepstatin). After centrifugation, aliquots of each sample were taken as input controls and the remaining supernatant was incubated with anti-FLAG antibody M2 coupled to protein-A-sepharose beads for 2 h at 4°C. The sepharose beads were collected by centrifugation and washed five times in 1 mL CoIP buffer. Finally, the immunoprecipitated proteins were recovered by boiling in 4× SDS sample buffer and protein complexes were analyzed by SDS-PAGE and Western blotting.
Saccharomyces cerevisiae Y153 was used in a two-hybrid system. Both the plasmid pGBT9 (Clontech, Mountain View, CA) encoding the GAL4-DB (Trp+) fusion and the plasmid pGAD424 (Clontech, Mountain View, CA) encoding the GAL4-A fusion (Leu+) were introduced into Y153 cells using a modified lithium acetate (LiAc) method. For this, cells were grown overnight in YAPD medium, pelleted and treated with LP-mix (40% w/v PEG 4000, 0.15 M LiAc, 10 mM Tris/HCL pH 7.5, 1 mM EDTA pH 8.0) and DMSO. Single-stranded carrier-DNA as well as both plasmids were added to the yeast cells. This step was followed by incubation at room temperature and subsequent incubation at 42°C. Thereafter, the cells were plated on minimal selection agar lacking Trp and Leu. For rapid in situ assays of lacZ expression from yeast colonies, an XGal filter assay was used. Nitrocellulose filters were laid onto the plate and allowed to wet completely, then lifted off the plate and placed in liquid nitrogen to permeabilize the cells. The filters were removed and placed cell side up in a petri dish containing Whatman Paper soaked with Z buffer containing β-Mercaptoethanol and XGal. The filters were incubated at 30°C and constantly analyzed for the development of a positive blue color. For quantitation of the ß-galactosidase activity in the yeast cells three colonies were picked and grown in medium also lacking Trp and Leu. The next day, the optical density was measured at 600 nm. After pelleting the culture, the cells were resuspended in Z buffer and permeabilized using chloroform and 0.1% SDS. The ß-galactosidase activity within the cells was assayed by the standard method using o-nitrophenyl-ß-D-galactopyranoside (ONPG) as substrate. The reaction was stopped by adding Na2CO3 and the absorbance was measured at OD405. The unit of ß-galactosidase was defined as (1.000×OD405)/(t×v×OD600) (t, reaction time [min]; v, culture volume [mL]). The ß-galactosidase activity for each sample was corrected for background by subtracting the signal of the empty vectors.
HFF cells were seeded into six-well dishes at a density of 3×105 cells/well and infected the following day with wt AD169 at an MOI of 0.01 and equivalent genome copies of AD169/hIE1 1–382 and AD169ΔhIE1. Triplicate samples of the infected cell supernatants were harvested at 2, 4, 6, 8 and 11 days after inoculation and subjected to lysis by proteinase K treatment. Thereafter, all samples were analyzed for the amount of genome copy numbers by quantitative real-time PCR (TaqMan-PCR) using an Applied Biosystems 7500 Real-Time PCR System (Applied Biosystems, Foster City, CA, USA) together with the corresponding software SDS (sequence detection system) [85]. For quantification of the viral DNA load, a sequence region within the gB gene locus was amplified using primers 5'gB_forw and 3'gB_rev along with the fluorescence labeled probe CMV gB FAM/TAMRA also directed against the gB gene region. In parallel, the cellular DNA amount was quantified using primers 5′Alb and 3′ Alb together with a fluorescence labeled probe, Alb FAM/TAMRA, specific for the cellular albumin gene. Real-time PCR was performed in 96-well plates being compatible with the ABI Prism sequence detector. For the determination of reference CT values (cycle threshold), serial dilutions of the respective standards (107–101 DNA molecules of gB or albumin) were examined by PCR reactions in parallel. The 20 µL reaction mix contained 5 µL sample or standard DNA solution together with 10 µL 2× TaqMan PCR Mastermix (Applied Biosystems, Foster City, CA, USA), 1.5 µL of each primer (5 µM stock solution), 0.4 µL of probe (10 µM stock solution), and 1.6 µL of H2O. The thermal cycling conditions consisted of two initial steps of 2 min at 50°C and 10 min at 95°C followed by 40 amplification cycles (15 sec at 95°C, 1 min 60°C). The viral genome copy numbers and albumin copy numbers were subsequently calculated using the sample-specific CT value when set into relation to the standard serial dilutions.
For analysis of complementation by immunofluorescence staining, control HFFi cells as well as HFFi cells expressing hIE1 or hIE1 1–382 were infected with AD169ΔhIE1 at an MOI of 0.01. Triplicate samples of infected cells were fixed at 48 h post infection and subjected to immunostaining of UL44. Images of approximately 500 cells per sample were taken and the number of UL44-positive cells was determined via measuring of mean gray values using the ImageJ software (version 1.47). For analysis of complementation by Western blotting, normal HFFi cells as well as HFFi cells expressing hIE1 or hIE1 1–382 were infected with AD169ΔhIE1 at an MOI of 0.05 in triplicate, and were harvested 72 h later for detection of UL44, UL69, IE1 and β-actin.
|
10.1371/journal.pntd.0004451 | Mating-Induced Transcriptome Changes in the Reproductive Tract of Female Aedes aegypti | The Aedes aegypti mosquito is a significant public health threat, as it is the main vector of dengue and chikungunya viruses. Disease control efforts could be enhanced through reproductive manipulation of these vectors. Previous work has revealed a relationship between male seminal fluid proteins transferred to females during mating and female post-mating physiology and behavior. To better understand this interplay, we used short-read RNA sequencing to identify gene expression changes in the lower reproductive tract of females in response to mating. We characterized mRNA expression in virgin and mated females at 0, 6 and 24 hours post-mating (hpm) and identified 364 differentially abundant transcripts between mating status groups. Surprisingly, 60 transcripts were more abundant at 0hpm compared to virgin females, suggesting transfer from males. Twenty of these encode known Ae. aegypti seminal fluid proteins. Transfer and detection of male accessory gland-derived mRNA in females at 0hpm was confirmed by measurement of eGFP mRNA in females mated to eGFP-expressing males. In addition, 150 transcripts were up-regulated at 6hpm and 24hpm, while 130 transcripts were down-regulated at 6hpm and 24hpm. Gene Ontology (GO) enrichment analysis revealed that proteases, a protein class broadly known to play important roles in reproduction, were among the most enriched protein classes. RNAs associated with immune system and antimicrobial function were also up-regulated at 24hpm. Collectively, our results suggest that copulation initiates broad transcriptome changes across the mosquito female reproductive tract, “priming” her for important subsequent processes of blood feeding, egg development and immune defense. Our transcriptome analysis provides a vital foundation for future studies of the consequences of mating on female biology and will aid studies seeking to identify specific gene families, molecules and pathways that support key reproductive processes in the female mosquito.
| Female post-mating behavior has important consequences for mosquito populations and their ability to transmit diseases. Male Aedes aegypti seminal fluid substances transferred during mating cause many important changes to female behavior and physiology, including blood feeding behavior, egg development, and oviposition. In an effort to understand how males induce these responses in Ae. aegypti females, we characterized the transcriptome changes that occur in the female reproductive tract at different time points after mating. We found several RNAs that are apparently transferred by the male, and 280 genes whose mRNA abundance in the female is affected by mating. The nature of the predicted products of many of these genes suggests roles in priming the reproductive tract for egg development, protecting the female against bacterial infections or processing the blood meal. This identification of mating-responsive genes provides information potentially useful for developing tools aimed at preventing disease transmission by manipulating female mosquitoes’ post-mating responses.
| Aedes aegypti is a major vector of arboviruses that impact human health, including those causing dengue and chikungunya [1]. Collectively, these neglected infections cause significant human morbidity and mortality. Among these arboviral threats, dengue is considered the most important, as it affects an estimated 390 million people worldwide per year, with 500,000 episodes of severe dengue and >20,000 dengue related deaths [2, 3]. In addition, chikungunya is sweeping across the globe, with nearly 1 million cases in the Americas in 2015 alone [4]. With no effective antiviral therapy, no commercially licensed vaccine, and no treatment for dengue or chikungunya, control efforts for these diseases must focus on developing tools for vector control that will ultimately reduce the burden of human infections [5]. Tools that target key processes in Ae. aegypti reproduction hold significant promise.
In recent years, progress has been made towards understanding the Ae. aegypti mating system. This work builds upon early observations that, after mating, Ae. aegypti females undergo profound physiological and behavioral changes, including increased egg development and oviposition rates [6, 7], changes in host-seeking and feeding behavior [8], and increased mating refractoriness [9, 10]. These changes largely result from the receipt of seminal fluid proteins (Sfps) that are transferred from males to females along with sperm during copulation [9, 11, 12].
Given the dramatic changes in a female’s physiology and behavior after mating, it is likely that mating triggers large changes in gene expression, yet nothing is known about mating-induced changes in gene expression in Ae. aegypti females. However, mating-induced transcriptome changes in the female have been studied at the whole-body level in other insects, including Drosophila melanogaster[13–17], Apis mellifera[18], Ceratitis capitata[19], and Anopheles gambiae[20], as well as in the D. melanogaster and An. gambiae reproductive tract (RT) [21–23].
Gene expression profiles of D. melanogaster whole females revealed that, soon after copulation (1–3 hours post-mating; hpm hereafter), a large number of genes showed small-magnitude changes in transcript levels. In contrast, by 6hpm, larger magnitude changes were observed for a smaller number of genes [15]. In addition, D. melanogaster females that receive Sfps, but not sperm, show differential expression of 160 genes at 1–3 hpm. Some immunity-related genes, especially those encoding antimicrobial peptides, were strongly up-regulated in response to Sfp receipt. Other categories of genes whose transcript levels were modulated by the transfer of Sfps include ones that encode enzymes, receptors, and transporters, some of which are involved in metabolism [14].
Microarray analysis of the transcriptome of the D. melanogaster female lower RT (reproductive organs below the oviducts) revealed differential expression of 539 genes in virgin versus mated females at 0, 3, 6 or 24hpm [21]. Interestingly, there was a pronounced peak in the number of differentially expressed RNAs at 6hpm, a time when most Sfps are at low levels or are no longer detectable [24, 25]. Many of these mating-regulated genes encode proteins with predicted functions in catalytic activity and in nucleic acid binding. Microarray analyses of RNAs from the oviduct of mated D. melanogaster females identified 432 transcripts expressed differentially in mated relative to unmated females. Among these RNAs, those from immune response genes showed a substantial increase in expression [22]. In An. gambiae, differential expression was detected for RNAs encoding various proteases in whole mated females at 2, 6 and 24hpm relative to virgin females. Some of those RNAs were also up-regulated in females that had mated to spermless males [26], indicating that their abundance was regulated by mating or seminal molecules rather than by sperm (analogous to results previously obtained in Drosophila[14]). Microarray analysis of the An. gambiae lower RT at 3, 12, and 24hpm showed that there is significant upregulation of genes involved in metabolic processes, particularly at 24hpm [23], in agreement with studies of expression profiles of the spermathecae that show increased expression of metabolic genes after 24 hours [27].
A recent paper reported transcriptome profiles from the fat body of Ae. aegypti after blood-feeding [28]. Another recent paper described gene expression changes in reproductive tissues of males and females before and after blood feeding [29]. However, post-mating transcriptome changes have not been explored for the female RT after mating. Identification of specific genes and gene classes whose expression is regulated post-mating is vital to understanding the reproductive biology of this disease vector.
We are especially interested in exploring the link between mating and downstream reproductive processes in the Ae. aegypti natural mating system. For this mosquito, the host serves as an encounter site for mating [30]. Males tend to patrol the space around the human host and intercept females as they fly in to take a blood meal [31]. For this reason, the first mating (and often only mating for this primarily monandrous species [32]), blood ingestion, and the beginning of the first gonotrophic cycle are likely to occur sequentially. Given that mating has already been shown to accelerate oocyte development, we predicted that the male might thereby “prime” the entire RT for these subsequent reproductive stages.
Here, we report mating-induced changes in mRNA levels in the Ae. aegypti female RT. Since we were specifically interested in the male’s potential to prime the female reproductive system, we examined changes that occurred in the female RT without the ovaries as this is the primary site of interaction between the male ejaculate and the female. Using short read RNA sequencing (RNAseq), we compared transcript abundances in the RTs of virgin females to females at three time points after copulation (0, 6 and 24hpm). We chose the 6 and 24hpm time points to capture intervals before females typically produce their first egg batch (∼72 or more hours depending on ambient temperature [33]) and to compare with studies conducted in Anopheles and Drosophila. To consider the potential for transfer of male-derived Sfp transcripts to females, as was observed in D. mojavensis[34], we also included a 0hpm time point. We identified a total of 364 transcripts that show significantly different abundance levels between the RTs of virgin and mated females. A subset of these transcripts is likely male-derived, as their abundance increases dramatically immediately after mating. Furthermore, we find that several gene ontology (GO) categories are enriched among transcripts that are differentially regulated in response to mating, especially among up-regulated transcripts at 6hpm and 24hpm. We discuss these findings in the context of known reproductive processes in other insects and as potential targets for vector control.
We used Aedes aegypti originally collected in Bangkok, Thailand (15.7193°N, 101.752°E) in 2011 and supplemented with field material from the same site in 2012 and 2014. Mosquitoes were reared as described in [35]. Individual pupae were transferred to vials to ensure virginity and sorted by sex upon eclosion. Two hundred individuals were transferred into 8L plastic cages, separated by sex, and held in single-sex groups until experiments commenced.
Matings were conducted as described previously [10]. Briefly, 5 to 6 day-old virgin Ae. aegypti males and females were used for each mating. Each virgin female was released into an 8L observation cage containing approximately 8 virgin males. Male and female couples were observed carefully, and copulating pairs were removed using a mouth aspirator after a minimum mating duration of 8 seconds to ensure successful copulation [36, 37]. In addition, a subset of females were examined for sperm in the spermathecae or in the bursa during dissections. Mosquitoes from the same cohort were used for the virgin samples and the three post-mating time points. Females in the 0hpm treatment were placed on ice immediately after mating. Females from the later post-mating time points were individually transferred into 50mL test tubes after mating and held in an environmental chamber at 28 ± 1°C with 71.9 ± 9.5% relative humidity and a photoperiod of 10h light:10h dark (with a 2h simulated dusk and dawn period). At the 6 or 24hpm, females were placed on ice and immediately dissected.
To test for transfer of male-derived transcripts to females during copulation, virgin females were mated with transgenic males that express eGFP in the accessory glands, driven by the promoter of the Sfp, AAEL010824 [35]. Mated females were placed on ice and immediately dissected after copulation.
To dissect reproductive tracts (RTs), females were anesthetized on ice and placed on a chilled glass slide in a droplet of phosphate-buffered saline (Sigma, St Louis, MO). Dissections were carried out by first dissecting the 0hpm samples, followed by the virgin samples and the 6hpm samples. The 24hpm samples were dissected around the same time the following day. Tissues for each replicate were pooled from several days of dissections. The RT included the bursa, the three spermathecae, the common oviduct and the two lateral oviducts. Care was taken to remove any extraneous tissues, including the adherent fat body and ovaries. Immediately after dissection, RTs were placed in 150μL of Trizol (Invitrogen, Carlsbad, California) and stored on ice. For RNAseq libraries, between 40 and 60 lower RTs were extracted to obtain ∼1μg of total RNA for each replicate of the virgin and post-mating treatments (three replicates for each of the four samples). To test for eGFP mRNA transfer from males, RNA was extracted from RTs of 4 or 5 females mated with AAEL010824-eGFP males.
Immediately after dissection, tissues were homogenized with a pestle, and an additional 350μL Trizol was added to each sample. Samples were incubated at room temperature for 5 min and stored at -80°C. RNA was then prepared by chloroform/isopropanol extraction and ethanol precipitation following the manufacturer’s protocol (Life technologies, Grand Island, NY). The concentration of RNA in each sample was measured using a Qubit spectrophotometer (Invitrogen, Grand Island, NY), and the quality of RNA was measured on a Fragment Analyzer (Advanced Analytical Technologies, Inc., Ames, IA).
From total RNA, mRNA was extracted using the poly-A mRNA Magnetic Isolation Module (New England Biolabs, Inc., Ipswich, MA). cDNA libraries for each replicate were prepared using the NEBNext Ultra RNA Library Prep Kit for Illumina (New England Biolabs, Inc., Ipswich, MA). NEBNext Multiplex Oligos for Illumina (New England Biolabs, Inc., Ipswich, MA) were ligated to each library prior to sequencing.
Samples were sequenced on the Illumina platform (HiSeq 2500, Cornell University Biotechnology Resource Center) in two rounds. In the first, all samples were sequenced in one lane and the reads were analyzed for consistency among replicates (see below and S1 Methods). It was found that two of the virgin repicates and one 24hpm replicate contained a large number of reads derived from rRNAs; thus, these samples were considered to be contaminated with ribosomal RNA, and were not included in the final analyses. Rather, in the second round of analysis, one of the virgin replicates and two replicates of each of the mated samples were re-sequenced at higher depth and used in subsequent analyses in place of the rRNA-contaminated libraries.
Single-end, 100bp reads were processed by trimming the first 10 bases and end bases below a quality Phred score of 20 (FASTX-TOOLKIT v.0.0.13). In addition, reads having an average quality score below 20 were discarded.
To generate the transcriptome used for differential expression analysis, each sample was mapped individually to the Ae. aegypti genome (release AaegL2, VectorBase) (Tophat v.2.0.9). The Ae. aegypti annotation file (release AaegL2.2) was used as a guide for known transcripts, allowing us to circumvent the need to heavily annotate the transcriptome de novo. Mapped reads from each sample were assembled into transcript fragments (Cufflinks v.2.2.1), and assembly annotations from each sample were merged to produce the gene annotation file used to define transcripts for differential expression analysis. Spliced transcript sequences were then reconstructed from the genome sequence using the merged gene annotation file (Cufflinks “gffread” utility, v.2.2.1). We also analyzed the data for changes in splice-isoforms. The final transcriptome consisted of 23,770 genes, which comprised 42,858 isoforms.
To perform differential expression testing among samples, each replicate was individually mapped to the transcriptome (Bowtie2 v.2.2.2) and the raw number of reads mapping to each transcript was estimated (RSEM v.1.2.8). The read counts were processed and normalized using the Trimmed Mean of M-values (TMM) method to obtain reads per kilobase of transcript per million mapped reads (RPKM) values. A quality filter was established by estimating a minimum read count (180 reads/transcript) in any given condition, above which a 2-fold change in expression between virgin and post-mating samples would be considered reliable; 10,031 transcripts satisfy this minimum read count threshold (S1 Methods). Additionally, to assess consistency among replicates, the fold change between replicates of each sample was examined; a total of 40 transcripts showed higher than 2-fold difference between replicates of the same sample and were excluded from further analysis. For each transcript that passed this filter, the fold change, p-value and false discovery rate (FDR) were calculated separately for each of the virgin-0hpm, virgin-6hpm, and virgin-24hpm comparisons (edgeR v. 3.2.4). The virgin-0hpm comparison revealed a set of transcripts that are likely male-derived, and were thus analyzed separately from later time point comparisons (these potentially male-derived transcripts were removed from subsequent analyses). Differentially expressed transcripts with similar expression profiles from the virgin-6hpm and virgin-24hpm comparisons were grouped by K-means clustering (R “Cluster” package).
An annotation database for the transcriptome produced here was created according to the Trinotate guidelines for transcriptome annotation (v. r20140708). First, transcriptome sequences were examined for the presence of coding sequences by in silico prediction of open reading frames (ORFs) with a minimum peptide length of 20 amino acids (Transdecoder r20140116). Second, homology to known transcripts was assessed by generating BLASTp and BLASTx reports against the SwissProt database (www.uniprot.org). Finally, homologous protein domains (hmmscan v.2.3.2), predicted signal peptides (signalP v.4.1), and predicted transmembrane domains (tmHMM v.2.0) were identified for each protein sequence. This comprehensive annotation database was used to query possible gene functions and to perform downstream Gene Ontology (GO) enrichment analysis. GO enrichment analyses were performed for the set of up- and down-regulated transcripts and for clusters with similar expression profiles (GOSeq:R package v.1.18.0; DAVID v.6.7).
To validate the expression profiles of differentially expressed transcripts identified through RNAseq, transcript abundance was measured at different time points for seven genes using qRT-PCR. Total RNA from RTs was isolated as described above (samples were independent from those used for RNAseq analysis). Prior to cDNA synthesis, RNA was treated with RNase-free DNAse (Clontech, Madison, Wisconsin). Reverse transcription was performed using 1μg of total RNA and oligo dT following the manufacturer’s instructions (Clontech, Madison, Wisconsin). Relative transcript levels of each gene were measured using quantitative PCR conducted with a CFV96 Real-Time System (Bio-rad, Hercules, California). For primer sequences used, see S1 Table. Amplifications were carried out in a 15μl reaction containing 7.5μl of iQ SYBR Green Supermix (Bio-rad, USA), 1μl cDNA and 300nM of each primer. Cycle differences between genes and RpS17 and Actin (ΔΔCT) were compared to generate the relative expression of the transcripts at different time points before and after mating. Two different housekeeping genes were included to ensure that mating did not influence transcript abundance of one of those genes. The fold-change in the RNAseq experiment and the qRT-PCR were compared using a Pearson correlation coefficient. In addition, qRT-PCR was used to test eGFP mRNA transfer from males to females during copulation, following the methods describe above. ΔΔCT for eGFP transfer was generated using RpS17 for normalization.
Aedes aegypti undergoes drastic physiological and behavioral changes after mating that result from the receipt of seminal fluid. Because the reproductive tract plays a crucial role in seminal receipt, sperm storage and ejaculate processing, and, ultimately, in egg development and oviposition, we were interested in understanding gene expression changes specifically in RT tissues. To examine post-mating transcriptome changes within the female RT, we compared transcript abundance levels in virgin females with those in mated females at 6 and 24hpm. We chose the 6hpm time point based on results from Drosophila, in which the peak of post-mating expression (the number of differentially expressed RNAs and/or the magnitude of fold-change) was highest [15, 21]. We chose to examine 24hpm to allow comparison to An. gambiae, whose peak differential expression was observed at this time [20], allowing us to identify transcriptome responses persistent throughout Culicidae. Finally, because male-derived mRNAs have been reported to be transferred during mating in D. mojavensis[34], we tested whether this phenomenon occurs in Ae. aegypti by comparing the RT transcript abundances of virgin females to those extracted from females immediately after mating (0hpm).
We measured abundance levels of 23,770 assembled transcripts (6,481 previously unannotated), 10,031 of which (1,143 previously unannotated) passed our quality control filter (described above and in S1 Methods). Significance tests for each gene that passed the filter were conducted for each of the virgin-0hpm, virgin-6hpm, and virgin-24hpm comparisons. We only considered transcripts that had a 2-fold or greater change in abundance between virgin and post-mating samples. In addition, a transcript was considered significantly differentially expressed if the significance test p-value was below 0.05 after correcting for multiple tests (Benjamini-Hochberg, [38]).
When considering differential expression across all three post-mating time points, we found a total of 364 transcripts to be significantly differentially expressed between virgin and post-mating RT samples (Fig 1A). The median fold-change among transcripts that are significantly up- and down-regulated across the three time-points was approximately 12-fold and 6-fold, respectively. The highest increases in abundance were observed at 0hpm (median fold-change ≈275 fold) whereas the increases in abundance at 6hpm and 24hpm were markedly lower (median fold-change ≈13-fold and 9-fold, respectively). Furthermore, several of the transcripts that increase at 0hpm maintain high abundance at later time-points. On the other hand, the median decrease in abundance across the three post-mating time-points was of lower magnitude than the increase in abundance, ranging from 4.5-fold to ∼8-fold. Overall, the majority of differentially expressed transcripts were found in the 24hpm time-point (Fig 1).
To glean information about alternative splicing, we also examined our data for changes in isoform abundance. However, isoform discrimination relies on the relatively small number of reads that span specific exon-exon boundaries, and therefore the subset of isoforms for which we had sufficient data to draw conclusions was lower than for total gene expression. Nevertheless, several interesting examples were identified where dramatic changes in relative isoform abundances are detected across the mating time course (S4 Fig). Broadly, the isoforms’ behaviors mirror those seen in the overall transcriptome data, insomuch as the abundance of some isoforms increases at specific time points, while others decrease. A further analysis of the splicing changes that accompany changes in post-mating gene expression will require deeper sequencing and/or improved techniques to fully characterize changes in alternative splicing.
To confirm the validity of differential expression results from the RNAseq analysis, we extracted new samples of RNA and used qRT-PCR to examine expression levels of seven genes selected based on their expression pattern in the RNAseq dataset (Fig 2). The seven genes include a predicted trypsin (AAEL006414), an oxidoreductase (AAEL009685), two cecropins (AAEL000621 and AAEL000611), a zinc metalloprotease (AAEL003012), a defensin (AAEL003841) and a gene with unknown function (AAEL000545). Melt-curve analysis revealed a single product for each tested gene, confirming the efficacy of primer amplification. In general, the fold change estimates from both analyses were in close agreement (Fig 2). Our qRT-PCR validated the reliability of our RNAseq data. For all seven transcripts, the direction of fold-change was the same in the qRT-PCR and the RNAseq analysis. Moreover, the magnitude of fold-change for any given mRNA in both analyses is within 2 standard deviations.
In our comparison of RT transcript levels between virgin females and females dissected immediately after mating (0hpm), we observed 76 transcripts whose abundance was >2-fold different between treatments (Fig 3). Sixteen of these RNAs decreased in abundance (between ∼4-fold and ∼61-fold) in the 0hpm sample relative to the virgin sample, but nearly all of them gradually increase in abundance in later time points; these include RNAs encoding several antimicrobial peptides. On the other hand, sixty transcripts showed increased abundance (between ∼5-fold and ∼1.4x104-fold) at 0hpm, with a median fold change of 275. These include 53 transcripts that are highly expressed in male reproductive tissues [29] (S5 Fig, Supplementary Results). Twenty of these transcripts encode known Ae. aegypti Sfps [39], including the previously described Sfp AAEL010824 [35]. The transfer of Sfps seen here is similar to that reported in D. mojavensis[34]. However, likely due to strong selection on Sfps in general [40], we found no clear homology between Sfp transcripts transferred in D. mojavensis and Ae. aegypti. Next, using RT-qPCR, we detected eGFP mRNA in the RT of wildtype females mated to AAEL010824-GFP transgenic males (Fig 4). These data, and the fact that new transcripts are highly unlikely to have been induced in the females in such quantity during the short time-frame of an Ae. aegypti mating, suggest that the RNAs detected as “up-regulated” at 0hpm were likely male-derived.
The expression levels of most of these allegedly male-derived transcripts follow a similar trend of exponential decay, implying that these transcripts are transferred to the female in one pulse and are then degraded at a uniform rate. One dramatic case is an unannotated transcript (XLOC019584) that has the highest abundance level at 0hpm (1.76 × 104 RPKM) and appears to be specifically expressed in Thai strain males, but not expressed in the Liverpool strain [29]. Mass spectrometry of male accessory gland extracts from the Thai strain (Villarreal, Avila, Wolfner, and Harrington, in prep) identified three peptides produced from isoforms of this mRNA (S9 Fig), consistent with XLOC019584 encoding a seminal protein. It is possible that the male-derived transcripts serve no function in the female, and are simply membrane or vesicle-associated byproducts of apocrine secretion from the male accessory glands.
After removal of allegedly male-derived transcripts (60 transcripts), we found 280 transcripts that were significantly up- or down-regulated between virgin and the later post-mating (6hpm and 24hpm) samples (S6 Fig). Because of the correction for multiple tests, the number of significantly differentially expressed transcripts in this second analysis is less than the number obtained if the number of male-derived transcripts were simply subtracted from the original number of differentially expressed transcripts (364). To examine the expression profiles and functional classes of transcripts that are differentially expressed at 6hpm and 24hpm, we clustered transcripts into groups by K-means clustering. First, we estimated K as the square root of half the number of transcripts [41], which yielded 12 clusters. Second, clusters with similar mean expression profile were merged (S7 Fig), such that peak abundance at any given time point is similar. We ultimately obtained a set of four clusters that summarize the observed variation in expression profiles among the differentially expressed transcripts (Fig 5A, S3 Table).
It has recently been shown that some genes in Ae. aegypti are under circadian regulatory control [42]. Because these patterns can influence influence interpretation of abundance changes in our study, we examined the circadian expression profile of the 280 transcripts identified here and found that only 10 of these show a circadian pattern of expression (S10 Fig). Therefore, the expression changes we report here are largely due to mating.
To better understand the biological significance of the changes we detected, we carried out GO enrichment analysis of the 150 transcripts that increased in abundance between virgin and mated-female RTs at 6hpm and 24hpm as a whole (Table 1) and as partitioned into the two clusters depicting different expression profiles (C1 and C2, Fig 5A, S2 Table). We also assigned GO terms for each transcript using Ensmbl and VectorBase (Fig 5B, S3 Table). Approximately 37% of these transcripts (57) encode proteins of unknown function (i.e. not classified as belonging to any specific biological process) (Fig 5B, S3 Table). The remaining transcripts largely encode proteins involved in cellular localization (17), proteolysis (26), and immunity (9). The latter were among the most differentially expressed, particularly at 24hpm; their mean increase in expression was ∼10-fold. Further, the largest categories of up-regulated genes with predicted function were those with binding (20) and peptidase activity (39) (S3 Table).
Two clusters derived from the K-means approach comprise the up-regulated set (Fig 5A). They are characterized by the time-point in which the fold-change was highest: Cluster 1 contains transcripts that are most abundant at 6hpm, while cluster 2 contains transcripts whose abundance was highest at 24hpm.
130 transcripts were down-regulated relative to virgin and mated females at 6hpm and 24hpm. Of these, 15 were down-regulated at 6hpm and 115 were down-regulated at 24hpm. The latter set contained 32 previously unannotated transcripts. GO analysis of the down-regulated transcripts shows that 44% (55 transcripts) are not classified as belonging to a specific biological process. The remaining transcripts encode proteins involved in binding (16 transcripts) and catalytic activity (40 transcripts) (Fig 5B, S3 Table). Two clusters derived from the K-means approach represent transcripts that have lowest abundance at 6 and 24hpm (Fig 5A).
Aedes aegypti females typically mate near the host, after which they take a blood meal and ultimately develop eggs. Therefore, we postulate that mating primes a female for these downstream processes, triggering expression changes that physiologically prepare her for blood meal digestion and oogenesis. In contrast with model organisms, such as D. melanogaster, relatively few genes in Aedes aegypti have had their functions elucidated experimentally. Therefore, we have speculated here about potential roles of genes in the post-mating response based on their putative functions, with the caveat that these suppositions must be tested.
Mating is directly linked to a female’s vectorial capacity via altered blood-feeding behavior and increased reproductive potential. Therefore, a thorough knowledge of mating-induced transcriptional changes may aid in developing tools that aim to prevent disease transmission by manipulating the post-mating response. Our dataset provides a framework by which this crucial life history milestone in Ae. aegypti can be examined. This study deepens our understanding of broad-scale gene expression changes after mating, and will serve as a launching point for future gene-specific investigations.
|
10.1371/journal.pgen.0030014 | Reductive Genome Evolution from the Mother of Rickettsia | The Rickettsia genus is a group of obligate intracellular α-proteobacteria representing a paradigm of reductive evolution. Here, we investigate the evolutionary processes that shaped the genomes of the genus. The reconstruction of ancestral genomes indicates that their last common ancestor contained more genes, but already possessed most traits associated with cellular parasitism. The differences in gene repertoires across modern Rickettsia are mainly the result of differential gene losses from the ancestor. We demonstrate using computer simulation that the propensity of loss was variable across genes during this process. We also analyzed the ratio of nonsynonymous to synonymous changes (Ka/Ks) calculated as an average over large sets of genes to assay the strength of selection acting on the genomes of Rickettsia, Anaplasmataceae, and free-living γ-proteobacteria. As a general trend, Ka/Ks were found to decrease with increasing divergence between genomes. The high Ka/Ks for closely related genomes are probably due to a lag in the removal of slightly deleterious nonsynonymous mutations by natural selection. Interestingly, we also observed a decrease of the rate of gene loss with increasing divergence, suggesting a similar lag in the removal of slightly deleterious pseudogene alleles. For larger divergence (Ks > 0.2), Ka/Ks converge toward similar values indicating that the levels of selection are roughly equivalent between intracellular α-proteobacteria and their free-living relatives. This contrasts with the view that obligate endocellular microorganisms tend to evolve faster as a consequence of reduced effectiveness of selection, and suggests a major role of enhanced background mutation rates on the fast protein divergence in the obligate intracellular α-proteobacteria.
| Genome downsizing and fast sequence divergence are frequently observed in bacteria living exclusively within the cells of higher eukaryotes. However, the driving forces and contributions of these processes to the genome diversity of the microorganisms remain poorly understood. The genus Rickettsia, a group of small obligate intracellular pathogens of humans, provides a fascinating model to study the genome downsizing process. In this article, we used seven Rickettsia genomes to reconstruct the genome of their ancestor and inferred the origin and fate of the genes found in today's species. We identify the process of gene loss as the main cause of genome diversification within the genus and show that the rate of gene loss, sequence divergence, and genome rearrangements are highly variable across the various Rickettsia lineages. This heterogeneity likely reflects the intricate effects of specialization to distinct arthropod hosts and critical alterations of the gene repertoire, such as the losses of DNA repair genes and the amplification of mobile genes. In contrast, we did not find evidence for the role of reduced population sizes on the long-term acceleration of sequence evolution. Overall, the data presented in this article shed new light on the fundamental evolutionary processes that drive the evolution of obligate intracellular bacteria.
| Intracellular bacteria that are strictly associated with multicellular eukaryotes possess small genomes, typically in the range of 1 Mb or less. This feature is a consequence of the reduction of originally larger genomes invariably accompanying the adaptation to parasitic/symbiotic lifestyles. The transition from a free-living existence to a close relationship with eukaryotic cells is a frequent theme in bacterial evolution and has been documented in mycoplasmas, phytoplasmas, chlamydias, and the α- and γ-proteobacteria [1]. The Rickettsia genus is a group of obligate intracellular, small, rod-shaped, α-proteobacteria that possess highly reduced genomes compared to those of their free-living relatives. Known Rickettsia are parasites of arthropods such as ticks and insects (lice and fleas) [2,3], in which they are presumably stably maintained in the population and can be vertically transmitted. Through bites or feces of the vectors, they can infect mammals that can become sources for the next lines of infected vectors. Many members of this genus cause mild to fatal diseases in humans.
The Rickettsia genus provides an excellent model to investigate the process of reductive evolution. Their genomes present substantial inter-species variations in size (1.1–1.5 Mb) and gene content (about 900–1,500 genes) as a consequence of the recent and ongoing genome degradation process [4]. However, they exhibit few recent gene transfers [5] and genome rearrangements [6], which help a fine reconstruction of genome evolution history. Complete genome sequences are publicly available for five Rickettsia species covering the three major genus sub-groups: the typhus group (TG), including Rickettsia prowazekii (the agent of epidemic typhus transmitted by louse [4]) and Rickettsia typhi (the agent endemic typhus transmitted by flea [7]); the spotted fever group (SFG), including Rickettsia conorii (the agent of Mediterranean spotted fever transmitted by tick [5]) and Rickettsia felis (the agent of flea-borne spotted fever [6]); and the last group currently represented by a sole species, Rickettsia bellii, associated with ticks [8]. The availability of these five Rickettsia genome sequences, as well as two new SFG Rickettsia genome sequences determined in our laboratory (i.e., Rickettsia africae, the agent of African tick bite fever and Rickettsia massiliae, the agent of a tick-borne spotted fever in Europe [9]) prompted us to carry out a comparative genomics analysis to investigate how genome reduction and other evolutionary processes have contributed to the diversity of the genus. In this study, we identified the genes conserved in seven Rickettsia, inferred the gene content of their ancestors, and investigated the evolutionary dynamics of genome changes that occurred during the evolution of the genus.
We first built initial sets of putative genes for the seven Rickettsia genomes using a common protocol to avoid potential biases across the originally published data that were generated by different gene identification methods. This study aims to investigate the evolution of Rickettsia genomes through the reconstruction of the genome of their last common ancestor. Thus, rather than generating maximal sets of potential genes, we designed a stringent gene identification method that maximally recognizes orthologous genes that can be used to reconstruct the ancestral genome (see Methods). This method has the advantage of minimizing the inclusion of erroneously predicted genes at the cost of potentially omitting small genes that are not well supported by comparative genomics. As they probably represent a very small fraction of the genomes, we believe this potential bias does not interfere with the conclusions drawn from this study.
The predicted genes were organized into orthologous gene groups (hereafter, referred to as Rickettsia genes [RIGs]) based on the reciprocal best BLAST hit criterion and the extensive colinearity between the genomes. We obtained 1,867 RIGs: 1,827 for proteins and 40 for RNAs (34 tRNAs, three rRNAs, one M1 RNA, one tmRNA, and one 4.5S RNA). Genes interrupted by frameshifts or internal stop codons were flagged as pseudogenes and considered as nonfunctional. Additional pseudogenes (i.e., highly degraded gene remnants) were identified within intergenic regions using TBLASTN (see Methods). In fine, each member of the RIGs was flagged as (1) full-length gene, (2) pseudogene, or (3) absent.
The number of predicted genes in the Rickettsia genomes varies from 846 (R. typhi) to 1,324 (R. felis) (Table 1). Coding fraction of genome varies from 69% (R. massiliae) to 79% (R. bellii). Numbers of pseudogenes are from 100 (R. bellii) to 274 (R. massiliae). The genomes of the TG Rickettsia are smaller and poorer in G + C than the genomes of R. bellii and the SFG Rickettsia (Table 1). Rickettsia palindromic elements (RPEs) are interspersed repetitive sequences of 100–150 bp, abundant in both the intergenic and protein-coding regions of Rickettsia genomes [10] (Table S1). We found that several RPE families exhibit lineage-specific proliferation (e.g., RPE-4 in R. bellii and RPE-7 in R. felis). In all investigated Rickettsia, the gene encoding the phenylalanyl-tRNA synthetase β chain (PheT) has a 63- to 84-bp insertion missing in other α-proteobacteria's orthologs (Figure S1). This Rickettsia-specific insert exhibits a significant sequence similarity with the RPE-7 repeats. This suggests that the proliferation of this RPE family started before the divergence of different Rickettsia lineages.
The seven Rickettsia genomes share 704 full-length orthologous protein-coding genes and 39 RNA-coding genes. These core genes correspond to 52%–86% of the predicted gene content of each genome and reflect the functions that have been maintained in all the Rickettsia lineages analyzed here. Of the 704 core protein genes classified into COG (cluster of orthologous group) categories, 546 are associated with a known biochemical function. The functions of the remaining 158 genes, including 40 uniquely found in Rickettsia, are poorly characterized or unknown, suggesting that basic biological features shared by Rickettsia species remain to be elucidated. Some of the core genes are associated with the pathogenicity or parasitic lifestyle of Rickettsia. They include genes encoding proteins related to the virulence of Rickettsia, such as the hemolysins (TlyA, TlyC) [11,12], the phospholipase D (Pld) [11,13], the OmpB outer membrane protein [14], the R. conorii putative adhesin (RC1281) orthologs [15], the parvulin-like peptidyl-prolyl isomerase (SurA) [16,17], and the components of the type IV secretion system (VirB/D) [18,19]. The core gene set also includes the genes for the ATP/ADP translocases, the hallmark enzymes of intracellular parasitism in Rickettsia and Chlamydia, as well as genes related to environmental stresses such as the guanosine pentaphosphate phosphohydrolase (gppA), the universal stress protein (uspA), superoxide dismutase (sodB), and cold-shock protein (cspA) genes. The core proteins were aligned and concatenated to determine robust phylogenetic relationships between the seven species (Figure 1). This phylogenetic tree served as a reference in the following sections.
We reconstructed gene repertoires of ancestral Rickettsia up to their last common ancestor, R0, using the maximum parsimony criterion (Figures 1 and 2). This reconstruction strongly supports the presence in R0 of 1,252 (1,213 protein-coding genes and 39 RNA-coding genes) out of the 1,867 RIGs. The seven Rickettsia genomes retained 66% (R. typhi) to 89% (R. bellii) of the 1,252 R0 RIGs (Figure 1 and Table 1), which constitute 81% (R. felis) to 98% (R. prowazekii and R. typhi) of the gene complement in the modern species. These figures indicate that the process of gene loss substantially contributed to the differences in gene contents between the modern Rickettsia, and that the TG genomes were mostly shaped by the reductive evolution from the R0 genome. The 615 remaining RIGs were specific to the R. bellii (211 cases) or the TG/SFG clades (404 cases). They may have been either (1) inherited from the R0 ancestor but lost in one clade, or (2) recently acquired by lateral gene transfer or gene duplication. Genes encoding transposases, which comprise 160 of the 615 RIGs, often occur in multiple, near-identical copies in the Rickettsia genomes, and, therefore, many have probably originated from recent duplications. The 65 RIGs of the R. felis plasmid have probably been acquired through a single transfer event. Among the remaining 390 chromosomal RIGs, we found only one case (metK, see below) of recent horizontal gene transfer that might have occurred after the divergence of R0, suggesting that many of those genes were already present in the last common Rickettsia ancestor. This quasi-immunity for detectable horizontal gene transfer within the Rickettsia genus contrasts with the many examples of lateral transfer detected prior to the R0 ancestor [8]. Our reconstruction indicates that the R0 genome contained at least 1,252, and most likely, about 1,650 genes.
Our reconstruction suggests that R0 was already parasitic, lacking most of the genes for the biosynthesis of amino acids and nucleotides and exhibiting few genes for carbohydrate degradation and transcriptional regulation. In contrast, it possessed five paralogs for ATP/ADP translocases, several genes for amino acid transporters, and the gene for RickA [20] responsible for the actin-based intracellular motility. The R0 ancestor exhibited only a few genes relevant to cofactor metabolism, among which we can cite the five full-length genes for the metabolism of biotin: a biotin synthase gene (bioB), a biotin (acetyl-CoA carboxylase) ligase gene (birA), a biotin–protein ligase gene (bpl1), and two genes (bioY1 and bioY2) for the BioY family protein involved in the synthesis of dethiobiotin. None of the modern Rickettsia possesses all of the five genes in a full-length state. The R0 genome also exhibits four genes for the folate metabolism: a bifunctional folate synthesis protein gene (folKP), a dihydrofolate reductase gene (folA), a tetrahydrofolate dehydrogenase/cyclohydrolase gene (folD), and a gene for the 5-formyltetrahydrofolate cyclo-ligase. Of the existing Rickettsia, only R. bellii possesses all four genes. The R0 genome possessed a sam gene for uptaking S-adenosylmethionine, which is conserved in all the existing Rickettsia [21]. In contrast, the metK gene for the synthesis of S-adenosylmethionine is inferred as absent in the R0 genome, while it was inferred as full-length in R1 (the common ancestor of TG and SFG). Our phylogenetic analysis suggests that a Rickettsia ancestor (before R1) acquired the metK gene through lateral gene transfer, probably from a γ-proteobacteria (Figure S2). The metK gene was then degraded in different lineages of TG and SFG Rickettsia [22,23]. It is interesting to notice that this gene has recently been implicated in the virulence of R. prowazeki strains causing typhus [24]. Finally, the newly sequenced genome of R. massiliae exhibits a cluster of putative genes for DNA transfers, which are homologous to the tra gene cluster previously identified in the R. bellii genome [8]. The order of several genes flanking to the tra gene clusters are conserved between R. bellii and R. massiliae. Accordingly, our computational method predicts the presence of the 13 genes for conjugal DNA transfer in the ancestral R0 genome. Given the mobile nature of these genes, we cannot rule out that R. bellii and R. massiliae independently acquired the tra genes after R0.
We examined the dynamics of genome reduction process based on the computational reconstruction of the ancestral gene repertoires. Here, we considered only gene losses among the minimally predicted 1,252 R0 gene set. A gene loss event was defined as a transition from full-length to degraded (pseudogene or absent) state of a RIG in a branch of the Rickettsia tree. Of the 1,252 genes, 491 (39.2%) have been lost in at least one branch. Overall, we inferred 970 different gene loss events over the whole phylogeny (Figure 1 and Table S2); many genes (310, 25%) have been lost in more than one branch.
Consistent with a large variation in genome size between different Rickettsia species, the estimated rate of gene loss is highly variable across the branches of the tree (Figure 1 and Table S2). The R. bellii and R. felis lineages underwent fewer gene losses than the other lineages. Their genomes are the largest among the sequenced Rickettsia. On the other extreme, the highest number of gene losses occurred in the branch leading to the TG possessing the smallest genomes of the genus (R. prowazekii and R. typhi). The R. massiliae, R. conorii, and R. africae lineages went through a notable acceleration of the rate of gene loss after their separation with the R. felis lineage.
We asked if the conservation of the 743 core genes is solely due to the small number of genome samples. To test this, we used a simulation based on the model M1 (see Methods), in which 970 losses are randomly distributed among 1,252 genes, retaining the number of gene loss per branch as estimated above. The model predicts an average of 524.7 universally conserved genes (standard deviation = 8.5; 1,000 Monte Carlo samplings), which is significantly lower than the number of core genes (743) in the real data (p ∼0). This suggests that there has been a strong evolutionary constraint to retain at least a subset of the core genes. In other words, the gene losses were confined to a subset of the R0 gene repertoire (i.e., dispensable gene set) containing genes not essential in the context of intracellular parasitism.
Next, we examined the notion that dispensable genes have widely different propensities of loss [25], which may vary in function of the genome context and specialization to a new niche. We used a simulation based on the model M2 (see Methods), in which gene losses are confined to the 491 dispensable gene set. Under equal propensities of loss (i.e., M2 with a2 = 1), the simulated distribution fitted the real data poorly (p < 0.05; Figure 3A). If we allow two classes of genes with different propensities of loss, the distributions fitted the real data better (a2 = 0.11; p > 0.05; Figure 3B). This suggests that the propensity to be lost varies across Rickettsia dispensable genes; some genes are more prone to inactivation than others or became dispensable in the course of evolution, for instance, as a result of the adaptation to a new ecological niche.
Finally, we examined the genome reduction process in terms of lost functions and chromosomal colocalization. Genes with unknown function or general function prediction (which includes many paralogous gene families like ankyrin protein or toxin-antitoxin protein genes), as well as genes involved in defense and signal transduction, have been lost significantly more often than expected by chance (Table 2). Inversely, losses were significantly underrepresented among genes involved in protein metabolism, nucleotide and ion transport and metabolism, cellular trafficking, energy production, and cell envelope biogenesis. Colocalization of gene loss events was also obvious from Figure 2. Using the predicted initial gene order in R1 as reference (see the genome rearrangement section), we confirmed that lost genes after the R1 ancestor were more frequently clustered in the R1 genome than expected by chance (Table 3). There are two possible reasons: first, gene loss may occur through deletion of large DNA segments encompassing several genes, as for example, during the Buchnera genome evolution [26]. Second, genes involved in a common pathway (e.g., operon) may undergo simultaneous degradation by small-scale mutations. This can arise if a metabolic pathway becomes superfluous as a result of changes in the life cycle, relaxing the functional constraints on the underlying genes colocalized on the chromosome.
Genome reduction correlates with an acceleration of sequence evolution [27]. We found that the TG core proteins globally diverged at a higher rate than their SFG counterparts (Figure 4A) consistent with the smaller sizes of TG Rickettsia genomes. The average rate for amino acid substitution was 2.43 times higher in the TG sub-tree than in the SFG sub-tree. Branch lengths estimated from the concatenated 4-fold degenerated (FFD) nucleotide positions of the core genes exhibit the same trend; i.e., the mean substitution rate at FFD sites is 2.32 times higher in the TG sub-tree than in the SFG sub-tree (Figure 4B). FFD sites are the positions in codon where any nucleotide change results in a synonymous mutation and are thus expected to be largely free from selection and to evolve at a pace similar to the mutation rate [28]. Hence, the faster protein divergence in TG is due in large part to a higher background mutation rate.
Increase of the mutation rate often correlates with the loss of genes involved in DNA repair pathways [29,30], which results in higher replication error rates [31]. Six RIGs belonging to DNA repair processes were lost in one or more Rickettsia (they include phrB, radC, mutM, mutT, and two putative alkylated DNA repair protein genes). Only mutM, mutt, and the alkylated DNA repair protein genes exhibit a pattern of presence/loss consistent with the higher mutation rate in TG (i.e., lost in both R. prowazekii and R. typhi, but retained in the other Rickettsia). The Escherichia coli mutT, mutM, and mutY genes protect the cell against the effects of the oxidative stress product, 8-oxoguanine, which can be incorporated during DNA synthesis and then paired with either A or C [32]. The mutY gene is present in all sequenced α-proteobacteria except in the order Rickettsiales (including the Rickettsia, Wolbachia, Anaplasma, and Ehrlichia genera). In E. coli, the mutTmutMmutY mutant strain produces 8.5 times more G:C to T:A transversions than the single mutY mutant strain [33]. These are consistent with the fast evolution and the AT enrichment for the TG genomes (mutTmutMmutY deficiency genotypes) relative to the SFG genomes (mutY deficiency genotypes) (Table 1). The variation of mutation rates among Rickettsia may also be due to their differences in generation times. Tick-associated SFG Rickettsia exhibit a longer generation time owing to their parasitic association with hard ticks that feed only two or three times during their life and have a very slow life cycle. The bacteria are in a quiescent state when ticks are not feeding and reactivated during the tick feeding [5]. In contrast, insect-associated Rickettsia (R. prowazekii, R. typhi, and R. felis) are agents of worldwide pandemics and exhibit a sporadic propagation at a very high rate by taking advantage of an arthropod-mammal cycle [3]. Thus, these bacteria may have an average generation time much shorter than tick-associated SFG Rickettsia.
In addition to a high mutation rate, the enhanced rate of protein divergence can be caused by a reduced efficiency of purifying selection [34,35]. The ratio of the level of nonsynonymous substitutions (Ka) to the level of synonymous substitutions (Ks), denoted by ω = Ka/Ks, is a classical measure of the magnitude and direction of selective constraints acting on protein sequences, with ω = 1, <1, and >1, indicating neutral evolution, purifying selection, and positive diversifying selection, respectively [36]. To compare the average selective pressures acting on the proteomes of Rickettsia and their relatives, we identified a set of 200 orthologous protein genes conserved in the genomes of three bacterial groups: Rickettsia, Anaplasmataceae, and free-living γ-proteobacteria. The Anaplasmataceae family is a group of obligate intracellular α-proteobacteria closely related to Rickettsia that includes the Wolbachia, Ehrlichia, and Anaplasma genera. For γ-proteobacteria, we used 13 genomes from the Vibrio, Photorhabdus, Salmonella, Shigella, and Escherishia genera (collectively referred to as the coli-group hereafter). Pair-wise ω ratios were estimated from the concatenated nucleotide alignments of the 200 conserved genes and were plotted against the level of divergence between the compared genomes measured by Ks (Figure 5A). For all of the three bacterial groups, the ω values are relatively high (ω = 0.18–0.88) for closely related genome pairs (Ks < 0.1) and decrease with increasing Ks.
It has been suggested that the obligate intracellular Buchnera accumulated higher fractions of nonsynonymous mutations than their free-living relatives [34,37], though this hypothesis has been recently questioned [30]. In contrast, the trajectories of ω for Rickettsia, Anaplasmataceae, and the coli-group converge toward comparable values (ω = 0.048–0.098 for Ks > 0.2). This suggests that, in the long term, Rickettsia and Anaplasmataceae retained similar proportions of nonsynonymous mutations as their free-living relatives. We cannot rule out that the convergence of ω in the three groups is due to saturation of nonsynonymous substitutions, though this is rather unlikely for comparisons with Ks < 2, as the Ka values are below 10%. Selection on synonymous codon usage in highly expressed E. coli genes can be another issue as this phenomenon tends to decrease Ks and therefore to overestimate ω. However, we could not observe any difference in ω across the three bacterial groups when strongly or weakly expressed E. coli homologs were analyzed separately (unpublished data). In summary, the elevated level of protein sequence evolution observed for Rickettsia and Anaplasmataceae [27,30,35] is mainly attributable to an increased background mutation rate rather than a modification of selective pressure.
A number of authors noted that the ω ratio is surprisingly high when very closely related bacterial sequences are compared [38–42]. Our data exhibit a similar trend (Figure 5A). The increase of ω with decreasing evolutionary distance is probably universal. As proposed by Rocha et al. [43], this time dependency of ω is probably due to a lag in the removal of nonfixed slightly deleterious nonsynonymous mutations. Sequencing errors can be another issue as their contribution to the rate estimates becomes larger when the number of true nucleotide differences is small; they bias the apparent ω toward one. However, we believe that this effect is negligible, as an independent evaluation of the sequencing error rate between two fully sequenced R. prowazekii strains indicates that less than one sequencing error occurs every 105 bases on average (unpublished data). Extrapolation of our results suggests that a substantial fraction of the nonsynonymous changes observed between closely related Rickettsia (and other bacteria) are nonfixed mutations (i.e., polymorphisms) in their respective populations and will eventually be eliminated by natural selection.
We explored, in a similar fashion, the relationship between gene loss and level of genome divergence (Figure 5B). The number of gene losses between any two genomes was taken as the sum of the loss events inferred along the branches that separate the two species on the Rickettsia phylogenetic tree (Figure 1 and Table S2). We normalized the number of gene loss by Ks (hereafter referred to as the Ω ratio) in order to account for the fact that a larger number of gene losses are expected between distantly related genomes. Interestingly, the trajectory of Ω against Ks is reminiscent of that observed for ω: the number of gene loss relative to the level of synonymous changes is higher between closely related genomes and decreases with the level of divergence. A simple explanation for the dependence of Ω on the level of divergence involves the effect of polymorphism, as in the case of the elevated ω for closely related bacteria. By extrapolation, a fraction of the gene losses identified between closely related Rickettsia genomes may represent nonfixed slightly deleterious mutations that will eventually be eliminated from the population by natural selection (i.e., by extinction of the genotype or back mutations). This hypothesis predicts the existence of genotypes in natural populations that lack the gene loss mutations recognized in the comparisons of closely related genomes.
We investigated the presence of such polymorphisms in gene/pseudogene status using several strains and isolates of R. africae and R. massiliae. We selected four R. africae and five R. massiliae pseudogenes that were recently degraded in the respective lineages. By aligning the sequences of the pseudogenes with their full-length counterparts, we identified precise positions of the gene loss mutations (point mutations or insertion/deletions) causing in-frame stop codons or frameshifts in the pseudogene loci. Using PCR and sequencing, we examined the presence and absence of these mutations in different strains and isolates (five for R. africae and seven for R. massiliae; Table S3). We could identify one case in R. massiliae where some strains/isolates exhibited genotypes lacking the examined mutations, likely encoding a functional gene product. This suggests that genes seemingly inactivated by only a few null mutations (i.e., “split genes” [5]) may represent, in fact, polymorphic alleles and still be intact and functional in other strains of the same species.
The remaining gene loss mutations were observed in all the examined strains and isolates. It is possible that some of those mutations represent common genotypes of the species (fixed genotypes) and others are polymorphisms. In the latter case, the failure to detect genotypes lacking gene loss mutations may simply be due to our limited sample size of the strains/isolates. Further analyses of the level of polymorphism will help to quantify the time frame as well as the taxonomic scale (i.e., strain or species level) on which purifying selection acts to eliminate slightly defective genotypes from natural populations of Rickettsia.
Except for R. bellii, all other Rickettsia genomes exhibit long-range colinearities with solely a handful of genome rearrangements (Figure 6). We inferred a genome rearrangement (“inversion”) scenario after the divergence from the R1, using a parsimony method (Figure S3). According to the inference, zero to nine inversions (about 22–877 kb) in each branch leading to six Rickettsia species explain most of the genome organization differences. Half of the inversions (eight cases over 15) are symmetric to the predicted terminus of replication. Such inversions have been frequently observed [44,45]. The high genome stability of the endosymbiont Buchnera and the lack of horizontal gene transfer during the past 150 million y have been attributed to the loss of genes involved in DNA uptake and recombination in the initial stages of endosymbiosis [46]. In contrast, gene loss is not overrepresented among genes involved in DNA repair and recombination in Rickettsia. The Rickettsia core gene set contains orthologs for most of the genes involved in recombination in E. coli [47], including the RecF pathway, RuvABC, RecA, and part of the RecBCD system. R. bellii, R. felis, and R. massiliae contain a large number of nearly identical copies of transposase genes. R. bellii exhibits a highly shuffled chromosome relative to other species. The R. felis lineage also exhibits a relatively frequent chromosome inversion (nine events). Thus, the high genome stability of most Rickettsia is probably linked to the lack of highly similar DNA repeats rather than the loss of key genes involved in recombination. There is no correlation between the number of inversions and the number of nucleotide substitution in each branch.
With an abundance of genomic data coupled to highly conserved gene colinearity, the genus Rickettsia offers a convenient model for studying the phenomenon of reductive genome evolution, a common theme in obligate parasites of multicellular eukaryotes. A tentative scenario of the evolution of Rickettsia is given in Figure 7 and detailed below.
The ancestor of Rickettsiales, i.e., the mother of the Anaplasmataceae and Rickettsiaceae clades, was probably already a cellular parasite since all of its known descendants are obligate intracellular. It underwent a founding evolutionary event that resulted in the split of the ribosomal operon at two loci, a unique feature among these bacteria [48]. The RPE repeats may have appeared at this stage, as one of their families is present in both Rickettsiales and Anaplasmataceae [49].
After the divergence of the Anapalasmataceae, recent studies evidenced many gene transfers between the ancestors of Rickettsiales and intra-amoebal bacteria [8], including a conjugative operon. These bacteria probably used conjugative plasmids for gene transfer, although plasmids have been described only in a member of SFG Rickettsia [6]. They suggest that the rickettsial ancestor initiated intracellular parasitism in unicellular eukaryotes like amoebae, and later adapted to multicellular eukaryotes. At this stage, several gene families were expanded, such as those associated with the stringent response (spoT and proP), membrane proteins (sca), and energy parasitic enzymes (tlc) [6]. We propose that the specialization to multicellular eukaryotes coincided with the beginning of genome reduction. The fact that the chlamydia-related obligate symbiont of amoebae has a large genome (2.4 Mb) compared to the related obligate intracellular human/animal pathogens (∼1 Mb), illustrates the consequence of this type of host transition on the genome size. In this perspective, sequencing the recently identified members of Rickettsiales living in amoeba [50] appears as a priority to better understand the early stage of genome evolution in Rickettsia.
The reconstructed genome of the Rickettsia ancestor (R0), although bigger than those of its descendants, was highly reduced (1,254–1,700 genes) lacking important biosynthesis pathways, but possessing many genes associated with parasitism. From this ancestor, we have shown that genome size differences between the modern Rickettsia mainly result from loss of genes presumably made dispensable in their current intracellular niche. This suggests that genome reduction is still an ongoing process in the Rickettsia genus [23]. As a matter of fact, we identified 100–274 pseudogenes in each genome. The massive gene loss has not been balanced by acquisition of new genes, leading to an inexorable contraction of the genome. We could not identify any recent lateral gene transfer, though paradoxically, the R0 ancestor was predicted to contain genes involved in conjugation. This could be due to the isolation of the bacteria in specific tissues of the host limiting the contact with sources of foreign genetic material. Nonetheless, the Rickettsia genomes underwent spreading of selfish DNAs, including lineage-specific multiplication of RPE-4 and RPE-7. Our data evidenced that RPE-7 was already present in the Rickettsia ancestor. Transposases were also intensely duplicated in R. felis and even more in R. bellii. In this study, we also demonstrated that evolutionary rates (i.e., the rates of nucleotide changes, gene losses, and chromosomal inversions) were highly variable across the different Rickettsia lineages. This heterogeneity likely reflects the intricate effects of specialization to distinct arthropod hosts and the critical alterations in gene repertoire, such as the losses of DNA repair genes and the amplification of mobile genes.
The following bioinformatic pipeline was used to build an initial gene set for every Rickettsia:
Step (1) For each genome, we first generated a starting dataset of open reading frames (ORFs) equal to or greater than 40 amino acids (aa) from start (ATG, GTG, or TTG) to stop codons. These ORFs were searched against the Swiss-Prot/TrEMBL sequence database (excluding Rickettsia sequences) [51] and the NCBI/CDD database [52] with the use of BLASTP and RPS-BLAST [53]. We also used SelfID [54] to build Markov models of genes for the Rickettsia genomes and generated lists of predicted protein-coding genes. ORFs shorter than 80 aa with no detectable homology (E-value < 10−3) in the databases were discarded. ORFs between 80 and 150 aa with no detectable homology and not identified as genes by SelfID were also discarded. Other ORFs (>150 aa) were kept.
Step (2) Pairs of ORFs overlapping by more than 30% of the size of the shorter ORF were further handled as follows: the ORF exhibiting database matches with lower E-value was kept irrespective of length; if no match was found, the longest ORF was kept.
Step (3) After identifying orthologous relationships among ORFs (see below), we further discarded the groups of orthologous ORFs shorter than 100 aa that did not exhibit detectable homology (E-value < 10−5) in the databases and were not present in at least two of the three major Rickettsia groups (i.e., SFG, TG, and R. bellii).
Step (4) Finally, we aligned orthologous ORFs to select the consensus start codons when applicable.
Step (5) tRNA genes were identified using tRNAscan-SE [55]. Other RNA genes were identified using BLASTN.
Step (6) RPEs were identified using Hidden Markov models [56] based on the previously identified RPE sequences [57].
Compared to the original annotation in GenBank, the numbers of putative protein-coding sequences are reduced in the initial gene set for four Rickettsia species (R. bellii, R. felis, R. conorii, and R. typhi) and increased for R. prowazekii (Table S4). The difference mostly concerns the addition or removal of small, predicted genes. Most (99%) of the predicted genes unique to the original annotations correspond to hypothetical proteins with no sequence similarity to proteins with known functions. In contrast, 64% of the predicted genes newly added in our annotation exhibit similarities to proteins with known function.
Orthologous relationships between RIGs were determined based on the reciprocal best BLASTP match criterion as well as the conservation of gene orders. Except for R. bellii, the Rickettsia genomes exhibit a nearly perfect colinearity and few genomic rearrangements. Hence, we examined and modified the orthologous relationships determined by reciprocal best matches by verifying that the order of orthologous genes were conserved in the TG and SFG Rickettsia genomes using the Genomeview software (S. Audic, personal communication). The orthologous gene groups were named RIGs for Rickettsia genes.
Next, we flagged some of the putative protein-coding sequences in the initial gene set as pseudogenes. A RIG of a species corresponding to multiple consecutive ORFs, or corresponding to a single ORF with length shorter than 50% of the size of the longest ortholog, was defined as a pseudogene. When a RIG was missing in one of the Rickettsia genome, we looked for gene remnants (i.e., highly degraded pseudogenes) in the corresponding orthologous genomic locus using the TBLASTN program, which performs protein against translated DNA alignments. For this search, the protein product of the longest orthologous gene was used as query. We set the E-value threshold at 0.01 and requested the size of high scoring pairs to be longer than 20 aa or 20% of the query size. This data is available in Dataset S1.
First, we inferred the gene content of the ancestral R0 genome. A RIG was considered present in the R0 genome if it was found in a full-length or pseudogene state in R. bellii and at least in one species from the SFG or TG (1,175 cases). In addition, when a RIG was present only in R. bellii or the TG/SFG clade, the protein product was searched against the GenPept database using BLASTP. The RIG was then considered to be present in the R0 genome if the score of the best match among α-proteobacteria (excluding Rickettsia sequences) was greater than the score of the best match against the other organisms (77 cases). Using this procedure, 1,252 of the 1,867 RIGs could be inferred as present in the R0 genome.
Of the 615 remaining RIGs, 275 RIGs exhibited a best BLAST match among organisms outside the α-proteobacteria (excluding Rickettsia sequences). Systematic phylogenetic analyses did not reveal any convincing case of recent horizontal gene transfer among the 275 RIGs (i.e., the Rickettsia proteins being anchored in a non-α-proteobacteria clade). In fact, the Rickettsia proteins were often the most distant sequences in the trees. The remainder (340) of the 615 RIGs did not match any sequence in database and can be therefore considered as hypothetical genes. Comparison of the GC percent at first, second, and third codon positions between the conserved 1,252 R0 genes and the 615 remaining RIGs did not reveal any significant differences.
Secondly, we determined the functional status of the 1,252 R0 RIGs in each of the ancestral nodes of the Rickettsia phylogeny using the maximum parsimony criterion with an irreversible transition model, which minimizes the total number of gene loss events. Concretely, a RIG was considered in a full-length state in an ancestral node if the RIG was also present in a full-length state in at least one of its descendant nodes. Otherwise, the RIG was flagged as lost (i.e., pseudogene or absent) in the ancestral node. Then, we determined the numbers of gene losses that occurred in each branch of the phylogeny by attributing a gene loss event to a branch when a RIG was inferred as full-length in the ancestral node and lost in the child node.
Finally, the ancestral states of the 615 RIGs, the presence of which could not be inferred in R0, were determined in the other ancestral nodes using the same criterion as above. A RIG was considered present in a node i (R1 or its descendant nodes) if it was found in a full-length or pseudogene state in at least one species for each of the two descendant clades. This data was used to generate Figure 1.
Simulations of the gene loss process were performed by distributing gene loss events among the descendants of the R0 genes along the phylogeny. For each branch, we distributed the same number of loss events as inferred from the real data. In addition, we took into account the constraints imposed by the phylogenetic context: a randomly chosen RIG could be lost along a branch if (1) it was not previously lost in the phylogenetic path up to the root (R0), and (2) it was not previously lost in any of the child phylogenetic paths. Losses of a RIG in the two branches of a bifurcation were avoided, otherwise it would be considered as a single loss in the ancestral branch according to the parsimony criterion. For a new gene loss event to be attributed to one of the available full-length genes in a branch, the probability of loss for any gene i was
where αi is a scaling parameter that reflects the propensity of loss of gene i (0 ≤ α ≤ 1) and n is the number of full-length genes available for deletion. In the random model M1, we assumed equal propensities of loss among genes (α = 1 for every genes). Under model M2, the simulation of the gene loss process was carried out among the 491 genes inactivated in at least one branch of the tree. We used the same procedure as above with the additional requirement that every starting gene had to be lost at least once in the simulation. This constraint was achieved using a preprocessing step in which a first gene loss event was attributed to each RIG along a randomly chosen branch. The probability of picking a given branch was equal to the proportion of gene loss events attributed to the branch in the real data relative to the entire phylogeny. In addition, the model M2 assumed two classes of genes with different propensities of loss. The first class contained half the genes with α = a1 = 1. The second class contained the remaining genes with α = a2 ≤ 1. Note that the ratio of propensities between the two gene categories is a2/a1 = a2. The goodness of fit of the model M2 to the real data was measured by comparing the distribution of the average number of loss events per gene obtained from 100 simulations to the distribution observed in the real data using the χ2 statistics.
Protein sequences were aligned with the MUSCLE program [58]. The corresponding gene sequences were aligned using the protein alignment as a guide. We reconstructed the Rickettsia phylogeny using the concatenated protein alignment and three methods of phylogenetic analyses. The neighbor joining and maximum parsimony trees were constructed using the MEGA3 software [59], and the maximum likelihood tree was searched using the PHYML program [60]. All three methods recovered the same tree topology. Branch lengths were estimated using the maximum likelihood approach implemented in the PAML package [61]. For the protein datasets, we used the JTT amino acid substitution matrix and a gamma distribution to account for variable rates of substitutions among sites. Because the Rickettsia genomes exhibit substantial differences in GC content, we employed the nonhomogeneous HKY + N2 model of Yang and Roberts [62] to estimate the branch lengths from the FFD sites. This model accounts for unequal base frequencies. For the ω ratio analysis, reciprocal best BLASTP hits were identified between the R. conorii genome and each of the eight Anaplasmataceae, 22 coli-group, and the six other Rickettsia genomes. For 198 R. conorii genes, we could identify a reciprocal best BLAST hit in every target genomes. These genes were considered ortholog and aligned on a codon basis. Gapped positions in alignments were removed in order to keep only homologous codons conserved in all species. Pair-wise Ks, Ka, and ω = Ka/Ks values were estimated from the concatenated alignment (55,792 codons) using the codeml program [61].
We identified colinear genome segments between the chromosomes of six species (R. conorii, R. africae, R. massiliae, R. felis, R. prowazekii, and R. typhi) based on the orthologous relationships of RIGs. The colinearity was used to represent each chromosome by a string of 27 signed characters (i.e., 1–27). The sign (plus or minus) of the characters represents the direction of the segment relative to the corresponding segment of R. conorii. Each character is associated with six genomic segments (from six species), of which the average size varies from 4.6 kb to 246 kb. These segments cover most (94.9% for R. massiliae to 98.8% for R. typhi) of each chromosomal sequence. We computed minimum number of inversions for each branch of the phylogenetic tree as well as the ancestral states of the strings, using these genome strings as a input for the GRAPPA release 2.0 [63] program with a fixed tree topology option. We also used GRIMM [64] for the inference of an optimal rearrangement scenario for each branch.
We selected four and five pseudogene loci in the genomes of R. africae and R. massiliae, respectively. Through sequence alignments with orthologous intact genes in closely related Rickettsia, we identified mutations causing pseudogenization (Table S5). We confirmed these mutations by examining the assembly of sequence reads for R. africae ESF5 strain and R. massiliae MTV5 strain. For each pseudogene locus, we designed a pair of primers to examine the polymorphisms in this region using five R. africae strains or seven R. massiliae strain/isolates (one reference strain, one strain growing on tissue culture, and five isolates from ticks). DNA was extracted by the QiAmp procedure (Qiagen, http://www1.qiagen.com). PCR experiments were performed using the Taq Phusion High-Fidelity DNA Polymerase (NEBiolabs, http://www.neb.com). Different elongation temperatures were tested to optimize the PCR fragments.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers corresponding to the Rickettsia genome records are as follows: R. africae (AAUY00000000), R. bellii (CP000087), R. conorii (NC_003103), R. felis (plasmid: NC_007109, chromosome: NC_007110), R. massiliae (AAVR00000000), R. prowazekii (NC_000963), and R. typhi (NC_006142).
The GenBank accession numbers of the Anaplasmatacea genomes are as follows: Anaplasma marginale (NC_004842), A. phagocytophilum (NC_007797), Ehrlichia canis (NC_007354), E. chaffeensis (NC_007799), E. ruminantium str. Gardel (NC_006831), str. Welgevonden (NC_005295), str. wMel (NC_002978), and Wolbachia pipientis str. wBm (NC_006833).
The GenBank accession numbers of the coli-group genomes are as follows: Escherishia coli str. CFT073 (NC_004431), str. K12 (NC_000913), str. O157:H7 VT2-Sakai (NC_002695), str. O157:H7 EDL933 (NC_002655), str. UTI89 (NC_007946), str. W3110 (NC_000091), Photorabdus luminescens (NC_005126), Salmonella enterica str. Choleraesuis (NC_006905), str. Paratyphi (NC_006511), str. typhi (NC_003198), str. Ty2 (NC_004631), S. typhimurium (NC_003197), Shigella boydii (NC_007613), S. dysenteriae (NC_007606), S. flexneri str. 2a 301 (NC_004337), str. 2a 2457T (NC_004741), S. sonnei (NC_007384), Vibrio cholerae (NC_002505), V. fischeri (NC_006840), str. YJ016 (NC_005139), V. parahaemolyticus (NC_004603), and V. vulnificus str. CMCP6 (NC_004459). |
10.1371/journal.pgen.1004039 | Truncation of Ube3a-ATS Unsilences Paternal Ube3a and Ameliorates Behavioral Defects in the Angelman Syndrome Mouse Model | Angelman syndrome (AS) is a severe neurodevelopmental disorder caused by maternal deficiency of the imprinted gene UBE3A. Individuals with AS suffer from intellectual disability, speech impairment, and motor dysfunction. Currently there is no cure for the disease. Here, we evaluated the phenotypic effect of activating the silenced paternal allele of Ube3a by depleting its antisense RNA Ube3a-ATS in mice. Premature termination of Ube3a-ATS by poly(A) cassette insertion activates expression of Ube3a from the paternal chromosome, and ameliorates many disease-related symptoms in the AS mouse model, including motor coordination defects, cognitive deficit, and impaired long-term potentiation. Studies on the imprinting mechanism of Ube3a revealed a pattern of biallelic transcription initiation with suppressed elongation of paternal Ube3a, implicating transcriptional collision between sense and antisense polymerases. These studies demonstrate the feasibility and utility of unsilencing the paternal copy of Ube3a via targeting Ube3a-ATS as a treatment for Angelman syndrome.
| Angelman syndrome (AS) is a devastating neurodevelopmental disorder diagnosed in young children, currently with no effective treatments. It is characterized by absence of speech, ataxia, intellectual disability, epilepsy, and a characteristic behavior of frequent laughter and smiling. The disease is caused by loss of the maternal allele of UBE3A, which is preferentially silenced on the paternal chromosome and expressed on the maternal chromosome in neurons due to genomic imprinting. It has been long proposed that by activating the originally silenced paternal allele of UBE3A, the disease may be cured. Here in our research, we demonstrated the feasibility of activating paternal Ube3a in mice by terminating the transcription of its antisense RNA Ube3a-ATS genetically. In the AS mouse model who additionally receives the terminated Ube3a-ATS allele from the paternal side, we observed restoration of Ube3a expression, amelioration of behavioral defects and reversal of the impaired long-term potentiation. We further studied the imprinting mechanisms of Ube3a and proposed a novel transcriptional collision model. These results provide solid in vivo evidence for a key regulatory role of Ube3a-ATS in the disease and open up an exciting possibility of a gene-specific treatment for Angelman syndrome.
| Angelman syndrome (AS) is clinically manifested by features of intellectual and developmental disability, absence of speech, ataxic movement, epilepsy, and unique behaviors such as frequent laughter and fascination with water [1], [2]. Despite absence of effective treatment currently, therapeutic development for Angelman syndrome could be potentially optimistic, since patients with AS have overall normal development and brain architecture early in life.
Genetically, the disease is caused by deficiency of an E3 ubiquitin ligase termed UBE3A, which participates in many important neuronal functions such as synaptic development, signal transduction, and plasticity [3]. The gene encoding UBE3A is among a handful of human genes that are subject to genomic imprinting. In neuronal cells, it is highly expressed from the maternal allele, but silenced on the paternal allele. Disruption of the maternal allele, through genomic deletion, paternal uniparental disomy, imprinting defects, or point mutations, leads to the absence of UBE3A expression in neuronal tissues and hence Angelman syndrome. Indeed, in all cases of the disorder, at least one copy of paternal UBE3A is intact. One could speculate that by correcting the expression level of UBE3A via activating the silenced paternal allele, the disease might be treated.
Imprinted genes usually form clusters in the genome and are controlled by the imprinting center (IC). On human chromosome 15q11–q13, paternally expressed genes, including MAGEL2, NDN, SNRPN, SNORD115 and SNORD116, are critical genes for Prader-Wiili syndrome (PWS) and form an 2-Mb imprinting cluster together with the AS gene UBE3A. Although not fully understood, it is generally believed that the PWS/AS region is regulated by a bipartite imprinting center composed of PWS-IC, which activates genes located in its proximity via looping and direct interacting with them, and AS-IC, which suppresses PWS-IC by transcription-mediated DNA methylation [4], [5]. As a result of combined action of both PWS-IC and AS-IC, the paternal and maternal alleles of NDN and SNRPN show very distinct epigenetic patterns of DNA methylation and histone modifications [6], [7], [8], which define the paternal alleles as transcriptionally active and maternal alleles as transcriptionally silent.
Imprinting of UBE3A, however, is not associated with differential DNA methylation at the promoter region [9], [10]. Instead, it is regulated by its antisense RNA, UBE3A-ATS, which is expressed from the paternally inherited chromosome in the brain [11], [12]. As part of the large non-coding transcript (Shng14) initiated from the Snrpn promoter in mice [13], Ube3a-ATS expression is always negatively associated with Ube3a sense transcript. For example, when the Snrpn promoter was deleted, with or without the Prader-Willi syndrome imprinting center (PWS-IC), the Ube3a-ATS level was found to be reduced, coupling with significant up-regulation of paternal Ube3a [12], [14]. On the other hand, when maternal Ube3a-ATS was activated through replacement of the mouse imprinting center (IC) with the human one, or deletion of the putative AS-IC, maternal Ube3a was found to be repressed to some extent [15], [16]. Recently, by terminating Ube3a-ATS transcription in neuronally differentiated ES cells, we have showed that paternal Ube3a can be activated to a comparable level as maternal Ube3a [12], suggesting a direct role of Ube3a-ATS in suppressing paternal Ube3a.
In the present study, we continue evaluating Ube3a-ATS as a potential therapeutic target for treating Angelman syndrome. By characterizing a novel mouse model expressing the truncated form of Ube3a-ATS, we provide the first in vivo evidence that eliminating Ube3a-ATS is sufficient to restore Ube3a expression and improve the abnormal behaviors in the AS mouse model. Mechanisms underlying paternal Ube3a silencing are also studied, and a hypothesis of transcriptional collision between Ube3a and Ube3a-ATS is proposed.
In order to test if suppression of Ube3a-ATS alone is sufficient to unsilence the paternal allele of Ube3a, mice with the Ube3aATS-stop allele were generated by inserting the triple SV40 poly(A) cassette [12] in between Snord115 and Ube3a (chr7:66573289 NCBI37/mm9) (Figure 1). This design aims to prevent overlap between Ube3a and Ube3a-ATS and to minimize its effect on expression of the snoRNA clusters. The inserted cassette also contains a neomycin selection marker in the opposite transcriptional orientation to Ube3a-ATS to facilitate and enhance transcriptional termination. The mice were backcrossed to C57/BL6 background for six generations before subsequent expression and behavioral analysis.
We first determined the effect of the termination cassette on the expression level of Ube3a-ATS and other genes located in the imprinting cluster. The Ube3a-ATS level downstream of the insertion site (Ube3a-ATS 3′, green arrows in Figure 1B) was found to be significantly down-regulated by qPCR analysis when the stop allele was inherited paternally, while maternal inheritance of the allele has no effect (Figure 2A). To exclude the possibility that the PCR amplification site is spliced out instead of terminated, a custom designed strand-specific microarray was further performed as previously reported [12]. A significantly lower level of Ube3a-ATS was detected beyond the stop cassette insertion site (Figure S1). Expression of most other imprinted genes located nearby, including Mkrn3, Magel2, Snrpn, Snord116, and Ipw remained unchanged in both Ube3aATS-stop/+ and Ube3a+/ATS-stop mice (maternal genotype precedes the paternal genotype), indicating that the imprinting status of the PWS/AS region is not disrupted by the insertion. The level of Ndn was found to be approximately doubled in Ube3a+/ATS-stop mice compared to the other two genotypes. It is interesting that similar observation has been found in delS-U/0.9 mice previously [12], which expresses Ube3a-ATS at a lower level due to Snrpn promoter deletion. The reason for the observed up-regulation is unclear.
Ube3a mRNA is doubled in the Ube3a+/ATS-stop mice, suggesting that paternal Ube3a may be unsilenced. To confirm this, male mice heterozygous for Ube3a-ATSstop were crossed with female mice heterozygous for Ube3aKO [17] (C57/BL6 background), which is a constitutive Ube3a knock-out allele (Figure 1 and 2B). In the progeny, littermates of wild-type (WT), Ube3aKO/+ (AS), Ube3a+/ATS-stop (stop), Ube3aKO/ATS-stop (AS/stop) were compared. In the AS/stop mice, Ube3a protein was found to be activated to ∼70% of the WT level in neocortex, ∼60% in hippocampus, and ∼50% in cerebellum. The incomplete activation may be due to leaky termination of Ube3a-ATS, as about 20% of Ube3a-ATS can still be detected in Ube3a+/ATS-stop mice (Figure 2A). Immunostaining with anti-Ube3a showed that in AS/stop mice, paternal Ube3a is expressed in most brain regions, including all layers of neocortex, CA1-3 and dentate gyrus of hippocampus, and Purkinje neurons of cerebellum (Figure 2C and S2). Its expression pattern is very similar to that of maternal Ube3a in the WT mice. The incomplete unsilencing of paternal Ube3a may be due to a smaller number of Ube3a positive neurons, or a lower expression level in each single neuron, or more likely a combination of both. Male mice heterozygous for Ube3aATS-stop were also crossed with female mice heterozygous for Ube3aYFP [18], which carries the C-terminal YFP tag (Figure S3). Since Ube3a-YFP is expressed as a fusion protein with a higher molecular weight, it can be easily distinguished from wild-type Ube3a protein by western blot. Inheritance of Ube3aATS-stop from the paternal side leads to biallelic expression of Ube3a, while in contrast, maternal inheritance of the allele had no effect.
Finally, the effect of Ube3aATS-stop on paternal Ube3a was compared with the other two alleles of del4.8 and del0.9. The allele of del4.8 removes 4.8 kb of Snrpn promoter and functions as a PWS-IC deletion, while the allele of del0.9 removes 0.9 kb of Snrpn promoter and is equivalent to a Snrpn promoter deletion (Figure 1) [19]. After crossing with female mice carrying genomic deletion over the Snrpn-Ube3a region (delS-U/+, Figure 1) [20], the mRNA and protein levels of paternal Ube3a were found to be the highest in delS-U/4.8 mice, intermediate in delS-U/Ube3aATS-stop mice and the lowest in delS-U/0.9 mice (Figure S4A, B). Interestingly, such order is in accordance with the suppression level of Ube3a-ATS (Figure S4C). Plotting of paternal Ube3a against Ube3a-ATS fits into the curve of exponential decay (R2 = 0.997, Figure S4D), suggesting that suppression of paternal Ube3a by Ube3a-ATS is “dose-dependent”.
We next tested whether inheritance of Ube3aATS-stop paternally can correct the phenotypic defects in the Angelman syndrome (AS) mouse model. To address this, male Ube3aATS-stop heterozygous mice were crossed with female Ube3aKO heterozygous mice [17] (C57/BL6 background) and the littermates of WT, Ube3aKO/+ (AS), Ube3a+/ATS-stop (stop), Ube3aKO/ATS-stop (AS/stop) were studied for various AS-related phenotypes.
Obesity is associated with a small portion of AS patients [2], [9] and constantly observed in many Angelman syndrome mouse models [15], [21], [22]. Ube3aKO/+ mice become overweight starting from three month of age, in both males and females (Figure 3A, p(WT vs. AS)<0.01 for 4, 5, 6 months of age, two-way ANOVA of repeated measures). Activation of paternal Ube3a in the AS/stop mice completely reversed the obese phenotype (p(AS vs. AS/stop)<0.05 for 4, 5, 6 months of age).
The marble burying test measures repetitive behavior as potentially analogous to an autistic phenotype. Interestingly, AS mice were found to be dramatically impaired in performing this task (Figure 3B, WT: 12.36±0.75, AS: 0.50±0.27, p(WT vs. AS)<0.001, one-way ANOVA with Newman-Keuls post-hoc test). AS/stop mice showed a slight but significant improvement over AS mice (AS/stop: 3.50±0.84, p(AS vs. AS/stop)<0.05).
Hyperactivity with short attention span is a pronounced problem in young children with AS. Different from humans, AS mice have been reported to display hypoactivity [22], [23]. When placed in an open field and allowed for exploration, AS mice showed significantly lower activity level as measured by total distance and central distance traveled, movement time, and vertical activity (Figure 3C). A slight trend of improvement was consistently observed in the AS/stop mice for these parameters. However, the difference between AS mice and AS/stop mice does not reach statistical significance (one-way ANOVA with Newman-Keuls post-hoc test).
Ataxia and movement difficulty is one of the most severe defects in human AS patients and AS mouse models [17]. AS mice display severe motor coordination defects during the accelerating rotarod test (Figure 3D, p(WT vs. AS)<0.05 for all eight trials, two-way ANOVA of repeated measures). AS/stop mice show restoration in the first few trials of accelerating rotarod, although they fail to improve in later trials (p(AS vs. AS/stop)<0.05 for trial 1–4). They also show full restoration of other motor defects during wire hanging test and dowel test, indicating a significant improvement of their motor coordination skills (Figure 3E, F, and Figure S5, p(AS vs. AS/stop)<0.01 for wire-hanging test and <0.001 for dowel test, one-way ANOVA with Newman-Keuls post-hoc test). It is noted that maternal inheritance of the Ube3aATS-stop allele does not affect the performance of the mice in all three motor tests (Figure S6), suggesting that the presence of neomycin cassette has minimal or no effect on motor coordination in mice.
Individuals with AS are frequently affected with specific cognitive deficits [1], [2] and Ube3aKO/+ mice are known to have learning and memory problems [17]. During a fear conditioning test, AS mice exhibited significantly less freezing behavior than did WT littermates (Figure 3G, WT: 38.66±5.78%, AS: 21.68±5.35%, p(WT vs. AS)<0.05, one-way ANOVA with Newman-Keuls post-hoc test). Remarkably, the freezing behavior displayed in the AS/stop mice is comparable to the WT mice, suggesting that long-term memory is fully restored (AS/stop: 44.82±3.85%, p(AS vs. AS/stop)<0.01).
Lastly, we studied long-term potentiation (LTP) at Schaffer collateral–CA1 synapses, using high-frequency stimulation as the LTP-inducing protocol [17], [24]. As expected, this protocol induced a stable LTP in WT slices but caused a decaying LTP in AS slices (Figure 3H, LTP at 120 min, WT: 78±6.8%, AS: 33±8.6%, p(WT vs. AS)<0.01, one-way ANOVA). Notably, the expression of paternal Ube3a reverses the LTP deficits (AS/stop: 63±9.1%, p(AS vs. AS/stop)<0.05). The LTP rescue in AS/stop slices cannot be attributed to abnormal basal synaptic transmission, since the relation of fiber volley versus stimulation intensity, initial slope of field EPSPs versus afferent volley size, and paired pulse facilitation were unaltered in these slices (Figure S7).
In developing therapies for treating AS via activating paternal UBE3A, it is important to understand the molecular mechanism underlying genomic imprinting of Ube3a. Promoters of both the paternal and maternal UBE3A remain unmethylated in human brains [9], [10], [25], therefore DNA methylation at the promoter cannot account for silencing of paternal UBE3A. In order to look for parent-of-origin epigenetic markers that may account for UBE3A imprinting, we first analyzed histone modifications of H3K4 trimethylation (H3K4me3) in human cerebellum tissues by ChIP-on-chip experiment (Figure 4A). In contrast to healthy controls, a PWS patient with a paternal class II deletion (common 4 Mb deletion from break point 2 to 3) lacked H3K4me3 at the SNRPN promoter, suggesting that this modification is paternal specific, as previously reported [8]. However, in AS patients with maternal class II deletion, the peak of H3K4me3 was still present at the UBE3A promoter, and was indistinguishable from control and PWS samples. Therefore H3K4me3 is equally distributed between the paternal and maternal promoters of UBE3A in human cerebellum, regardless of the mono-allelic expression pattern. This conclusion from human was later supported by a recent ChIP-seq study in mice [26], in which equal enrichment of H3K4me3 and H3K27 trimethylation at both parental promoters of Ube3a was observed.
We next measured binding of the transcription preinitiation complex (PIC) at the Ube3a promoter by chromatin immunoprecipitation (ChIP). The PIC is a large protein complex composed of RNA polymerase II, TATA binding protein (TBP), TFIIB, and many other proteins assembled at the promoter of active genes. ChIP with antibody against RNA polymerase II was performed in brain samples of F1 hybrid of C57 (female) crossed with B6.Cast.Chr7 (male), which carries Mus. musculus castaneus chromosome 7 on the Mus. musculus domesticus C57BL/6 background. Single nucleotide polymorphisms (SNPs) between the two lines allow detection of parental specific alleles. In contrast to the Snrpn promoter, from which only the transcriptionally active paternal allele was precipitated, both parental alleles of the Ube3a promoter can be detected in the same IP fraction (Figure 4B). ChIP in F1 hybrids of the reciprocal cross and ChIP with anti-TFIIB and anti-TBP revealed the same result (Figure S8). Altogether, the results suggest that the PIC is able to be properly assembled at the promoter of both paternal and maternal Ube3a alleles, regardless of theirs different expression status.
Since paternal Ube3a shows multiple features of an active gene as we demonstrated above, we considered the hypothesis that it is actually transcriptionally active despite the absence of mature mRNA. To test this, a mouse model carrying a deletion from Ube3a to Gabrb3 (delU-G, Figure 1A and 4C) was used [27]. Since the deletion covers the promoter of Ube3a, only maternal Ube3a RNA is present in the paternal deletion +/delU-G mice and only paternal Ube3a RNA is present in the maternal deletion delU-G/+ mice. We set the RNA copy number of maternal Ube3a in +/delU-G mice equal to 1 across different portions of the gene and used it as the reference to calculate the relative RNA copy number of paternal Ube3a in delU-G/+ mice or total Ube3a in WT mice. Consistent with the known mono-allelic expression pattern, the mature mRNA of paternal Ube3a (quantified by qPCR using primers spanning exon-exon junction) is about 0.2 copy at both the 5′- and 3′-portions in delU-G/+ mice (Figure 4C, ex1-4, ex4-6, and ex12-13). However, when pre-mRNA of paternal Ube3a was quantified (by strand-specific qRT-PCR [28] using tagged primers directed to introns), it is around one copy at the 5′ portion of Ube3a (black bars of int1, int3, and int4.2 in Fig. 4C) and drops to about 0.2 copy as the primers are moved to the 3′ portion of Ube3a (black bars of int4.4, int6, and int12 in Fig. 4C). Altogether, our data supports a model that paternal Ube3a is transcribed at a comparable level as maternal Ube3a from the promoter, but later becomes suppressed during the process of transcription elongation.
Airn and Kcnq1ot1, two antisense RNAs playing a regulatory role in their respective imprinting cluster, are known based on FISH analysis to be localized around the transcribed regions [29], [30], consistent with their functional roles. Ube3a-ATS has been shown to be localized exclusively to the nucleus [12], [31], but the subnuclear detail was unknown. To address this question, a combined RNA/DNA FISH was performed in mouse brain sections. Signals of Ube3a-ATS form a single bright dot inside the nucleus (Figure 5A) and can be observed in multiple regions throughout the brain including olfactory bulb, neocortex, hippocampus, cerebellum, and hindbrain. Interestingly, the signal co-localizes with only one of the two foci formed by Ube3a DNA signal (Figure 5B). Such co-localization is not random overlapping between the DNA and RNA probes because Ube3a-ATS does not overlap with the control DNA probe (targeting an irrelevant gene on mouse chromosome 4). Therefore, similar to Airn and Kcnq1ot1, Ube3a-ATS remains located proximate to its transcription site after it being synthesized.
Patients with Angelman syndrome suffer from developmental delay, speech impairment, and epilepsy. Therapies for AS are limited and focus mainly on symptomatic management [2]. Recently, topoisomerase inhibitors have been identified as the first compounds to successfully unsilence paternal Ube3a in mice [32], [33]. In the current research, we investigated a potential therapeutic strategy by activation of the silenced paternal allele of UBE3A via suppressing its antisense RNA. Previous studies have defined Ube3a-ATS as the negative regulator of Ube3a imprinting [12], [14]. However, it was unknown if depletion of Ube3a-ATS without modulating other epigenetic factors is sufficient to activate paternal Ube3a. This question is crucial in determining whether knock-down of Ube3a-ATS is a suitable strategy for treating AS. By generating a mouse model with Ube3a-ATS being prematurely terminated, we observed unsilencing of Ube3a in multiple brain regions, implying that the antisense RNA plays a regulatory role in modulating Ube3a imprinting. We then compared mice which express paternal Ube3a on the maternal Ube3a knock-out background (AS/stop) with AS and WT mice. The AS/stop mice exhibit complete reversal of obesity, motor tests of wire-hanging and dowel walking, fear conditioning defect, and plasticity-related electrophysiology. They also display slight but significant improvement in the tests of accelerating rotarod and marble burying. Therefore, our research confirmed the clinical benefit of activating paternal Ube3a in treating Angelman syndrome and provided a mouse model as the positive control for future drug testing. Given the conservation of the PWS/AS region between mouse and human, activation of paternal UBE3A through inhibiting UBE3A-ATS expression/transcription should be a promising strategy for developing AS therapy.
One important question in activating paternal UBE3A is how much UBE3A protein is needed to achieve phenotypic improvement in AS patients. In the mouse model of AS/stop, we observed some phenotypic reversal, such as obesity, and cognitive deficits. However, their performance during accelerating rotarod and marble bury test is only partially or moderately improved and their decreased locomotive activity is not restored. This may be due to the incomplete activation of paternal Ube3a, which is quantified to be 50–70% of the WT level in different parts of the brain by western blot. Some of the behavioral phenotypes might be more sensitive to the protein level of Ube3a and therefore are more difficult to reverse. Another possibility is the interference from the remaining neomycin cassette. However, paternal inheritance of the cassette on the WT background does not affect mouse behaviors, and maternal inheritance of the cassette does not change Ube3a expression and rotarod performance in mice, suggesting that the presence of the selection marker has no or minimal effect on Ube3a function. Among the human UBE3A mutations that have been reported so far, there is a striking preponderance of frameshift and nonsense mutations [34]. It is possible that individuals with less pathogenic missense mutation in UBE3A display some, but not all, clinical features associated with AS and thus are excluded from AS diagnosis and research. A patient with C21Y missense mutation located outside the HECT domain of UBE3A has been reported to have a less classical phenotype [35], suggesting that partial activity of UBE3A may be beneficial.
Another relevant issue is to understand the molecular mechanism underlying UBE3A imprinting. Interestingly, several pieces of evidence have suggested that the paternal allele of UBE3A/Ube3a is transcriptionally active. For example, the promoter of paternal UBE3A is unmethylated [9], [10], [36], modified with active histone markers (Figure 4A), and bound with transcription pre-initiation factors (Figure 4B). Indeed, Ube3a pre-mRNA can be detected equally from the 5′-portion of both paternal and maternal alleles in mice (Figure 4C). Therefore, paternal Ube3a is transcriptionally active and its suppression may occur during the process of transcription elongation. The previous observation of “biallelic” expression pattern at the 5′-portion of mouse Ube3a by SNP analysis is consistent with this conclusion [37].
As demonstrated in this and many other studies, Ube3a-ATS has a direct role in silencing paternal Ube3a. However the detailed mechanism is unclear. Research on the other two imprinted ncRNA Airn and Kcnq1ot1 has raised two different working models, promoter occlusion and RNA-directed targeting. When silencing the overlapping gene in embryonic tissues, Airn transcribes through the Igf2r promoter and precludes binding of RNA polymerase II to the Igf2r promoter [38]. In contrast, when silencing the respective non-overlapping genes in extraembryonic tissues, the RNA product of Kcnq1ot1 or Airn will bind to trans-acting protein factors and induce repressive higher-order chromatin changes [29], [39], [40], [41]. Can either of the two models be applied to Ube3a-ATS? Promoter occlusion is unlikely to be the cause of Ube3a imprinting since paternal Ube3a promoter is transcriptionally active. Components of PIC such as RNA polymerase II, TBP, and TFIIB are found to bind paternal and maternal Ube3a equally. Currently, it is unknown whether the RNA product of Ube3a-ATS is essential in mediating Ube3a imprinting. However, Ube3a-ATS has very low homology between mouse and human, and is quickly degraded [12], implying a low functional importance of the RNA product.
Here we proposed an alternative hypothesis of transcriptional collision as the mechanism for Ube3a-ATS mediated Ube3a imprinting (Figure 6). Our previous strand-specific microarray data revealed a significant decrease of Ube3a-ATS RNA signal around intron 4 of Ube3a, although the transcript remains detectable until ∼40 kb upstream of Ube3a promoter [12]. Interestingly, this is around the same region where the pre-mRNA level of paternal Ube3a becomes suppressed. Therefore, on the paternal chromosome, Ube3a sense and antisense RNAs are transcribed head-to-head at a relatively high level until the polymerases reach intron 4, where both drop to a lower level. These findings are similar to what has been described for transcriptional collision occurring during convergent transcription [42]. Such collision will result in stalling, dissociation of both polymerases, and abortive transcription of both. Research in budding yeast has demonstrated both in vitro and in vivo that convergent transcription will result in collision of the two opposing polymerases [42], [43]. The collision event can also be detected by atomic force microscopy in vitro when two promoters are aligned convergently on a linear DNA template [44]. Currently, we still lack direct evidence to support the Ube3a transcriptional collision hypothesis. It will be necessary to test it in the future by mapping RNA polymerase II stalling sites along Ube3a using GRO-seq or NET-seq technology [45], [46].
All animal procedures were performed in accordance with NIH guidelines and approved by the Baylor College of Medicine Institutional Animal Care and Use Committee (IACUC). All human studies were performed in accordance with NIH guidelines and approved by the Baylor College of Medicine Institutional Review Board (IRB).
The insertion cassette composed of SV40 triple poly(A) signal and neomycin selection marker was inserted downstream of Ube3a by gene targeting in wild-type AB2.2 ES cells. After microinjecting into blastocysts of C57/BL6 mice, high percentage male agouti chimeras were obtained and germline transmission was established. The lines were then backcrossed to C57/BL6 mice for more than six generations. PCR genotyping was developed with TS-F (TTCCCAGTGCTGAGACTAAAG), TS-R (CCACAATCTGAA-CCCTAAAAC) and SV40-R (AAAAGGGACAGGATAAGTATG).
Total RNA was prepared with miRNeasy Mini Kit (Qiagen, Valencia, CA). On-column DNase treatment was performed for all the samples. The cDNA was generated using 0.2–1 µg of total RNA with SuperScript III First-Strand Synthesis System (Invitrogen, Carlsbad, CA), and qRT-PCR was performed using Applied Biosystems StepOnePlus Real-Time PCR System and SYBR Green Master Mix (Applied Biosystems, Carlsbad, CA). Primers used are listed in Table S1.
Western blot against Ube3a and β-tubulin was performed as previously described [12]. Quantification was performed based on densitometry with ImageJ.
Tissue preparation and immunohistochemistry were performed by Neuropathology Core of Baylor College of Medicine, as previously described [47]. Immunostaining was carried out with Rabbit polyclonal anti-Ube3a (1∶500, A300-352A, Bethyl Laboratories, Montgomery, TX) and horseradish peroxidase conjugated goat anti-rabbit (1∶200, Dako Inc., Carpinteria, CA). The localization of the antibody was visualized using diaminobenzidine (DAB, 0.5 mg/ml, Vector Laboratories Inc., Burlingame, CA) as a chromogen.
A battery of behavioral tests was performed using a protocol previously described and used in Behavioral Core facilities at Baylor College of Medicine [27], [48]. A detailed protocol for each test is described in Text S1. Tests start when the mice are 2 month-old in both males and females and the order of tests are kept the same as listed in the supplementary material. The interval between two tests is one week, except wire hanging, dowel tests and rotarod were performed in two consecutive days.
Horizontal hippocampal slices (350 µm) were cut with a Leica (VT 1000S) vibratome (Buffalo Grove, IL) from brains of WT, AS and AS/stop mice in 4°C artificial cerebrospinal fluid (ACSF) and kept in ACSF at room temperature for at least one hour before recording, as previously described [49], [50]. Slices were maintained in an interface-type chamber perfused (2–3 ml/min) with oxygenated ACSF (95% O2 and 5% CO2) containing in mM: 124 NaCl, 2.0 KCl, 1.3 MgSO4, 2.5 CaCl2, 1.2 KH2PO4, 25 NaHCO3, and 10 glucose. Bipolar stimulating electrodes were placed in the CA1 stratum radiatum to excite Schaffer collateral and commissural fibers. Field EPSPs were recorded at 30–31°C, with ACSF-filled micropipettes. The recording electrodes were placed in the stratum radiatum and the intensity of the 0.1 ms pulses was adjusted to evoke 40–50% of maximal response. A stable baseline of responses at 0.033 Hz was established for at least 20 min. Tetanic LTP was induced by using two 1 s, 100 Hz tetani, 20 s apart at baseline stimulus intensity, as previously described [17].
Postmortem brain tissues from control, PWS, and AS individuals were obtained from NICHD Brain and Tissue Bank for Developmental Disorders from University of Maryland School of Medicine.
ChIP-on-chip analysis was performed as previously described [51]. Immunoprecipitation was performed with Protein A Dynabeads (Invitrogen) coated with normal rabbit IgG or anti-H3K4me3 antibodies (17-614, Millipore) according to manufacturer's instructions. Precipitated and input DNA was amplified with GenomePlex Complete Whole Genome Amplification (WGA) kit (Sigma, St. Louis, MO) and labeled with Cy3 (input DNA) or Cy5 (ChIP DNA) using BioPrime Array CGH Genomic Labeling System (Invitrogen). The DNA was then applied to a custom designed human chromosome 15q11.2–q12 focused array (Agilent, Santa Clara, CA), with genomic tilling probes covering regions from MAGEL2 to GABRB3 (chr15:21,361,151–25,487,147, genome build NCBI36/hg18) in the 4X44k format. Hybridization, wash and scanning were performed according to manufacturer's instructions. The image files were processed with Agilent Feature Extraction software using protocol CGH-v4_95_Feb07 and further analyzed with Agilent G4477AA ChIP Analytics 1.3 software.
Brain tissues of 50 mg from newborn mice was chopped into fine pieces, crosslinked with 1% formaldehyde in DMEM and lysed in SDS lysis buffer (50 mM Tris, 10 mM EDTA, 1% SDS). The lysate was then sonicated (Fisher Scientific 500 Sonic Dismembrator) and centrifuged. The supernatant was collected and combined with IP buffer (2 mM Tris, 15 mM NaCl, 0.2 mM EDTA, 0.1% Triton X-100, 1× proteinase inhibitor). Immunoprecipitation was then performed with Protein G Dynabeads (Invitrogen) coated with anti-pol II (05-623, Millipore), anti-TFIIB (sc-225, Santa Cruz Biotechnology), or anti-TBP (MAB3658, Millipore) overnight.
Immunoprecipitated DNA was PCR amplified with Snrpn or Ube3a promoter primers, purified with MinElute PCR purification kit (Qiagen) and analyzed by Sanger sequencing to identify allelic SNPs. Alternatively, unpurified PCR products were digested with BsaI for the Snrpn promoter or BbsI for the Ube3a promoter, and analyzed by eletrophoresis on a 1.5% agarose gel.
Total RNA of 200 ng from cortices of newborn mice was used as the input in the analysis. The cDNA synthesis was performed using tagged gene-specific primers in the RT reaction to detect Ube3a pre-mRNA in a strand-specific manner [28], and then amplified in the SYBR Green q-PCR system using the tag as the reverse primer and locus specific forward primer. All primers used are listed in Table S2.
Tissue preparation and RNA FISH were carried out by RNA In-Situ Hybridization Core at Baylor College of Medicine as previously described [52]. Briefly, brains of adult mice were embedded in O.C.T., fresh frozen, and sectioned sagittally at 25 µm thickness. After paraformaldehyde fixation, acetylation, and dehydration, the slides were assembled into flow-through hybridization chambers and placed into a Tecan (Mannedorf, Switzerland) Genesis 200 liquid-handling robot. The DIG labeled RNA probes were prepared with in-vitro transcription and corresponded to chr7:66,530,657–66,531,391(NCBI37/mm9). Primers for DNA template synthesis are SP6-Ube3a-int5.1F ATTTAGGTGACACTATAGAAGCGAAGATGAGTCAG-TTTGGTTTT and T7-Ube3a-ex6.1R TAATACGACTCACTATAGGGAGATTCTGAGTCTTCTTCCATA-GC). The T7 promoter was used to generate Ube3a-ATS probe. Hybridized probes were detected by a dual amplification strategy and visualized by Alexa488 conjugated streptavidin [52].
After RNA FISH, the slides were washed in 2XSSC at 37°C for 15 min, dehydrated in 70%, 85%, 95% ethanol at −20C for 2 min each and denatured in 70% formamide/2XSSC at 70°C for 2 min. After washing with 70%, 85% and 100% ethanol, the slides were air-dried before hybridization with the DNA FISH probe. The probe was prepared from Ube3a BAC clone bMQ311i10 (Source BioScience, UK) or Lepre1 (chr 4) with FISH Tag DNA Red Kit (Invitrogen) and hybridized to the sections at 37°C overnight. The slides were washed in 50% formamide/2X SSC solution twice for 8 min at 42°C, once in 2XSSC for 8 min at 37°C and mounted with SlowFade Gold antifade reagent (Invitrogen).
Statistical analysis was performed with GraphPad Prism 5 (GraphPad Software, Inc. La Jolla, CA). One-way ANOVA with Newman–Keuls post-hoc test and two-way ANOVA with repeated measures were used.
|
10.1371/journal.pntd.0004238 | Differential Impact of LPG-and PG-Deficient Leishmania major Mutants on the Immune Response of Human Dendritic Cells | Leishmania major infection induces robust interleukin-12 (IL12) production in human dendritic cells (hDC), ultimately resulting in Th1-mediated immunity and clinical resolution. The surface of Leishmania parasites is covered in a dense glycocalyx consisting of primarily lipophosphoglycan (LPG) and other phosphoglycan-containing molecules (PGs), making these glycoconjugates the likely pathogen-associated molecular patterns (PAMPS) responsible for IL12 induction.
Here we explored the role of parasite glycoconjugates on the hDC IL12 response by generating L. major Friedlin V1 mutants defective in LPG alone, (FV1 lpg1-), or generally deficient for all PGs, (FV1 lpg2-). Infection with metacyclic, infective stage, L. major or purified LPG induced high levels of IL12B subunit gene transcripts in hDCs, which was abrogated with FV1 lpg1- infections. In contrast, hDC infections with FV1 lpg2- displayed increased IL12B expression, suggesting other PG-related/LPG2 dependent molecules may act to dampen the immune response. Global transcriptional profiling comparing WT, FV1 lpg1-, FV1 lpg2- infections revealed that FV1 lpg1- mutants entered hDCs in a silent fashion as indicated by repression of gene expression. Transcription factor binding site analysis suggests that LPG recognition by hDCs induces IL-12 in a signaling cascade resulting in Nuclear Factor κ B (NFκB) and Interferon Regulatory Factor (IRF) mediated transcription.
These data suggest that L. major LPG is a major PAMP recognized by hDC to induce IL12-mediated protective immunity and that there is a complex interplay between PG-baring Leishmania surface glycoconjugates that result in modulation of host cellular IL12.
| Leishmaniasis is a group of parasitic diseases caused by intracellular protozoa belonging to the genus Leishmania, pathological manifestations ranging from self-healing cutaneous forms to severe visceral infections that result in death. These clinical outcomes are dictated by the Leishmania species initiating the infection and are influenced by early responses of host immune cells, which ultimately initiate an IL12 mediated immune response in resolving infections. Like the diseases themselves, the magnitude of IL12 induction in hDCs is Leishmania-species and strain specific, where species that elicit visceral disease do not induce IL12, while most cutaneous disease-causing L. major strains induce robust IL12 responses and confer life-long immunity. The molecular mechanisms that mediate the ability of these innate immune cells to discriminate between pathogens remain elusive and have been primarily investigated in murine model systems. Here we identified L. major LPG as a major PAMP that induces IL12 in hDCs. Elucidation of this critical component of human immunity to L. major has ramifications for leishmaniasis vaccine development.
| Leishmaniasis constitutes a group of vector-borne parasitic diseases that affects approximately 12 million people worldwide and results in diverse clinical pathologies [1]. The causative intracellular protozoa belonging to the genus Leishmania, generally dictate disease outcome in a distinct species-specific manner. Visceral leishmaniasis may result from infection with Leishmania donovani parasites that disseminate throughout the body, manifesting into fatal systemic disease if left untreated. In contrast, Leishmania major, which is a causative agent for cutaneous leishmaniasis, produces ulcerative lesions localized at the site of sand fly vector inoculation. In the majority of L. major patients, lesions heal within several months, conferring life-long acquired immunity [2]. Recovery of cutaneous leishmaniasis with a strong immune response can be attributed to early cellular activities that occur following initial entry of the parasites into host cells.
Leishmania parasites have evolved mechanisms to survive within host cells and mediate infectivity in sand fly vectors through the interaction of their cellular surface coat molecules. The Leishmania surface coat is densely packed with glycosylphosphatidylinositol (GPI)-anchored glycoconjugates, including lipophosphoglycan (LPG), proteophosphoglycans (PPGs), glycosylinositolphospholipids (GIPLs), and glycoprotein 63 (GP63) [3–5]. Together these molecules provide a protective barrier for parasites to persist within the host environment [6]. LPG is one of the most intensely studied Leishmania surface molecules, in both the sand fly vector and vertebrate hosts, playing a distinct role in modulating host immune function [7] and even vectorial capacity of various sand fly species [8]. LPG is polymorphic among Leishmania species and developmentally regulated [6]. One dominant feature of LPG, the phosphoglycan repeating unit [Gal-Man-P] (PG), contains species-, strain-, and stage-specific modifications usually on the Gal residues [9–13]. The number of PG repeat units almost doubles during metacyclogenesis [14] and LPG is dramatically down regulated in the amastigote stage [15]. Thus, the role of LPG in mammalian infections is limited to the initial period of invasion and establishment of infection by metacyclic promastigotes.
Protective immunity to cutaneous leishmaniasis requires a robust IL12 driven type 1 helper T-cell (Th1) mediated response that produces high levels of interferon-gamma (IFNG), which ultimately promotes anti-microbicidal production of nitric oxide (NO) and reactive oxygen species (ROS) that destroy invading pathogens [16,17]. Dendritic cells (DCs) and macrophages are among the major cell sources of IL12, whose bioactive secretion is dependent upon the covalent linkage between the p40 (IL12B) and p35 (IL12A) subunits [18]. The ability of Leishmania to selectively suppress IL12 production, as first established by using murine macrophages [19,20], occurs through the transcriptional inhibition of the IL12B promoter [21] and is one immune evasion strategy employed by parasites to establish infection. Phagocytosis of Leishmania parasites by murine DCs induces IL12, driving the differentiation of Th1 cells to elicit their effector function [22–27]. The precise role of different DC subsets during murine infection in vivo is discordant depending on the Leishmania strain utilized, the infection route, and the timing of analysis [28,29]. A role for DCs early in infection has been identified in vivo, however, as DCs carrying Leishmania antigen produce IL12 within 8 hours following infection [30]. The murine DC IL12 response can be altered depending on the biochemical composition of the parasite surface, as evidenced by a study demonstrating that infection with L. major LV39c5 lpg2−, a mutant that lacks phosphoglycan (PG)-containing molecules and other LPG2-dependent metabolites [31], induced IL12B in bone marrow derived mouse DCs (BMDCs) co-stimulated with anti-CD40 and IFNG [32]. This effect along with the long-term persistence of these parasites likely account for why vaccination with these LV39c5 lpg2− parasites protects mice against L. major wild type (WT) challenge [33].
Remarkably, hDCs exhibit a dynamic range in IL12 production in response to Leishmania infection that is largely dependent upon the nature of the infecting species or strain. L. donovani fails to elicit IL12, whereas a general induction of IL12 is observed during L. major infections [34]. However, IL12 production also varies across L. major strains. Strains LV39 and SD do not induce IL12, whereas Friedlin V1 (FV1), IR173, IR176, and CC-1 strains elicit high levels of IL12 [34,35]. These differences are not well-correlated with LPG structural polymorphisms, as L. major LV39cl5 bears a highly poly-galactosylated LPG [36], while L.major SD synthesizes an unsubstituted LPG similar to that of L. donovani [37]. Several groups have reported differences in lesion pathology following in vivo infection with these same L. major strains. For example, L. major FV1 infected C57BL/6 mice develop lesions that eventually heal over time, whereas mice infected with L. major SD produce non-healing lesions [38]. BALB/C IL4RA knockout mice are resistant to L. major IR173 strain but susceptible to L. major LV39 strain [39]. Moreover, while L. major FV1 strain infected BALB/C mice quickly develop lesions, L. major LV39c5, a clonal derivative of the LV39 strain, elicits slower lesion development. Hybrid crosses of L. major FV1 x LV39c5 segregate at a 1:1 ratio into “fast” or “slow” virulence progeny [40]. These differential host responses to variant intra-species strains of L. major have important implications for the parasite strain-specific factors that could dictate disease persistence versus healing and induction of immunity.
In this study, we focus on elucidating whether parasite surface molecules are associated with the robust cytokine response observed in hDCs using the ‘high-IL12 inducing’ L. major FV1 strain. We generated parasite mutants lacking LPG alone, as done previously with the ‘low-IL12 inducing’ L. major LV39c5, through inactivation of the LPG1 galactofuranosyl transferases required for LPG core synthesis. Mutants generally lacking in all PG-containing structures were generated through inactivation of the Golgi GDP-mannose nucleotide sugar transporter gene, LPG2 [31]. This approach is powerful for probing the role of LPG as it allowed us to assess the impact of LPG deficiency in the context of the parasite, rather than through exogenous and relatively artificial routes. A second advantage is that multiple mutants provided a means to discriminate between LPG effects and those of molecules that bear structures related to or shared with those found in LPG. Notably, the PG repeating units present on LPG also are abundant on secreted molecules, such as acid phosphatases and other PPGs, which can be anchored to the parasite surface through glycosylphosphatidylinositol (GPI). Inactivation of LPG1 results in a parasite lacking LPG alone but otherwise normal in GIPL and PPGs levels [41].
Our results demonstrate that hDC infection with the LPG-null L. major FV1 lpg1− mutant resulted in significantly diminished IL12B mRNA, relative to FV1 WT parasites, indicating that LPG is essential for stimulating host IL12 production. However, the PG-null L. major FV1 lpg2− mutant infected DCs exhibited an increase in IL12B expression, suggesting that PGs and/or other LPG2-dependent metabolites may suppress IL12 induction. These results suggest that L. major parasites balance stimulatory and inhibitory effects on the host cells to establish infection.
The study protocol was approved by the University of Notre Dame Institutional Review Board in compliance with all applicable Federal regulations governing the protection of human subjects (Human Subjects Assurance #M1262). The research was deemed exempt under exemption #4. The samples were purchased from Central Indiana Regional Blood Center, Indianapolis, IN and no identifying information was provided.
Monocytes were isolated from healthy human donor buffy coats (Central Indiana Regional Blood Center, Indianapolis, IN) by enriching for CD14+ cells using a magnetic bead separator (AutoMACs, Miltenyi Biotech systems, Germany). Monocytes from each donor were cultured in 6-well plates at a concentration of 106 cells/2ml of RPMI-complete media (10% heat-inactivated FBS, 2mM l-glutamine 100U/ml, 1% penicillin/streptomycin) and supplemented with recombinant human IL4 (40U/ml, Peprotech, NJ) and granulocyte-macrophage colony-stimulating factor, GMSCF (1000U/ml, Peprotech, NJ) on days 0, 3, and 6 to allow differentiation into immature DCs. Cells were harvested, washed one day before infection to remove any residual cytokines, and assessed for DC marker CD1A to confirm a homogenous population of immature DCs. All parasite strains were cultured at 26°C without CO2 in M199 medium containing 10% heat-inactivated FBS [42]. Metacyclic promastigotes were isolated according to previously described methods [43] and opsonized by treatment with 5% human serum for 30 min at 37°C. DCs were then infected at a concentration of 10 parasites per 1 DC in RPMI-complete media. As we previously demonstrated that the peak of IL12B expression occurs at 8 hours post L. major infection [44] and to avoid the complication that mutant parasites might be degraded at later time points as previously observed [31,41], samples were typically harvested at 8 hours post-infection. For kinetic analyses we focused on the early time points following infection (2, 4, 8, or 24 hours). Cytospins were prepared at the conclusion of each experiment and Diff-quick stained (Fischer Scientific, Pittsburgh, PA) for visual analysis by light microscopy. Uninfected and infected DCs (100 total) were counted to calculate the infection rate (% infected DCs) and the parasite indices (# parasites per 100 cells) for each infection sample. All parasite and human cell cultures tested negative for mycoplasma (PCR detection, Takara) and tested below the limits of detection for endotoxin (<0.25U/ml) (Limulus Amoeboctye Assay, Endosafe, Charleston, NC).
Leishmania major strain Friedlin clone V1 (MHOM/IL/81/Friedlin) and L. donovani strain 1S (MHOM/SD/62/1S) were grown in M199 medium containing 10% heat-inactivated FBS [45]. Methods for electroporation of logarithmic phase promastigotes and plating on semisolid media to obtain clonal lines were as described previously [46].
L. major FV1 lpg1− mutants were obtained by a gene disruption strategy, in which autonomous drug resistance cassettes were inserted within the LPG1 coding region [41]. The methods and constructs used were the same as in the prior study generating the L. major LV39c5 lpg1− mutants [41]. In the first round, plasmid B2947 DNA was digested with restriction enzymes XhoI and HindIII to yield the LPG1::HYG targeting construct, conferring selective resistance gene to hygromycin B (hygromycin phosphotransferase). 10μg of DNA was used for electroporation and parasites were plated on semisolid medium containing 50μg/ml of hygromycin B. Clonal parasite lines were obtained at typical frequencies and screened for the presence of the expected heterozygous LPG1 and LPG1::HYG insertion by PCR (S1 Fig, S1 Table). Several clones were inoculated into susceptible BALB/C mice (107 stationary phase, footpad) and recovered after 1 month; such mouse passaged lines are designated as ‘M1’. These heterozygotes underwent a second round of transfection; electroporating 10μg of LPG1::PAC, conferring a selective resistance gene to puromycin (puromycin acetyltransferase), derived from BamHI digestion of plasmid B2949, and followed by plating parasites on semisolid media containing 50μg/ml hygromycin B and 30μg/ml puromycin. Clonal lines bearing disruptions in both LPG1 alleles, and lacking unmodified LPG1 (^LPG1::HYG/^LPG1::PAC), were identified by PCR analysis and confirmed by Western blot analysis and agglutination tests. Several clones were inoculated into susceptible BALB/C mice (107 stationary phase, footpad) and recovered after 1 month (M1). For simplicity, these lines are referred to as FV1 lpg1−. To generate complemented ‘add back’ lines, several FV1 lpg1− clonal lines were electroporated with the LPG1 expression plasmid pSNBR-LPG1::NEO (B3340), conferring an episomal selective resistance gene to the aminoglycoside antibiotic G418 via expression of the neomycin phosphotransferase gene NEO, and clonal lines were recovered by plating on semisolid media containing 50μg/ml HYG, 30μg puromycin, and 12μg/ml of G418. Successful transfection was established by PCR tests and restoration of LPG expression by western blot, and agglutination tests. Formally, the genotype of such lines is (^LPG1::HYG/^LPG1::PAC/+pSNBR-LPG1), which for simplicity is referred to as FV1 lpg1−/+LPG1. Sibling clonal lines displayed similar phenotypes and one representative FV1 lpg1− line (cl2.10, M1), and its complemented offspring (cl2.10 AB3, M1), designated FV1 lpg1−/+LPG1 were used in the experiments.
L. major FV1 lpg2− mutants were obtained by a gene replacement strategy; where the drug resistance gene ORFS replaced the LPG2 coding region. In the first round, plasmid B3950 was digested with XhoI I, yielding the LPG2::HYG targeting construct; 10μg was used for electroporation and cells were plated on semisolid medium containing 50μg/ml of hygromycin B. Clonal lines were obtained at typical frequencies and screened for the presence of the expected heterozygous LPG2 and LPG2::HYG insertion by PCR (S2 Fig, S1 Table). Several clones were inoculated into susceptible BALB/C mice (107 stationary phase, footpad, M1). These heterozygotes underwent a second round of transfection, electroporating 10μg of LPG2::SAT, conferring a selective resistance gene to nourseothricin (streptothricin acetyltransferase), derived from XhoI, HindIII digestion of plasmid B6598, followed by plating on semisolid media containing 50μg/ml hygromycin B and 100μg/ml nourseothricin. Clonal lines bearing disruptions in both LPG2 alleles and lacking unmodified LPG2 (ΔLPG2::HYG/ ΔLPG LPG2::SAT) were identified by PCR analysis, and confirmed by Western blot analysis and agglutination tests. For simplicity, these lines will be referred to as FV1 lpg2−. To generate complemented ‘add back’ lines, several FV1 lpg2− clonal lines were electroporated with the LPG2 expression plasmid pXG-LPG2::NEO (B4296) and clonal lines recovered by plating on semisolid media containing 50μg/ml HYG, 100μg/ml SAT, and 15μg/ml of G418. Successful transfection was established by PCR tests and restoration of LPG and the PPGs region expression by western blot, and agglutination tests. Formally, the genotype of such lines is (ΔLPG2::HYG/ ΔLPG LPG2::SAT/+pXG-LPG2), which for simplicity is referred to as FV1 lpg2−/+LPG2. Sibling clonal lines displayed similar phenotypes and one representative FV1 lpg2− line (cl6.1A, M1), and its complemented offspring (cl6.1A AB15, M1), designated FV1 lpg2−/+LPG2 were used in the experiments.
Plasmid B6598 was generated by a fusion PCR strategy. Briefly, the 5’LPG2 flanking sequence, 3’LPG2 flanking sequence, LPG2 ORF, and selected drug marker, SAT ORF were amplified by PCR and inserted into the pGEM-T-Easy vector by TA cloning according to manufacturer’s instruction (Promega, Madison, WI) and transformed into E. coli. Its structure was confirmed by DNA sequencing. The primers used for constructing B6598 are provided in S1 Table.
For Western blot analysis of PG-containing molecules, parasites were grown to logarithmic phase and harvested for cell lysate preparation in 4X Lamelli buffer (50 mM Tris-HCl pH 6.8, 2% SDS, 10% Glycerol, 1% 2-mercaptoethanol, 12.5 mM EDTA, and 0.02% Bromophenol Blue). Samples were separated on 10–12% SDS-PAGE gels at a concentration of (3.5x106 cells/well) and transferred onto methanol activated nitrocellulose membrane for 3 hrs at 60V, 4°C. Ponceau staining was performed to assure macromolecule transfer prior to blocking in 5% milk overnight. Membranes were stained with primary mouse monoclonal anti-sera WIC79.3 antibody (1:1000), recognizing galactosylated Gal-Man-P repeats on LPG, and detected using a goat anti-mouse HRP conjugated secondary antibody (1:5000) (Invitrogen, Carlsbad, CA). Membranes were developed using West-Pico detection solution assay (Thermo Scientific, Rockford, IL) and an X-ray film developer.
LPG was isolated from 109 L. major FV1 metacyclic promastigotes as previously described, with minor modification [47,48]. Cellular membranes were disrupted by sonicating pelleted cells suspended in a cold chlorform:methanol:water (1:2:0.8) solution, centrifuged (5000rpm, 10min, 4°C), and the top de-lipidated layer containing the majority of GIPLs and phospholipids was removed. The remaining insoluble material was quick-dried under stream of N2 and further extracted with two rounds of 9% 1-butanol extraction to release LPG molecules into the top aqueous layer. Hydrophobic interaction chromatography was performed to purify LPG molecules from the Leishmania surface coat. Briefly, LPG-containing butanol extracts were pooled and added to a 20% Octyl-Sepharose column that was pre-equilibrated with (5% proponal, 1M ammonium acetate). A desalting gradient (5%-60%) was applied to the column to elute LPG fractions utilizing the fraction collector, (BioRad Fraction Collector, Model 2128). LPG was detected by thin layer chromatography (TLC) and quantified by phenol sulfuric assay. Sample fractions were spotted on silica containing TLC plate. Glycan determinants were visualized by spraying the plate with orcinol (0.5mg/ml in 95% ethanol), dried, and sprayed with 75% sulfuric acid. All LPG containing fractions were pooled and dried in speed-vacuum at room temperature. Lyophilized LPG was resuspended in water and quantified by a colorimetric phenol-sulfuric assay [49]. Purification of LPG molecules was confirmed by a standard Stains-All protocol. Briefly, 5–10μg of LPG was boiled in 2X Loading Dye and loaded onto 10% SDS PAGE gel, running at 140V (room temperature). Gels were fixed in 25% 2-propanol and stained with stains-all solution (Fluka Analytical, Switzerland) containing 10% formamide, followed by destaining with 40% ethanol. Bands were visualized under white light, based on the observation that LPG molecules give rise to a blue colored complex (wavelength– 649nm) [50]. WIC79.3 western blot analysis was utilized to confirm LPG purification. Lyophilized purified LPG was resuspended in serum free RPMI and a titration of LPG (0.5μg, 1μg, and 10μg) was used for the hDC infection assay.
Relative levels of human gene transcripts were determined by qRT-PCR. Total RNA from uninfected or Leishmania-infected DCs was isolated using an RNeasy kit (Qiagen, Valencia, CA) and 1μg of RNA per infection sample was used to generate cDNA using SuperScript III Synthesis (Invitrogen, Carlsbad, CA) according to manufacturer’s instructions. For analysis of IL12B, IL12A, IRF1, IRF8, TNF, IL10, IL1B, SOCS3, TNFAIP3, and HPRT (hypoxanthine-guanine phosphoribosyltransferase) mRNA expressions, qRT-PCRs were conducted utilizing SYBR Green PCR Master Mix (Applied Biosystems by Life Technologies, Carlsbad, CA) according to manufacturer’s protocol and detected with an ABI 7900HT Fast Real-Time PCR System (Applied Biosystems by Life Technologies, Carlsbad, CA). All human primer sequences were designed by Integrated Design Tools (IDT) and used at a concentration of 5μM per reaction (S2 Table). For select analysis of IL12B, IL12A, and GAPDH (glyceraldehydes 3-phosphoate dehydrogenase) mRNA expressions, PCR reactions were setup employing TaqMan pre-developed assay kits (Life Technologies, Foster City, CA) and determined using an ABI 7500 Real Time PCR System (Applied Biosystems, Foster City, CA). For each gene, relative numbers of mRNA copies were determined by the ΔΔCT method [42].
Total RNA was isolated 8 hours post-infection from four additional donors’ uninfected monocyte-derived DCs and DCs infected with L. major FV1 WT, FV1 lpg1−, FV1 lpg1−/+LPG1, FV1 lpg2−, and FV1 lpg2−/+LPG2 using RNeasy kits (Qiagen, Valencia, CA). RNA 6000 Nano kits (Agilent Technologies, Santa Clara, CA) were used to determine total RNA integrity on a Bioanalyzer 2100 instrument (Agilent Technologies, Santa Clara, CA). 25ng of high quality RNA was converted to double stranded cDNA using a TransPlex Complete Whole Transcriptome Amplification kit (Sigma-Aldrich, Saint Louis, MO). RNA degradation, double stranded cDNA purification, and cDNA precipitation was conducted following NimbleGen Gene Expression Array user’s guide protocols (Roche-NimbleGen, Madison, WI). A Nanodrop ND-2000 (Thermo Fisher Scientific, Waltham, MA) was used to determine total RNA and double stranded cDNA concentrations. Sample cDNAs were Cy3-labeled using NimbleGen Single Color Labeling Kit (Roche-NimbleGen, Madison, WI) per manufacturer's recommendations. Labeled cDNAs were hybridized to 12-plex NimbleGen Homo sapiens Expression Arrays (platform GPL16025), featuring 140,096 probes, representing 21,269 genes and transcripts, using Hybridization LS and Wash Buffer Kits (Roche-NimbleGen, Madison, WI) per manufacturer's recommendations. Image acquisition of arrays was performed using a NimbleGen MS 200 Microarray Scanner (Roche-NimbleGen, Madison, WI), at a 2 micron resolution. NimbleGen array image data were processed using NimbleScan version 2.5 (Roche-NimbleGen, Madison, WI) to extract intensity values for each gene. NimbleScan software automates the pre-processing of NimbleGen microarray image data, including identifying the location of each probe, extraction of intensity data from the image, background correction, and obtaining expression summary values for each gene using a probe-level summarization robust multi-array average method (RMA). Probes with intensity values greater than twice that of background were retained for downstream analysis. Log2 normalized expression ratios for each gene were calculated between infected samples and paired uninfected samples. Z-scores were calculated between infected and uninfected samples as previously described [51]. Briefly, Z-score = (log2(infected intensity value/inter-quartile mean of uninfected intensity values)Gi−average(log2(infected intensity value/ inter-quartile mean of uninfected intensity values)Gi…Gn) / standard deviation(log2(infected intensity value/ inter-quartile mean of uninfected intensity values)Gi…Gn). An absolute Z-score value of 1.96 may be inferred as significant (p<0.05) [51]. Complete array data generated in this study are accessible at the NCBI Gene Expression Omnibus database (accession GSE59766). Gene expression data of RMA normalized raw microarray probe hybridization fluorescence values, where at least one sample value was twice that of background resulted in 12,911 genes.
Genes that displayed significant differential expression from FV1 WT, FV1 lpg1−, or FV1 lpg2− samples compared to uninfected samples on NimbleGen microarrays were fed into the Short Time-series Expression Miner (STEM) program [52,53]. Briefly, log2 ratio values for each of four donors were loaded into the program as repeated data, where FV1 WT data represented a “time point 1”, FV1 lpg1− data represented “time point 2”, and FV1 lpg2− data represented “time point 3”. The datasets were clustered using the STEM clustering method with minimum correlation values of 0.6. The genes from the resultant model expression profile containing IL12B were used for downstream enrichment analysis in the Web-based Gene Set Analysis Toolkit (WEBGESTALT) [54,55] with a simple list of 233 official gene symbols as input. KEGG Pathway enrichment was conducted on that list of genes with similar expression profiles to that of IL12B using the following parameters: protein-coding EntrezGene database as a reference set and a hypergeometric test with Benjamini and Hochberg multiple test adjustments. Pathways with an adjusted p-value < 0.01 and a minimum of three genes found were considered significant. The same list of gene symbols was input to The Database for Annotation, Visualization and Integrated Discovery (DAVID) v6.7 Functional Annotation Tool [56] and transcription factor binding sites for each gene were identified using protein interaction enrichment. The annotations were cross-referenced to report the most common transcription factor binding sites found in the IL12B gene and genes with IL12B-like expression between DC samples infected with L. major FV1 LPG mutants.
All statistical tests were performed using Graph Pad Prism version 5.0 (Graph Pad Software, San Diego, CA). Statistical analysis was performed using Log2 transformed ΔΔCT values using a paired Student’s T-test. Differences were considered significant at p<0.05.
First, we confirmed that the IL12B mRNA expression in hDCs infected with L. major strains FV1 WT was greater than DCs infected with LV39c5 WT. We demonstrated that L. major FV1 induced approximately 15-fold greater amounts of IL12B than L. major LV39c5 (Fig 1A) at 8 hours post infection, the optimal time for peak IL12B mRNA expression following L. major infection [44]. These data support previous work that illustrated that the hDC IL12 response is strain-specific, and also that infection with L. major FV1 promotes a high induction of IL12 and that LV39c5 is similar to the LV39 strain tested previously [34]. We also confirmed that the increased IL12B expression observed during L. major FV1 WT infections were significantly associated with the infective metacyclic promastigote stage, whereas smaller effect was observed with the non-infective procyclic promastigote stage, and no response was elicited by amastigotes (Fig 1B). These data are consistent with prior studies indicating that IL12 induction depends on the life cycle stage of Leishmania parasites [57,58].
To determine whether the enhanced IL12B production observed following infection with L. major FV1 metacyclic promastigotes was an LPG-dependent response, we first assessed the role of purified LPG on the hDC IL12B response. Human DCs were cultured in the presence of varying amounts of purified metacyclic L. major FV1 LPG for 8 hour and then assessed for IL12B expression. At lower concentrations (0.5, 1 μg), LPG induced a slight increase over uninfected samples, while at a higher concentration (10 μg) a significant 15-fold induction of IL12B mRNA was observed (Fig 1C), indicating that LPG alone is capable of stimulating IL12B production. Albeit to a lower level than what is observed with L. major LPG, purified L. donovani LPG induced a significant increase in IL-12B expression in 2 out of 3 donors (S5 Fig). Due to variation amongst the human donors, however, this difference was not statistically significant.
To probe the role of LPG and related PGs in host cell IL12 responses in the context of a Leishmania infection, we generated parasites lacking LPG alone (FV1 lpg1−) or all PGs (FV1 lpg2−) (Table 1). As L. major strain FV1 is disomic for chromosomes 25 and 34 bearing LPG1 and LPG2 respectively, two rounds of gene targeting were required to generate null mutants (S3A and S3B Fig). PCR tests confirmed the loss of LPG1 (S1B Fig) and LPG2 (S2B Fig) ORFs in the FV1 lpg1− and FV1 lpg2− mutants, respectively. Similarly, PCR tests confirmed the generation of the planned genetic alterations for the LPG1 disruption (FV1 lpg1−) (S1C Fig) and the LPG2 replacement (FV1 lpg2−) (S2C Fig). Complemented ‘add back’ lines were generated by introducing episomal constructs expressing the LPG1 or LPG2 genes into their respective null mutants (S3A and S3B Fig, bottom), which were confirmed by PCR and drug sensitivity tests. Western blot analysis with an anti-PG anti-sera (WIC79.3) showed that LPG expression alone was lost in the FV1 lpg1− mutant (S3C Fig, lane 6) and restored in the complemented FV1 lpg1−/+LPG1 line (S3C Fig, lanes 4 and 5). Similarly, Western blot analysis with WIC79.3 verified the absence of both PPGs and LPG in the FV1 lpg2− mutant (S3C Fig, lane 2), and their restoration in the complemented FV1 lpg2−/+LPG2 line (S3C Fig, lane 3).
To explore the role of LPG on the IL12 response elicited from L. major infected hDCs, we quantified the relative amount of IL12B mRNA in hDCs after 8 hours of infection with FV1 WT, FV1 lpg1−, and FV1 lpg1−/+LPG1 parasites. Compared to FV1 WT, FV1 lpg1− infected hDCs displayed a substantial decrease in IL12 expression (3.2 fold; Fig 2A) that was restored to levels approximately twice more than WT in the complemented FV1 lpg1−/+LPG1 line, perhaps consistent with a slight elevation of LPG in this line (S3C Fig, lanes 4 and 5). Our results indicate LPG plays a key role in IL12 induction in hDCs, consistent with the stimulatory effect seen with purified LPG (Fig 1B).
Conversely, FV1 lpg2− infected hDCs, relative to FV1 WT, displayed a significant increase in IL12B expression, that returned to comparable FV1 WT levels in the complemented FV1 lpg2−/+LPG2 line (Fig 2A). This observation was unexpected as FV1 lpg2− lacks LPG as well as other PGs, including PPGs (Fig 2C, lane 2). We considered the possibility that differences in infectivity between the WT and lpg2− could contribute to this result as L. major Lv39c5 lpg1− and lpg2− mutants exhibit reduced survival in peritoneal macrophages [41,59]. While parasite survival was slightly elevated in FV1 lpg2− infections, a comparable fraction of DCs were infected (Fig 2B and 2C), indicating, the differences observed in IL12 induction are likely not related to parasite survival in hDC under the conditions tested.
Thus, our studies showed that LPG is associated with increased IL12 production when tested biochemically (purified) or genetically (FV1 lpg1−), while paradoxically lpg2- which also lacks LPG showed increased production. These data invoke the possibility LPG2-dependent molecules, such as phophoglycans including PPGs or other metabolites [60] may play a suppressive role on IL12 production. Alternatively, the loss of all LPG2-dependent structures may reveal another PAMP on the parasite surface that is able to induce IL12. Either scenario indicates a complex balance and interplay between parasite glycoconjugates and host cells.
A kinetic analysis of these phenomena was conducted in DCs across four time points: 2, 4, 8, and 24 hours post-infection with FV1 WT and knockout mutants (Fig 3A). By 2 hours post-infection, FV1 lpg2− mutant infected hDCs induced slightly more IL12B compared with FV1 WT infected DCs. Albeit at higher expression levels than FV1 WT, FV1 lpg2− induced a similar kinetic IL12B mRNA response that declined by 24 hrs post infection. FV1 lpg1−, on the other hand, induced little to no IL12B mRNA (Fig 3A). Similarly, FV1 lpg2− induced a quicker and more robust IL12A response compared to FV1 lpg1− and FV1 WT infections (Fig 3B). There were no differences between the WT and mutant strains for expression of the IL12 homolog IL23A (Fig 3C), suggesting that LPG and PGs regulate IL-12 production rather than IL-23.
In addition to IL12, DCs are strong producers of other Th1 proinflammatory cytokines. TNF, for example, is significantly up-regulated in L. major infected hDCs [61]. We determined the relative fold induction of TNF in hDCs following infection with FV1 lpg1− and FV1 lpg2− mutants. We demonstrated that FV1 lpg1− induces significantly less TNF mRNA compared to WT or FV1 lpg1−/+LPG1 add back infections (Fig 4A), similar to the pattern of IL12B expression (Fig 2A). Infection with FV1 lpg2−, however, was not statistically different compared to WT infection. The effect LPG has on both IL12 and TNF may contribute to the overall skewing of L. major towards a predominant Th1 response during cutaneous leishmaniasis.
IL10 is generally implicated as a powerful inhibitor of IL12 production [62], and neutralizing IL10 promotes the ability of L. major parasites to establish IL12 production [63]. Here we quantified the IL10 mRNA levels in hDCs infected with our mutant parasites to determine whether the failure of FV1 lpg1− to elicit sustained host IL12 induction relative to FV1 WT is due to the over-expression of IL10. The IL10 expression elicited from hDCs infected with FV1 lpg1− or FV1 lpg2− did not differ from WT induced expression levels (Fig 4B), suggesting the mechanism by which these mutant parasites modulate IL12B expression is not dependent upon IL10.
To further assess the influence of LPG and PPGs on host immunological responses, we infected additional DCs with L. major FV1 WT, mutants, and complemented strains, collecting mRNA at 8 hours post-infection. cDNA generated from these samples was hybridized to NimbleGen Homo sapiens Expression Microarrays. Expression of ten genes (IL12B, IL1B, IL8, TLR4, TLR2, FKBP4, SOCS3, SMOX, FCGR1A, and TNFAIP3) correlated significantly using qRT-PCR (p<0.000001, Spearman correlation coefficient = 0.784), validating the array values (S4 Fig). Gene transcript expression values were transformed to Z-scores and those genes that were significantly differentially expressed compared to uninfected cells (Z-score ≥ 1.96) were retained for downstream analysis. Hierarchical clustering of 730 genes that were expressed differently than FV1 WT infections in at least one mutant infection revealed that the complemented strains clustered more closely to the WT strains than their respective mutant strains (Fig 5A). Compared to uninfected cells, similar numbers of genes were regulated by infection with FV1 WT (771), FV1 lpg1−(717) and FV1 lpg2− (740) (Fig 5B). Infection with FV1 WT resulted in more genes being up-regulated than either mutant strain (FV1 WT—524; FV1 lpg1-—444; and FV1 lpg2−—449). Notably, the magnitude of regulation (either up or down) was less during infection with FV1 lpg1− compared to either FV1 WT or lpg2− (Fig 5A and 5C), suggesting that this strain enters hDC in a silent fashion.
To assess the pathways involved in the regulation of IL12 by LPG, we utilized STEM and identified 233 genes that exhibited expression patterns similar to IL12B in response to infection with FV1 WT, FV1 lpg1−, and FV1 lpg2−. Overall lpg2− resembled WT while lpg1−differed (Fig 6). Pathway enrichment revealed 22 significantly enriched pathways, mostly belonging to the immune response or infectious disease categories (Table 2). The most striking observation was the enrichment of three pathways: Cytokine-Cytokine Receptor Interactions, JAK-STAT Signaling and Toll-like Signaling, in which all the genes were down-regulated by infection with FV1 lpg1− compared to FV1 WT and FV1 lpg2− (Fig 6). Although the lpg2- pathway genes did not reflect any significance in this initial analysis compared to WT, future analysis of enriched pathways by criteria other than IL12 expression could reveal significant pathways enriched by lpg2- infection.
The most common transcription factor binding sites present in the promoters of genes regulated similarly to IL12B were identified using the DAVID functional annotation tool [56]. Not surprisingly, binding sites for transcription factor families known to regulate IL12B were identified, including, Octomer-binding transcription factor (OCT), Nuclear Factor Kappa B (NFκB), Interferon Regulatory Factor (IRF), cAMP Response Element Binding protein (CREB), and CCAAT/Enhancer Binding Protein families [64–68] (Table 3).
Production of IL12 relies on the nuclear translocation and cooperative binding of IRF-1 and IRF8 to IFNG-activated sequences (GAS) found within the IL12B promoter [18]. We previously demonstrated that L. major infection of hDC results in the early activation of NFκB transcription factors resulting in the transcriptional induction and nuclear translocation of IRF-1 and IRF-8 and, ultimately, IL12 production [42]. To delineate the effect of FV1 lpg1− and/or FV1 lpg2− on the upstream transcriptional features that regulate IL12B expression, we assessed IRF1 expression in hDCs and observed that infection with FV1 mutants up-regulated IRF1, but not significantly more compared to WT induced levels (Fig 7A). This result suggests that the different IL12B responses displayed during FV1 lpg1− and FV1 lpg2− DC infections are not influenced by IRF1 expression. IRF8 mRNA levels, however, were regulated by LPG. Infection with FV1 lpg1− resulted in a reduction of IRF8 that is restored following infection with the FV1 lpg1− add back strain (Fig 7B). Infection with FV1 lpg2− did not significantly affect IRF8 expression.
The major focus of this study was to investigate whether the enhanced IL12 immune response observed in L. major FV1 WT infected hDCs is dependent upon parasite LPG; as previous studies have implicated LPG plays a major role in modulating immune function in murine cells [31,69,70], as well as in human mononuclear cells [71–73]. First, we showed that, for this strain, infection with metacyclic promastigotes induces a high IL12B response (Fig 1B), compared to procyclic promastigotes and amastigotes, consistent with prior studies [57,58]. Additionally, we demonstrated that purified LPG stimulates an IL12B response in hDCs (Fig 1C). Similar studies utilizing purified L. major LPG from another strain have also highlighted the stimulatory effect LPG has on IL12 in human PBMCs [72].
To assess the role of surface molecules in situ, we employed genetic strategies to generate parasite mutants devoid of LPG (FV1 lpg1−) or PG molecules and other LPG2-dependent metabolites (FV1 lpg2−) in the L. major strain FV1 background (S3A Fig). Previous studies on the ‘low hDC IL12, L. major strain LV39c5 mutant parasites established several roles for LPG and PGs in regulating immune function [31–33,41,60]. For example, LV39c5 lpg2− induces IL12 in mouse BMDCs co-stimulated with anti-CD40 or IFNG [32,33]. In the absence of co-stimulation, however, there was no significant difference between IL12 elicited from LV39c5 WT or LV39c5 lpg2− parasites. We observed a similar result in our hDC assay where there was little difference in IL-12 induction between LV39c5 WT, LV39c5 lpg2−, and LV39c5 lpg2−/+LPG2 infections (S6 Fig). Compared to FV1 WT, LV39c5 WT does not induce the same robust levels of IL12B (Fig 1A, S6 Fig).
Here, we generated LPG1 and LPG2 knockout mutants in the ‘high hDC IL12’ L. major FV1 background strain, in order to directly assess the parasite-derived molecular factors that contribute to the robust hDC IL12 response elicited by this strain of L. major. Our data demonstrated that the FV1 lpg1− mutant does not induce a high amount of IL12B transcript in hDCs as compared to FV1 WT (Figs 2A and 3A). Consistent with this observation, we showed that application of purified LPG was able to induce significant IL12 expression (S5 Fig), with both metacyclic L. major LPG which bears abundant PG side chain modifications, and L. donovani LPG, which is unmodified.
In contrast, and somewhat surprisingly given its similar lack of LPG, FV1 lpg2− up-regulates the IL12B response (Figs 2A and 3A) relative to FV1 WT. While in macrophage and animal infections the lpg1- and lpg2- mutants are typically attenuated [41,59], in our studies the survival of the WT and two mutant parasites did not differ significantly in DC survival over the course of these studies (Fig 2B and 2C). One explanation for this finding is that in L.major strain FV1, LPG and other LPG2-dependent glycoconjugates play inverse roles in stimulating the IL12 response in human DCs. One candidate for such an inhibitory LPG2-dependent molecule are the proteophosphoglycans (PPGs), which remain intact in the lpg1− mutant. Compared to LPG, little is known about the function of PPGs on host cell immune response, with evidence supporting roles as both an inhibitor or enhancer depending on the species and study [74–77]. PPGs vary structurally across species both in their PG and protein composition, and their large size and tendency to form polymeric aggregates renders their study more challenging [78]. Clearly, the development of mutants lacking only PPGs would be beneficial for future studies to directly assess the role these molecules have on the host cell response. Interestingly, amastigotes do not express significant amounts of the ‘pro-IL12’ LPG but do express high levels of PPG, which may further contribute to their inability to stimulate IL12 expression in hDCs. Importantly, the LPG2-dependent effect was also observed in the ‘low hDC IL12’ LV39 line, where ablation of LPG2 similarly resulted in increased IL12 production (S6 Fig)
Thus our data cause us to infer the presence of other LPG2-dependent PAMPs beyond LPG, with PPG as a possible candidate, and acting in an inhibitory fashion. The potential dominance of these inhibitory LPG2-dependent PAMPs provides an explanation for the conundrum that while all Leishmania species express LPG, despite that many do not induce IL12 [79]. Potentially, the strength of these suppressive LPG2-dependent PAMPs/processes may vary in different species and/or strains.
As it has been established that IRF1 and IRF8 are up-regulated in L. major infected hDCs and positively regulate IL12B gene expression [42], we assessed whether FV1 lpg1− or FV1 lpg2− affected the expression of IRF1 and IRF8. Interestingly, FV1 lpg1− parasites caused a significant decrease in IRF8 expression compared to WT (Fig 5A), indicating that LPG may influence the induction of IL12B by targeting upstream IL12B associated transcription factors that mediate its expression. Although IRF1 and IRF8 are known to cooperatively regulate IL12B gene transcription [42,80], we report that the FV1 lpg1− mutant does not affect IRF1 expression compared to WT at 8 hours post-infection (Fig 4B). The distinct expression phenotypes exhibited by IRF1 and IRF8 following infection with FV1 lpg1− may be due to the difference in regulation of these two transcription factors. IRF1 is ubiquitously expressed, whereas IRF8 is preferentially expressed in immune cells and in response to activating signals. Furthermore, IRF1 and IRF8 can be differentially expressed in hDCs [81]. To bind target DNA sequences, IRF8 must bind to another transcription factor, compared to other IRF family members that can bind DNA sequences alone [82]. It is possible that infection with FV1 lpg1− reduces the amount of IRF8, which in turn inhibits the capacity of other transcription factors, such as IRF1, to form heterodimeric complexes that bind the IL12B promoter. These data suggest that LPG and not other PGs, enhance the IL12B response by a common mechanism involving IRF8.
Like IL12, L. major induces TNF in both human macrophages and DCs [61]. We therefore evaluated the relationship between parasite derived PG-bearing molecules on TNF using our LPG and PG null mutants. Our results demonstrate that the lpg1− mutant exhibits a significant decrease of TNF expression, similar to the reduction observed for IL12B (Fig 5A). Interestingly, the promoter regions for IL12B and TNF have similar transcription factor binding sequences, namely NFκB and ETS sites; the latter containing ISRE sequences that promote gene transcription upon IRF8 complex binding [83]. Therefore, it is possible that the reduction in TNF expression observed during FV1 lpg1− infection (Fig 5A) may also be IRF8-specific. A murine study demonstrated that cholera toxin (CT) inhibits plasmacytoid dendritic cellular IL12 by blocking the ability of IRF8 to bind to the ISRE sequence within the IL12B promoter, while IRF1 phosphorylation and subsequent binding to its DNA target sequence remained unaffected [84]. It is feasible that a similar mechanism exists in L. major infected cells, whereby IRF8 is specifically targeted for induction downstream of parasite LPG binding, subsequently leading to the induction of IL12B and TNF. Altogether, our data indicates that L. major FV1 skews the hDC response in an LPG-dependent manner towards a Th1-like polarization characterized by an increase in IL12 and TNF production which may be regulated by a common mechanism involving IRF8. A recent study demonstrated that macrophage induction of IL12B is controlled at the level of IRF8, which is specifically targeted for activation downstream of TLR4 in concert with Notch signaling pathways [85]. Interestingly, TLR4 [86] and other TLRs [87–91] have been implicated in recognition of parasite LPG.
An alternative explanation for the lack of an IL12 signal observed in the FV1 lpg1− infections may be a consequence of other functionally active PG-containing molecules, such as the PPGs which remain intact in the lpg1− mutant. These PPGs could provide an inhibitory IL12 signal. This theory is supported by our results demonstrating that FV1 lpg2−, which lacks both LPG and PPGs, induces higher levels of IL12 compared to WT (Fig 2A), suggesting that some PG-containing molecules actually inhibit IL12 responses. In addition, amastigotes, on which LPG expression is drastically down-regulated and high levels of other PG containing glycoconjugates are highly expressed [15], do not induce IL12 (Fig 1B). Compared to LPG, little is known about the function of PPGs on host cell immune response. Previous work illustrating the ability of PPGs to induce complement activation by triggering the mannose binding protein pathway [76] and their inability to elicit CD4+ T-cell response in murine bone marrow derived macrophages [74], concludes that PPGs may contribute to the chronic infections observed during L. mexicana infections. However, it has been demonstrated that L. major PPGs require IFNG priming to induce TNF and NO production in murine macrophages [77]. In human PBMCs, PPGs cause an induction of IL10 and to a lesser extent NO and IL12 [75]. Although these studies provide conflicting implications for PPGs role as either inhibitor or enhancer of immune response, it is difficult to compare studies because the repertoire of PPGs structure varies across species [78]. Additionally, the use of purified PPGs can be problematic because the amount of purified PPGs added is often higher than what is biologically present during an actual infection, therefore the development of mutants lacking only PPGs would be beneficial for future studies to directly assess the role these molecules have on the host cell response. We measured IL-10 mRNA levels in our mutant-infected DCs, because of the generally inhibitory effects of IL-10 on IL12 [62]. However, IL10 expression exhibited between FV1 lpg1−, FV1 lpg2−, and WT infected hDCs did not differ (Fig 5B), ruling out one theory that the decrease in IL12B expression observed during FV1 lpg1− could be consequence IL10 overproduction. Another explanation for the induction of IL12 by FV1 lpg2−, is the possibility that the absence of all surface and secreted PGs reveals a molecular pattern or some other molecule that induces IL12.
Our microarray analyses of FV1WT, FV1 lpg1−, and FV1 lpg2− infected hDCs revealed that FV1 lpg1− enter hDC in a relatively silent fashion as indicated by the overall down-regulation of significantly expressed transcripts, (Fig 6), and the overall reduction in genes belonging to cytokine and TLR related gene pathways, (Fig 7). Altogether these data suggest that a lack of LPG molecules results in silent entry and that LPG is a major pattern recognized by pattern recognition receptors on DCs. As with the IL12 response, the absence of all PGs appears to either release some sort of repression or reveals a molecular pattern that compensates for the lack of LPG, highlighting the complexity of DC pattern recognition receptor interactions in controlling host responses to Leishmania infection. Future analyses focusing on FV1 lpg2- mutant infections may reveal pathways uniquely regulated by PGs.
This work adds to the growing set of genetically modified parasites (lpg1−, lpg2− in the L. major FV1 background) providing biologically relevant tools for assessing the role of parasite surface glycoconjugates on cellular function in human and mouse model systems, as well as, provides insight into the complex interplay of LPG and other PG molecules on the cellular immune response elicited following L. major infections by global gene expression analyses.
|
10.1371/journal.pcbi.1004861 | Sensory Agreement Guides Kinetic Energy Optimization of Arm Movements during Object Manipulation | The laws of physics establish the energetic efficiency of our movements. In some cases, like locomotion, the mechanics of the body dominate in determining the energetically optimal course of action. In other tasks, such as manipulation, energetic costs depend critically upon the variable properties of objects in the environment. Can the brain identify and follow energy-optimal motions when these motions require moving along unfamiliar trajectories? What feedback information is required for such optimal behavior to occur? To answer these questions, we asked participants to move their dominant hand between different positions while holding a virtual mechanical system with complex dynamics (a planar double pendulum). In this task, trajectories of minimum kinetic energy were along curvilinear paths. Our findings demonstrate that participants were capable of finding the energy-optimal paths, but only when provided with veridical visual and haptic information pertaining to the object, lacking which the trajectories were executed along rectilinear paths.
| Recent studies have shown that when learning novel dynamics in the context of reaching movements, people often ignore energetic optimality in favor of Euclidean geometric optimality, preferring rectilinear paths over mechanically optimal trajectories. Although an explanation could be that sensory-motor coordination ignores energetic cost, another possibility is that different sensing modalities need to be in agreement before the brain will optimize energetic cost during motion. We provide evidence for this latter perspective, by showing that when provided congruency and consistency of visual and haptic feedback, participants take into account both geometric and mechanical properties of a manipulation task. However, when visual and haptic feedback are inconsistent, participants revert to the rectilinear paths seen in previous studies. We conclude from these observations, that when transporting an external object, sensory agreement between vision and touch guides the optimization of the kinetic energy exchanged during movement between the arm and the object.
| One of the most established findings in planar multi-joint reaching movements is that hand trajectories tend to be executed along straight paths with a bell-shaped velocity profile [1–3]. Given that there are theoretically infinite paths that the hand could take for reaching from one point to another, the presence of this consistent feature in reaching movements has been used to suggest that the nervous system chooses this solution because it is “optimal” in some way. Mathematical optimization has been considered as an appealing principle to explain observed biological movements. Optimization requires an objective function, or cost, that includes the quantities being minimized. The choice of cost has received much attention in the study of neural information processing, in particular, by the motor system [4–8]. The components of an objective function generally fall into two main types: kinematic and dynamic. While the former relates only to the geometry of motion, the latter relates to the forces that cause the motion. Despite fundamental differences between the two types, objective functions consisting purely of one or the other have been similarly successful in predicting data obtained from unperturbed planar reaching movements. Adaptation studies have attempted to distinguish between kinematic and dynamic costs by introducing perturbations to these movements. It has been shown [9, 10] that in the presence of kinematic perturbations, participants chose to move the hand along curved paths so as to produce visually straight trajectories. Similarly, under the dynamic perturbation caused by forces depending upon the velocity of the hand, subjects learned to recover straight hand trajectories through repeated practice of reaching movements [11]. More recently, to evaluate if mechanical energy costs play a role in motor learning, a custom force field was designed in a way that the path of minimum mechanical energy was substantially different from the straight path [12]. Under this situation, participants returned to straight line reaches even after they experienced moving along the energy optimal path. These studies suggest that the tendency to move the hand on a straight line in planar movements is strong and persistent, arising under a variety of dynamic and kinematic perturbations, and indicate that the kinematic costs are either necessary [13] or sufficient [14] components of the cost function.
However, these previous studies were typically focused on unconstrained movements of the hand in free space and when a force field disturbed these movements. In the majority of earlier studies a cursor was used as a visual image of the system under control [11, 12, 15]. This representation makes all spatial directions visually equivalent and moving the cursor on a straight line appears to be an economical approach. Moreover, the haptic feedback (i.e. the contact forces experienced during movements) were predominantly in the form of force fields. In these experiments, there were no features in the visual scene or in the shape of the cursor that could be associated with the forces experienced by the subjects. We hypothesized that this dissociation between sensory modalities elicits a compensatory strategy where movements are channeled to restore the kinematics of the unperturbed hand motion. Conversely, congruence between feedback modalities, representing the action upon an identifiable external object is expected to result in energy efficient strategies. In this case, motor learning leads to a progressive optimization of the energetic costs of movements rather than a process towards recovering a straight invariant trajectory. Depending on the object's dynamics, the resulting trajectory may systematically and substantially deviate from the straight line. Therefore to test this prediction, we used an object manipulation task where there is a well-defined relation between the visual and haptic feedback.
Participants executed goal-directed reaching movements in the horizontal plane while holding the end point of a virtual planar double pendulum in the absence of gravitational effects. The choice of the double pendulum was motivated by the fact that the energy-optimal trajectories for moving this object were along curved paths, allowing us to tease apart the relative importance of kinematic and dynamic costs.
The energy-optimal paths for moving this system were calculated as follows. The total energy for this system consists only of the kinetic term. The path of least kinetic energy between any two double pendulum configurations is a solution to a two-point boundary value problem—leaving the initial and final velocities as free variables- of the unforced system and it is generally curved in shape (Fig 1). However, this path is a purely geometric quantity and from a control perspective, it is not an admissible solution because the velocity requirements are not satisfied. To find an admissible solution we used optimal control theory with the only running cost of effort, defined as the force being applied to the object. Expectedly, the path of minimum energy and the effort optimal trajectory had similar shapes (see S1 File). The energy (mechanical work) that is acquired by the object upon point-to-point maneuvers was calculated by
E=∫0T|Fv|dt
(1)
where F represents the force applied to the object, v is the velocity of the hand and T is the movement time. The energy that was required to move along the straight path and the path of least energy between each pairs of targets is included in Fig 1. These values are obtained from a minimum jerk trajectory with the movement duration of 1 sec.
Participants were randomly divided in three groups. All received both visual and haptic feedback. Participants in Group 1 (n = 8) received veridical visual and haptic information of the double pendulum. For Group 2 (n = 8), only the visual feedback was manipulated. Participants in this group were presented with a circular cursor representing the moving extremity of the pendulum. They did not see the linkage but the haptic feedback corresponded to the entire mechanism. Participants in Group 3 (n = 8) could see the entire linkage while the haptic information was manipulated so as to emulate the isotropic inertia of a point mass. In this scenario, our hypothesis made an explicit prediction on the trajectory formation: If participants were provided with full vision of the manipulated object together with a congruent haptic feedback (Group 1), they would integrate the geometric structure with interaction forces to converge to the curvilinear paths of minimum energy. In contrast, if participants only received visual feedback of the endpoint (Group 2), or haptic feedback corresponding to point-mass dynamics (Group 3), then movements would be executed along rectilinear paths because of the lack of consonance between the sensing modalities.
For each trial, movement initiation and termination were identified using 10% of peak velocity threshold. Participants in all the three groups started the experiment by moving along straight line trajectories. However, the trajectory divergence between Group 1 and the remaining groups started after the very first few trials. We found that with practice, all participants that received congruent visual and haptic feedback progressively moved towards producing curved trajectories that were similar to the path of minimum energy. This gradual adjustment suggests that the problem of finding the energetically optimal trajectory was solved via gradient descent beginning from the straight line trajectory typical of the freely moving hand, In contrast, all participants that were presented with incongruent feedback continued to move along rectilinear paths (Fig 2). We quantified the similarity of executed trajectories to both straight line and least energy paths using discrete Fréchet distance (DFD) [16]. The Fréchet distance between two curves is the minimum cord-length that is sufficient to join two points traversing each curve with arbitrary speeds without backtracking. Intuitively, imagine a dog walking along one curve and the dog’s owner walking along the other curve and they are connected by a leash. Both can change the speed and even stop at arbitrary positions with arbitrary durations but neither are allowed to move backwards. The Fréchet distance between the two curves is the length of the shortest leash that connects the man to the dog at all time. One-way ANOVAs on DFD from the straight path and DFD from the least energy path during the last block revealed a significant group effect on both distances. Dunnett’s post-hoc tests showed that Group 1 was significantly further from the straight path than the two other groups (p <0.01). Similarly Group 1 was significantly closer to the path of minimum energy compared to Group 2 (p <0.01).
One feature in the result is that although the participants in Group 1 show greater curvature, they did not completely converge to the energy efficient path. We speculate that this may be due to the fact that participants in Group 1 did not have any explicit knowledge about the mechanical properties of the object and the geometric shapes of the effort optimal trajectories. They derived these trajectories solely based on the sensory information. Therefore, considering noise and model uncertainties in sensory transduction and neural computation, they were expected to move at larger distances from the paths of minimum energy and exhibit greater variability in their movements in comparison with participants in the remaining groups who moved along straight paths and had explicit kinematic plan for executing their movements.
Our results suggest that when learning novel dynamics, if the visual representation is a cursor or a shape that is indicative of isotropic dynamical structure, this impoverished representation provides a strong bias towards Euclidean representations of the configuration space, where all directions are equivalent and straight lines are the natural geometrical paths for joining two points. In this situation, participants experience a mismatch between expected and sensed forces under the assumption that they are moving the arm in free space. This mismatch between sensory information triggers a compensatory strategy where subjects fight the force field to recover the straight unperturbed trajectory. However, if subjects are provided with any visual information suggestive of an external object being manipulated, with non-Euclidean dynamical structure to begin with (because of the non-isotropic, position dependent inertia tensor at the contact point), then they attribute the haptic feedback to the visual image and with practice try to develop a representation of the object’s configuration space. This harmony between sensing modalities promotes a control policy that requires less effort to perform the task. Work on remapping finger movements has highlighted that when subjects learn a novel task of manipulating a kinematic chain by continues finger motions, movement trajectories are formed along the geodesics (i.e., paths of minimum length) corresponding to the geometrical structure of that object [17]. Subjects in Group 1 and 3 were both provided with the same visual feedback but at the end of the experiment they moved along different paths, each corresponding to the path of minimum energy of the object that they were manipulating (double pendulum haptics vs point mass haptics). This result confirms that curved trajectories observed in Group 1 is not a solution to the kinematic problem but it is a progressive optimization of the energy exchanged with the object. Here, effort was defined as the force that subjects applied to the system to move it between target positions. Although we did not measure the metabolic effort (i.e. the physiological energy cost), a recent study found that indeed the metabolic effort is reduced during force field adaptation [18].
The conclusions of previous studies on energy optimization in human motor control are mixed. Some studies suggest that the motor system is capable of minimizing the energetic costs of free limb movements both in arm reaching movements [19] as well as locomotion [20], while other studies of learning novel dynamics suggest that the motor system does not take into account the energy when executing movements [12, 14]. When moving a limb or manipulating an object the energy optimal solution depends on the mechanical properties rather than the visual representation. However, there is considerable amount of evidence that the movement control policy and consequently trajectory formation both in free reaching movements and in object manipulation depends remarkably on the visual feedback.
Straight line trajectories are found typically in studies of free arm movements when the sight of the arm is obstructed and the subjects are presented with a cursor. However, it has been shown that trajectories of the free reaching movements of congenitally blind and even blind folded individuals to haptic targets are more curved than movements made by subjects to the same target positions under visual guidance [21, 22]. Similarly, subjects performed curved motions when they were instructed to reach to physical targets with their arm rather than reaching with a cursor to a virtual target [19]. These studies suggest that in free limb movements, humans can flexibly alter movement behavior between geometric and energetic optimally depending on the feedback. In contrast to free movements, moving in a force field provides an additional challenge to the nervous system because in this case, the effort optimal trajectory not only depends on the mechanics of the body, but also on the dynamical properties of the field. It has been demonstrated that the representation of the dynamics of a manipulated object also depends on the visual representation [23] and that only specific and meaningful visual cues can promote proficient switching between different mechanical tasks [24]. Recent studies on learning novel dynamics reported that the motor system ignores energetic costs in favor of geometric optimality. However subjects in these studies were exposed to a force field with the representation of a cursor [12] or with no visual feedback [13]. In the latter study the visual feedback (cursor) was provided only at the beginning and at the end of each trial. Here, we showed that in object manipulation, the visual motion of the object resulting from an applied force is a critical piece of information for the brain to represent the dynamics. Given our finding that subjects learned to minimize the energy transfer with an object having anisotropic position-dependent inertial properties, we observe that the most common objects being transported by our hands have isotropic position-independent translational inertias. They are therefore characterized by straight-line kinetic energy geodesics when moving on the horizontal plane. Thus, we speculate that the tendency to perform straight-line planar movements of the hand may be a baseline behavior emerging from the experience of transporting such objects while optimizing the energy exchanged with them. It has been shown [19] that the arm trajectory in a 3D pointing movement is along the geodesic path that is obtained through the minimization of the kinetic energy on the configuration space of the arm. The Euler- Lagrange equations are equivalent to the equations of geodesic motion on a Riemannian manifold. We extended the computational model in [19] to learning novel dynamics, we showed that trajectory formation hinges on the consistency between feedbacks representing the system under control, and how feedback variations can lead to remarkably different behaviors. The demonstrated results provide insights into studies on adaptation, effort minimization and object manipulation by the human motor system.
Twenty four right-handed volunteers (12 female) participated in the experiment. All participants were neurologically intact and had no prior knowledge of the experimental procedure. The study protocol was approved by Northwestern University’s Institutional Review Board and all the participants signed an informed consent form. Participants were positioned in front of a horizontal mirror and held the handle of a planar, two degree of freedom robotic manipulandum with their right hand. The mirror prevented the participant’s view of their hand and the robot. A LED monitor was positioned above the mirror with the same vertical distance as the distance between the robot and the mirror. This setup caused the visual information to appear at the same height as the hand. The display was calibrated so that the visual feedback of the hand was overlaid on its true position.
Participants performed goal-directed reaching movements to three targets (diameter = 3 cm). Targets were presented in a block structure, with randomized order within each block. After reaching to each target, participants maintained the position for 500 ms before the next target appeared. The experiment consisted of 10 blocks and in each block participants performed 48 reaching movement (16 reaches per target). Participants could rest between blocks. During all these reaching movements, the manipulandum was either connected to the endpoint of a virtual double pendulum with the mechanical properties that are listed in Table 1 or a virtual 15 kg point mass, by means of a virtual spring-damper (K = 2200 N/m, B = 65 N.s/m). Position and velocity of the manipulandum handle were computed from instrumented encoders at the frequency of 1 kHz to provide haptic feedback of the forces resulting from moving the double pendulum or the point mass. The manipulandum was equipped with electric motors with the peak torque of 82 Nm. However, the maximum torque that the robot was asked to generate in the fastest recorded trial in this experiment was about 15 Nm. Data were recorded at the rate of 100 Hz. Participants were randomly divided into three groups of equal size (n = 8 per group): Group 1, where participants received both the visual and haptic feedback of the double pendulum. Group 2, where participants received the visual feedback of the moving extremity of the pendulum that was held in their hand, in form of a circle (diameter = 1.5 cm) and the haptic feedback of the double pendulum. Group 3, where participants received the visual feedback of the double pendulum with haptic feedback of the point mass. One of the participants in Group 3 revealed that he was familiar with the purpose of the study and was replaced by another participant.
|
10.1371/journal.ppat.1007737 | A streptococcal Fic domain-containing protein disrupts blood-brain barrier integrity by activating moesin in endothelial cells | Streptococcus equi subsp. zooepidemicus (SEZ) is a zoonotic pathogen capable of causing meningitis in humans. The mechanisms that enable pathogens to traverse the blood-brain barrier (BBB) are incompletely understood. Here, we investigated the role of a newly identified Fic domain-containing protein, BifA, in SEZ virulence. BifA was required for SEZ to cross the BBB and to cause meningitis in mice. BifA also enhanced SEZ translocation across human Brain Microvascular Endothelial Cell (hBMEC) monolayers. Purified BifA or its Fic domain-containing C-terminus alone were able to enter into hBMECs, leading to disruption of monolayer barrier integrity. A SILAC-based proteomic screen revealed that BifA binds moesin. BifA’s Fic domain was required for its binding to this regulator of host cell cytoskeletal processes. BifA treatment of hBMECs led to moesin phosphorylation and downstream RhoA activation. Inhibition of moesin activation or moesin depletion in hBMEC monolayers abrogated BifA-mediated increases in barrier permeability and SEZ’s capacity to translocate across monolayers. Thus, BifA activation of moesin appears to constitute a key mechanism by which SEZ disrupts endothelial monolayer integrity to penetrate the BBB.
| Streptococcus equi subsp. zooepidemicus (SEZ) is an important animal pathogen and can cause meningitis in humans. Little is known about how this Group C streptococcal species penetrates the blood-brain barrier (BBB). We identified bifA, a gene that is critical for SEZ to cause meningitis in mice and to penetrate a human brain endothelial monolayer in a tissue culture model. BifA’s Fic domain enables the protein to enter into endothelial monolayers and to bind to moesin, a cytoskeletal regulatory protein, leading to its activation. Preventing moesin activation abolished BifA-induced barrier leakiness and SEZ’s capacity to penetrate a monolayer barrier. Together, our findings suggest that SEZ meningitis depends on BifA, a Fic-domain protein that manipulates moesin-dependent signaling to modulate BBB permeability.
| Streptococcus equi subsp. zooepidemicus (SEZ) is a Lancefield Group C opportunistic pathogen capable of infecting a broad range of animal species, including humans [1]. The most significant burden of disease caused by SEZ is in farmed animals, including horses, cows and pigs [2]. However, human SEZ infections have been reported globally and are often linked to consumption of unpasteurized milk or contact with infected animals. Meningitis is the most common clinical manifestation of human infection with SEZ and can be fatal [3, 4].
SEZ, like most streptococci, is an extracellular pathogen [2] and to cause meningitis, these organisms must penetrate the blood-brain barrier (BBB), a functional barrier established in part by the endothelial cells lining the brain microvasculature. This highly selective barrier between the brain and the circulatory system acts as an important protective mechanism, excluding blood-borne pathogens and toxins from the central nervous system [5]. While relatively high pathogen concentrations in blood are thought to be a prerequisite for organisms to traverse the BBB, different pathogens appear to rely on varied mechanisms to penetrate this barrier [5]. Diverse factors facilitating pathogen adhesion to brain capillary endothelial cells have been identified and both transcellular and paracellular routes for pathogens to cross the BBB have been reported [6, 7].
Although SEZ virulence factors that facilitate pathogen adhesion to host tissue and immune evasion have been identified [8–10], there is little knowledge of the factors and mechanisms that enable SEZ to penetrate the BBB. In previous research, we sequenced and compared the genome sequence of a virulent SEZ strain (ATCC35246, isolated from a dead pig) to those of non-virulent SEZ strains, to identify potential virulence-linked genes [11]. Several loci in the ATCC35246 isolate appeared to have been acquired through horizontal gene transfer. One such region (pathogenicity island II) contained a gene (SeseC_01334) that is predicted to encode a protein carrying an N-terminal RhuM domain and a C-terminal Fic domain. These two domains are linked to virulence in other pathogens. Fic (filamentation induced by cyclic AMP) domain-containing proteins are present in many animal and plant pathogens [12]. Often these proteins are delivered via type III or type IV secretion systems (T3SS, T4SS) directly into the cytosol of host cells, where they manipulate host signaling pathways via covalent modification of target proteins. Though Fic proteins induce varied modifications in their targets (e.g., AMPylation, UMPylation, phosphorylation and phosphocholination have been described), they all share a consensus 9 amino acid core, HxFx(D/E)(A/G)N(K/G)R, with the histidine residue exhibiting the greatest conservation [12]. Since Fic domain proteins are linked to pathogenicity, we investigated whether SeseC_01334 (here re-named BifA, for brain invasion factor) contributes to SEZ virulence.
We show that BifA is critical for SEZ to disrupt the BBB and to infect the mouse brain. Furthermore, this Fic-domain protein is required for SEZ to penetrate a tissue culture model of the BBB. BifA’s Fic domain enables the protein to enter into and to disrupt the barrier function of brain endothelial monolayers. BifA targets moesin and leads to its phosphorylation. Inhibition of moesin phosphorylation or knockdown of moesin expression prevented BifA-mediated increases in monolayer permeability and SEZ’s capacity to penetrate a monolayer barrier. Collectively, our findings reveal that SEZ meningitis depends on BifA, a Fic-domain protein that disrupts BBB function by manipulating moesin-dependent signaling.
We previously found that SEZ ATCC35246 contains 2 purC homologues, SeseC_00028 and SeseC_01334. The later locus was presumably acquired by horizontal transfer because its G+C content (34.86%) differs from the chromosomal G+C content (41.65%). Notably, although SeseC_01334 bears some similarity to SeseC_00028, it also features an additional C-terminal Fic domain (S1 Fig), which in several other bacteria has been linked to pathogenicity [12], and an N-terminal RhuM domain that SeseC_00028 lacks. To investigate if SeseC_01334 (here renamed bifA, for brain invasion factor A) is required for SEZ ATCC35246 virulence, we generated a bifA deletion mutant strain (ΔBif) as well as a complemented strain (CBif), in which BifA was expressed from a plasmid in the ΔBif background. Using an established murine model of SEZ infection [13], mice were inoculated via intraperitoneal (i.p.) injection with WT or ΔBif strains. There was ~100× more WT colony forming units (CFU) than ΔBif CFU recovered from the brains of infected mice (Fig 1A). In contrast, there were less marked differences in numbers of WT and ΔBif CFU recovered from the lung and kidney and in the liver and spleen, the number of ΔBif CFU recovered tended to exceed those of the WT (Fig 1A). Thus, BifA may be particularly important for SEZ colonization of the brain. Furthermore, all WT-challenged mice died by 2 days post-infection (dpi), whereas mice challenged with ΔBif survived until 5 dpi (Fig 1B). Complementation of BifA in the ΔBif mutant restored its lethality to WT levels (CBif, Fig 1B) as well its capacity to colonize the brain (Fig 1C). Despite the differences in the virulence of the WT and ΔBif strains, they had very similar in vitro growth curves (S2A Fig), suggesting that an intrinsic growth defect is not the explanation for the in vivo attenuation of the mutant. Together, these observations show that BifA promotes SEZ’s lethality and its capacity to enter into and/or proliferate in the brain.
For SEZ to colonize the brain it must traverse the BBB. We used an Evans Blue (EB) dye permeability assay [13] to assess the integrity of the BBB in mice inoculated with SEZ. EB was administered to mice 18 hours post infection (hpi) with WT, ΔBif or CBif and then the brains were dissected 2 hours later (Fig 1D). The brains of mice infected with WT SEZ had significantly greater amounts of detectable EB than the brains of mice infected with ΔBif (Fig 1D); bifA complementation partially restored the capacity of ΔBif to disrupt the BBB (Fig 1D). Thus, the marked defect of the ΔBif strain to colonize the brain may, at least in part, be explained by the reduced capacity of this strain to penetrate the BBB. Consistent with this hypothesis, we found that there was a much lower ratio of CFU recovered from the CSF vs the blood 12 hour after infection with ΔBif vs the WT strain (Fig 1E), even though there were very similar numbers of WT and ΔBif organisms recovered from blood at this point (S2B Fig). The absence of bifA appears to account for the reduced capacity of ΔBif to access the CSF, since this defect was not observed in the complemented strain (Fig 1E). Furthermore, the WT and ΔBif strains had indistinguishable capacities to proliferate in blood (S2B Fig). One consequence of BifA’s apparent capacity to promote SEZ disruption of the BBB may be the severe cerebral hemorrhage that was observed in the brains of animals infected with WT and CBif, but not in those infected with ΔBif (S3 Fig). Moreover, using a transwell assay, we found that WT and CBif had a greater capacity to traverse human brain microvascular endothelial cell (hBMEC) monolayers than ΔBif (Fig 1F). Together, these observations suggest that BifA promotes SEZ virulence and brain pathology by enabling the pathogen to transit the BBB.
Prediction of protein structure using the THHMM server suggested that BifA lacks transmembrane helical domains and is likely a hydrophilic protein. We found that BifA could be detected in SEZ culture supernatants (Fig 2A), raising the possibility that it might directly interact with host cells to modulate BBB integrity. Consistent with this idea, we found that BifA could be detected as cytoplasmic foci inside cultured hBMEC cells after exposure to supernatant derived from WT but not ΔBif SEZ (Fig 2B).
The full-length and N- and C-terminal portions of BifA (ΔFic and ΔRhuM, respectively) were purified along with a BifA mutant containing an H247A substitution in the Fic domain (Fig 2C). This mutation was shown to ablate the catalytic activity of other Fic domain containing proteins [14]. Purified full length BifA was taken up into cultured hBMEC cells where it was detected as cytoplasmic foci by immunofluorescence microscopy (Fig 2D). Notably, the concentration of BifA found in the culture supernatants (~18ug/ml, S4 Fig) used above, were similar to the final concentration of purified BifA used to detect BifA entry into hBMEC cells in Fig 2D and to modulate monolayer permeability in experiments described below. Neither ΔFic nor BifA H247A were detected inside the hBMEC cells, whereas the intracellular amount and distribution of the ΔRhuM BifA variant was similar to full length BifA. Thus, the activity of BifA’s Fic domain appears required for the protein to enter host cells, but its RhuM domain is dispensable for this function.
We also tested whether BifA was sufficient to enable latex beads coated with the protein to enter hBMEC cells. Transmission electron microscopy revealed that beads coated with full length BifA or the ΔRhuM truncated variant enabled latex bead internalization (Fig 2E). Beads coated with ΔFic or H247A BifA were not internalized into cells any more than uncoated beads. Together, these observations indicate that BifA can mediate its own entry into hBMEC cells, and that entry appears dependent on a functional Fic domain.
To test whether a functional Fic domain was important for SEZ virulence in vivo, we inoculated mice with the ΔBif strain complemented with BifA lacking the Fic domain (CBif (ΔFic)) or the H247A allele (CBif (H247A)). These strains were similarly attenuated as ΔBif in lethality (Fig 1B) and brain colonization (Fig 1C). These observations strongly suggest that BifA’s Fic domain is required for robust SEZ virulence.
Since BifA appears to promote SEZ’s capacity to transit the BBB, we tested whether BifA treatment altered the barrier integrity of hBMEC monolayers. Using penetration of EB as a gauge of barrier disruption [15], addition of either full-length BifA or the ΔRhuM truncation variant to hBMEC monolayers resulted in time-dependent increases in barrier permeability, which became apparent as early as 15 minutes after addition of BifA or ΔRhuM (Fig 3A). In contrast, addition of the H247A BifA mutant or the ΔFic truncation variant to the transwells did not alter the monolayers’ barrier function.
Live microscopy of monolayers was carried out to monitor the effects of BifA and BifA H247A treatment on monolayer integrity (Fig 3B and S1–S6 Movies). In some parts of the BifA-treated monolayers, the hBMEC membranes between adjacent cells appeared to retract by ~15–30 min after addition of the protein and frank gaps in the monolayer, which widened through time, became evident by ~120 min after treatment (Fig 3B from S1 Movie and Fig 3C from S2 Movie). In contrast, addition of H247A BifA to monolayers did not result in detectable morphologic changes in the hBMEC cells compared to the untreated monolayer (mock) over a 3 hours period of observation (Fig 3B, S3–S6 Movies). Additional studies to elucidate the molecular mechanism(s) by which BifA disrupts the integrity of hBMEC monolayers are required. However, interruption of tight junctions could contribute to the permeabilization of the monolayers, since we found that cellular levels of the tight junction protein, zona occludens-1 (ZO-1), decreased after addition of BifA (Fig 3D, S5 Fig).
We used a SILAC-based comparative ‘pull-down’ approach to identify BifA binding partners. For these studies, BifA-GFP was expressed in HEK293T cells and the proteins that precipitated along with BifA were identified by mass spectrometry (Fig 4A). One of the top hits among the 19 candidate BifA-interacting protein identified (S1 Table) was an ERM family protein, which was enriched ~1.8-fold in the BifA-GFP vs the GFP pull-down. ERM family proteins include Ezrin, Radixin, and Moesin, which function in endothelial cells as well in other cell types as critical regulators of the actin cytoskeleton [16]. These proteins are capable of binding to integral membrane proteins through their N-terminal FERM domains and filamentous actin through their C-terminal Ezrin Radixin Moesin association domain (ERMAD). By virtue of these dual binding capacities, ERM proteins regulate actin polymerization at the cell cortex, where they provide a critical link between the cell membrane and cytoskeletal components [17]. Since the dominant ERM family protein in hBMEC is moesin [16], we focused subsequent studies on BifA’s potential interaction with moesin.
To confirm that BifA interacts with moesin in hBMEC cells, co-immunoprecipitation (co-IP) experiments were performed with lysates from hBMEC expressing HA-tagged moesin and several BifA variants. The epitope-tagged moesin co-IPed with full-length BifA and the ΔRhuM truncation variant, but not with the H247A BifA or ΔFic variants (Fig 4B). Thus, BifA’s interaction with moesin in hBMEC cells appears to depend on its Fic domain. Similar co-IP experiments were carried out using cells transfected with tagged variants of moesin (Fig 4C), to determine which of the moesin domains is required for BifA interaction. BifA co-IPed with a moesin variant lacking the FERM domain but not with a variant lacking the ERMAD domain (Fig 4D), suggesting that the FERM domain is dispensable for the BifA-moesin interaction. In addition, we found that purified BifA could interact with moesin in hBMEC lysates, while BifA H247A could not (S6 Fig).
Moesin’s ERMAD domain includes a highly conserved threonine residue (T558) that is phosphorylated during activation [17]. Substitution mutants in the moesin T558 residue that are predicted to be phosphoablative (T558A) or phosphomimetic (T558D) were generated to begin to address whether T558 phosphorylation modulates BifA-moesin interaction. Interestingly, BifA precipated less moesin T558A than WT moesin or moesin T558D, which precipitated with BifA at least as well as WT moesin (Fig 4D), suggesting that BifA’s interaction with moesin is enhanced by phosphorylation of moesin T558.
Surface plasmon resonance analyses with purified BifA and moesin proteins were performed to further characterize BifA’s interaction with moesin. These studies demonstrated that BifA could bind moesin in isolation from other proteins, and thereby indicate the interaction is direct (Fig 4E). Consistent with previous results, minimal binding of the H247A or ΔFic BifA variants to moesin was detected with this assay (S7A Fig). Additionally, the binding affinity of BifA for the moesin T558A mutant was less (T558A, KD = 7.366×10−7 M) than that of T558D (KD = 1.078×10−8 M), which had an even higher affinity than the wild-type protein (Fig 4E and S7B Fig). Collectively, these observations demonstrate a direct interaction between BifA and moesin that is dependent on their respective Fic and ERMAD domains and that is likely enhanced by activation (T558 phosphorylation) of moesin.
Since several Fic domain-containing bacterial toxins are reported to lead to the phosphorylation of their respective target proteins [18], we tested whether BifA promotes moesin phosphoryation. We monitored moesin T588 phosphorylation following addition of different BifA variants to hBMEC cells by immunoblotting with an antibody that recognizes phosphorylated moesin T588 (p-Moesin). Addition of either full length BifA or the BifA ΔRhuM variant to cells led to moesin T588 phosphorylation in a time-dependent fashion but did not alter total cellular moesin levels (Fig 5A and 5C). In contrast, no changes in moesin phosphorylation or levels were detected when the H247A BifA or ΔFic variants were added to cells (Fig 5B and 5D). Similar results were obtained from immunoblots of lysates electrophoresed with Phos Binding Reagent Acrylamide, which alters the electrophoretic mobility of phosphorylated proteins (S8 Fig). These observations indicate that treatment of hBMECs with internalizable and moesin-binding variants of BifA promotes moesin phosphorylation.
ERM family proteins are phosphorylated by protein kinase C (PKC) [19]. We used NSC305787, a small molecule inhibitor of PKC phosphorylation of ERM family proteins [20], to investigate whether BifA-induced phosphorylation of moesin was dependent on PKC. When hBMEC cells were pre-treated with NSC305787 for 30 minutes before addition BifA, there was no induction of moesin phosphorylation (Fig 5E), consistent with the idea that BifA induction of moesin phosphorylation depends on PKC.
Moesin phosphorylation leads to activation of small G proteins, such as RhoA and Rac1 [21], that regulate actin cytoskeletal and membrane protrusion dynamics [22], phenotypes that could be pertinent to BifA-induced changes in brain endothelial cells and BBB permeability. RhoA and Rac1 activation are controlled by their conversion from GDP- to GTP-bound states [23], and we used immunoblots to monitor GTP-RhoA and GTP-Rac1 levels in BifA-treated cells. There was a time-dependent increase in GTP-RhoA that was associated with moesin phosphorylation in BifA-treated cells (Fig 5F). GTP-Rac1 levels were also increased during BifA treatment (S9 Fig). When BifA-induced moesin phosphorylation was blocked with NSC305787, GTP-RhoA formation was abrogated (Fig 5F), consistent with the idea that RhoA activation by BifA is dependent on moesin phosphorylation.
We next tested whether moesin phosphorylation was required for BifA-mediated barrier disruption. Monolayers pre-treated with NSC305787 did not exhibit increased permeability after addition of BifA (Fig 6A). Similarly, NSC305787 pre-treatment led to marked reduction in SEZ translocation across hBMEC monolayers (Fig 6B). Furthermore, we used siRNA to knockdown (KD) moesin in hBMECs (S10 Fig) to further investigate the requirement of moesin for BifA action. By itself, moesin KD did not alter the barrier function of the hBMEC monolayer, as these cells remained impermeant to Evans Blue dye (Fig 6A). However, the moesin KD cells exhibited significantly less permeabilization after BifA treatment, in marked contrast to control hBMEC monolayers treated with BifA (Fig 6A). Moreover, there was a marked reduction in the ability of SEZ to translocate across the moesin KD monolayer vs the WT monolayer, phenocopying the effects of NSC305787 (Fig 6B). Since NSC305787 inhibits all ERM family protein phosphorylation, the similarity of the phenotypes observed in the NSC305787 treated and moesin KD cells supports that idea that moesin is the dominant ERM protein in hBMECs. Thus, both blockade of moesin phosphorylation and moesin depletion are sufficient to protect cells from BifA-dependent bacterial translocation across the hBMEC monolayer. Collectively, these observations suggest that BifA disruption of hBMEC monolayer barrier integrity relies on moesin-dependent signaling pathways.
We found that the virulence of SEZ ATCC35246 depends on BifA, a Fic domain-containing protein encoded in a pathogenicity island. Deletion of bifA reduced its lethality in mice as well its capacity to disrupt the BBB, to colonize the brain and to traverse hBMEC monolayers in tissue culture. BifA bound to moesin, a host protein that regulates cytoskeletal processes. Inactivation of BifA’s Fic domain eliminated its capacity to enter hBMEC monolayers, increase monolayer permeability, and to bind to moesin. BifA activation of moesin appears to be critical for BifA’s modification of monolayer permeability, since either moesin knock down or pharmacologic inhibition of moesin activation abolished BifA-mediated increases in hBMEC permeability and SEZ penetration of a hBMEC monolayer. Collectively, our findings suggest that by usurping moesin-dependent signaling, BifA enables SEZ to efficiently penetrate the BBB.
Our observation that addition of BifA to hBMEC monolayers induced formation of gaps and increased monolayer permeability is consistent with the idea that BifA enables SEZ penetration of the BBB by disrupting the integrity of the brain endothelial monolayer, a critical constituent of the BBB. Concordant with this model, BifA’s action appears dependent on the associated phosphorylation of the ERM protein moesin, which is known to have diverse consequences that include activation of signaling pathways involved in cell adhesion, migration and invasion [21, 24–26]. In particular, we found that moesin phosphorylation following BifA treatment was linked to formation of RhoA-GTP, the active form of this small G protein known to regulate multiple cytoskeletal processes [27, 28]. Formation of RhoA-GTP is likely a consequence of moesin phosphorylation, since blockade of moesin phosphorylation with NSC305787 inhibited the generation of RhoA-GTP (Fig 5F). Notably, activation of RhoA has been shown to promote the dissolution of tight junctions, which serve to limit paracellular permeability between endothelial cells [29]. Loosening of tight junctions and additional factors (e.g. adherens junctions) that increase the adherence of adjacent cells in the brain endothelium could open a paracellular route for SEZ movement across the BBB (S11 Fig).
Additional pathogens manipulate RhoA signaling to reduce the integrity of the BBB. For example, E. coli K1’s CNF toxin’s modulation of RhoA activity is thought to be important for this common agent of neonatal meningitis to cross the BBB [5]; however, in this case, RhoA activation is thought to enable this pathogen to traverse the BBB via a transcellular route. RhoA activation by the brain-invasive fungal pathogen Cryptococcus neoformans also facilitates its traversal of the BBB [30]. Additional Fic domain-containing proteins are known to catalyze post-translational modification of Rho family proteins (e.g. AMPylation of Rho proteins by Vibrio parahaemolyticus VopS [12]), but to our knowledge, other Fic toxins have not been reported to modify BBB function.
Both the mechanisms of BifA release from SEZ and uptake into eukaryotic cells require elucidation. In contrast to several Fic domain-containing virulence factors described in other pathogens (e.g. VopS), BifA delivery into the eukaryotic cytosol does not rely on a bacterial type III or type IV secretion system. In this regard, BifA functions as a traditional bacterial toxin, mediating its own uptake into host cells. Similar to BifA, IbpA, a Fic-domain containing protein from the cattle pathogen Histophilus somni doesn’t require additional bacterial factors for uptake into bovine cells, where it AMPylates Rho family proteins [12]. It will be particularly interesting to determine whether the receptor(s) and pathways that mediate BifA uptake into host cells modulate its downstream function(s).
The characterized Fic domain-containing proteins produced by other pathogens catalyze post-translational modifications of target proteins upon entry into the eukaryotic cell cytosol [12]. Fic domain-containing proteins can directly phosphorylate their targets; e.g., Doc phosphorylates its target EF-Tu [31]. However, although we found that BifA directly binds to moesin (Fig 4) and that BifA treatment of endothelial cells resulted in elevated levels of phosphorylated moesin (Fig 5), we did not directly demonstrate that BifA phosphorylates moesin. The observation that the PKC kinase inhibitor NSC305787 blocked the induction of moesin phosphorylation in cells treated with purified BifA suggests that BifA may not directly phosphorylate moesin, but could instead promote its phosphorylation indirectly. For example, BifA could enhance the activity of a host kinase, akin to the action of the Pseudomonas syringae Fic-like T3SS effector AvrB, which leads to the phosphorylation of the plant immune regulator RIN4 by promoting the activity of the endogenous kinase MPK4 [32]. Alternatively, our observation that BifA binds to the phosphorylated form of moesin with greater affinity than to the non-phosphorylated form (Fig 5), raises the possibility that BifA stabilizes phospho-moesin, leading to its accumulation.
Interestingly, despite a high degree of overall conservation among sequenced SEZ isolates, bifA homologues are not found in other SEZ genomes. BifA is encoded in a SEZ ATCC35246 pathogenicity island, suggesting that this critical SEZ Fic-domain containing virulence factor was likely acquired via horizontal gene transfer, and thus that lateral gene exchange was a key step in the evolution of SEZ ATCC35246 as a pathogen. Acquisition of bifA alone may be sufficient to enhance BBB penetration by related organisms. BifA homologues are not present in other well-characterized meningeal pathogens, e.g. Group B streptococci, consistent with the idea that different CNS pathogens rely on distinct factors to traverse the BBB [5]. However, BifA homologues with intact Fic domains are present in a variety of other Gram-positive as well as Gram-negative organisms, many of which are usually thought of GI tract commensals, suggesting that BifA-like proteins may carry out functions beyond diminishing the integrity of the BBB. Finally, BifA’s capacity to increase BBB permeability may have medical applications in delivery of drugs and other agents to the brain.
Streptococcus equi subsp. zooepidemicus ATCC35246 (SEZ) was isolated from a dead pig in Sichuan Province, China. SAICAR gene SeseC_01334 (Genbank) was re-named bifA. A bifA deletion mutant and complemented strain were constructed using pSET4s, a Streptococcus-E. coli temperature sensitive suicide shuttle vector and expression plasmid pSET2 respectively (SI Appendix) [33]. SEZ was cultured in Todd Hewitt Broth (THB) (Becton, Franklin Lakes, NJ, USA) at 37°C. Human brain microvascular endothelial cells (hBMECs) were purchased from ScienCell Research Laboratories (Catalog #1000). HEK293T cells were purchased from American Type Culture Collection (ATCC number CRL-3216). Cells were cultured in DMEM (Gibco, Grand Island, NY, USA) with 10% fetal bovine serum (FBS) (Gibco, Grand Island, NY, USA) in a 37°C incubator containing 5% CO2.
The vectors pAcGFP-C and pCMV-C-HA were used for the respective expression of BifA and moesin in eukaryotic cells respectively. E. coli BL21 (DE3) plysS was used to express recombinant BifA and its variants with the pCold-SUMO vector; E. coli Rosetta (DE3) was used to express moesin and its variant proteins with pGEX-6p-1. The bifA gene was PCR amplified from SEZ genomic DNA and subcloned into the pAcGFP-C vector. The moesin cDNA was PCR amplified from human cDNA and subcloned into the pCMV-C-HA vector. For expression of His- or GST-tagged proteins, bifA was subcloned into pCold-SUMO vectors and moesin was subcloned into pGEX-6p-1. The mutations in bifA and moesin were constructed by PCR mutagenesis using the ClonExpress II One Step Cloning Kit (Vazyme Biotech Co., China). The plasmid constructs were verified by Sanger sequencing. E. coli DH5α was used for propagation of plasmids. All plasmid information and primers are listed in S2 Table.
An allele exchange vector for deletion of bifA was created by PCR amplification of fragments upstream and downstream of the bifA gene with primers of bifA-up-fwd/bifA-up-rev and bifA-dwn-fwd/bifA-dwn-rev, using SEZ ATCC35246 genome as template. The upstream and downstream PCR products were mixed 1:1, and primer pair bifA-up-fwd/bifA-dwn-rev were subjected to fusion PCR amplification. The fusion fragment was purified, digested with appropriate endonucleases, and then cloned into the same sites of the temperature-sensitive S. suis-E. coli shuttle vector pSET4s [33]. Plasmids were electroporated into SEZ (Bio-Rad, Gene Pulser Xcell, Voltage: 2300V, Capacitance: 25μF, Resistance: 200Ω, Cuvette: 1mm) and mutant isolation was carried out as described [34]. The bifA gene was amplified using the primer pair bifA-pSET2-fwd and bifA-pSET2-rev using SEZ genomic DNA as the template, and then inserted into the pSET-2 plasmid. This plasmid was used to complement the bifA deletion in the ΔBif strain. Templates containing mutant versions of bifA were amplified from the expression vectors used above and subcloned into pSET-2. The inserts in all plasmids were confirmed by Sanger sequencing.
All animal experiments were performed with protocols approved by the College of Veterinary Medicine of Nanjing Agricultural University for Research Protection Standing Committee on Animals in accordance with Science and Technology Agency of Jiangsu Province guidelines (SYXK2017-007). All efforts were made to minimize animal suffering.
Four-week-old female BALB/c mice, purchased from the Comparative Medicine Center of Yangzhou University, were used for all animal work. In bacterial load determination of different organs in Fig 1A, mice (10/group) were i.p. challenged with 1×105 CFU of WT or mutant SEZ. In the mortality experiments shown in Fig 1B, mice (20/group) were i.p. challenged with 5×105 CFU of WT or mutant SEZ. For bacterial load determinations in Fig 1C, mice (10/group) were i.p. challenged with 5×105 CFU of WT or mutant SEZ and CFU counts from the brain were determined 2 days later. For pathology (S3 Fig), brains were harvested and embedded in paraffin and sectioned for hematoxylin and eosin staining, 48 hours after intravenous injection of 1×106 CFU of WT and mutant SEZ.
Evans Blue (EB) leakage was used to assess BBB permeability as described [13]. In these experiments, mice were challenged i.v. with 5×106 CFU of WT or mutant SEZ, and 18 hour later, 100μl of 2% EB was administered i.v. Two hours later, brains were dissected, photographed and then dried at 56°C in aluminum foil for two days. Formamide was used to extract the EB out of the tissue and EB amounts were determined as absorbance at OD620.
To purify His-tagged BifA and its variants, cultures of BL21 (pCold-SUMO-bifA) was grown to OD600 = 1.0 in Luria-Bertani (LB) broth at 37°C, 180 rpm then induced with IPTG at a final concentration of 1 mM at 16°C, 180 rpm for 24 hours. Bacterial cells were collected by centrifugation. Cells were lysed in the lysis buffer (10mM Tris-HCl, pH 8, containing 100 mM NaCl and 20mM imidazole) by sonication. The cell lysate was centrifuged and the supernatant was used for purification. SUMO tag was digested with SUMO protease (Thermo, Waltham, MA, USA). Primary purification was performed using the Histrap HP column (GE Healthcare, Piscataway, NJ, USA), followed by Superdex 200 10/300 GL gel filtration (GE Healthcare, Piscataway, NJ, USA) using an AKTA Protein Purifier (GE Healthcare, Piscataway, NJ, USA).
To purify non-tagged moesin for SPR, cultures of BL21 (pGEX-6p-1-msn) were grown until OD600 = 0.8 in LB at 37°C, 180 rpm then with IPTG at a final concentration of 1 mM at 30°C, 180 rpm for 5 hours. Cells were lysed in the lysis buffer (PBS, 140 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 1.8 mM KH2PO4, pH 7.3) by sonication. The lysates were used for purification with GSTrap HP 5 ml column (GE Healthcare, Piscataway, NJ, USA) followed by tag digestion with Precission Protease and GSTrap FF 1 ml column (GE Healthcare, Piscataway, NJ, USA), Sephadex 10/300 (GE Healthcare, Piscataway, NJ, USA) was used to final purification in AKTA Protein Purifier (GE Healthcare, Piscataway, NJ, USA).
WT hBMEC or moesin knock-down hBMEC were seeded on the apical side of collagen-coated polytetrafluoroethylene (PTFE) 3 μM pore-size membranes (Corning Incorporated, Corning, NY, USA) for the bacterial penetration assay and 0.4 μM pore-size membranes (Corning Incorporated, Corning, NY, USA) for the barrier integrity assay. Cells were grown for 7 to 10 days to form intact monolayers. Barrier integrity was assessed with 0.4% Evans Blue solution [35].
Bacterial cells (1×106 CFU) were added to the upper chamber of transwells containing hBMEC and incubated at 37°C in 5% CO2 for 2 hours. The 100 μl medium from both sides of the transwells was collected and spread on THB agar plates for CFU determination [36]. To examine the effect of BifA on barrier integrity, 10 μg/ml BifA was added to the upper chamber of transwells. Transwell inserts were then transferred to a fresh plate containing Hanks Balanced Salt Solution (HBSS) in the bottom chamber and 50 μl of 0.4% Evans Blue solution in PBS was added to the upper chamber. Transwell inserts were incubated at 37°C in 5% CO2 for 40 min, and the permeability was assessed by colorimetric quantification at OD600 nm of the bottom chamber as described [37].
SEZ was grown to an OD600 = 0.6 in THB media at 37°C with vigorous shaking (180 rpm). Bacteria were diluted 1:100 in DMEM media and grown for 12 h at 37°C with vigorous shaking (180 rpm) and then culture supernatants were isolated by centrifugation. Proteins were precipitated from supernatants with TCA-acetone as described [38, 39].
Approximately 106 sulfate-modified fluorescent red polystyrene latex beads (0.1 μm mean particle size, Sigma-Aldrich, St. Louis, MO, USA) were suspended in 200 μL of 25 mM 2-(N-morpholino) ethanesulfonic acid (MES) buffer, pH 8.0. Purified BifA or BifA variants were dissolved in 10 μL 25mM phosphate buffer (pH 7.2) and incubated with the suspended beads at 4°C overnight with gentle mixing. We sequentially added 25mM MES buffer (150 μL) every 15 min until the original volume was diluted 200-fold. The coated beads were collected by centrifugation at 3000 × g for 20 min, washed twice in MES buffer and resuspended in DMEM without FBS and then sonicated for dispersal. Dot blot assays were used to confirm protein coating of the beads [40].
For transmission electron microscopy, coated beads were incubated with hBMEC (100:1) for 4h at 37°C. Extracellular beads were removed by washing with PBS and then samples were fixed in 2.5% paraformaldehyde and 0.1% glutaraldehyde in 0.05 M cacodylate buffer, pH 7.3. Then, 0.03% CaCl2 was added to the mixture. After fixation, the cells were washed with 0.1 M cacodylate buffer, and pelleted by centrifugation. Low melting point agar was pre-embedded and stained with 1% uranyl acetate in 0.1 M maleate buffer, then dehydrated in ethanol. Ultrathin sections were cut, stained with lead citrate and examined using a JEM 1400-PLUS electron microscope (JEOL, Tokyo, Japan) [41].
The hBMEC were seeded onto 15 mm Glass Bottom Cell Culture Dishes (Corning, NY, USA) and treated with 10 μg/ml BifA or BifA variants for 2 hours. Cells were then fixed with 4% paraformaldehyde followed by 0.1% Triton X-100 permeabilization buffer and blocked with 5% BSA in PBS-Tween. Mouse polyclonal anti-BifA antibody, Alexa 488-conjugated goat anti-mouse antibody (Jackson Immunoresearch, West Grove, PA, USA), rabbit anti-Moesin antibody (Abcam, Cambridge, MA, USA) and Alexa 594-conjugated goat anti-rabbit antibody (Jackson Immunoresearch, West Grove, PA, USA) were used at 1:2000 in PBS containing 1% BSA. Primary antibodies were incubated for 2 hours and secondary antibody for 1.5 hours at room temperature. 4,6- Diamidino-2-phenylindole (DAPI) was used to detect cell nuclei. Plates were washed three times with Phosphate buffered saline with Tween-20 (PBST) with shaking to wash out unbound antibodies. Images were obtained on a laser scanning confocal microscope (LSCM) (ZEISS, Japan).
The hBMEC cells were cultured on 6-well Glass Bottom Plates (∅35mm, Cellvis, CA, USA) for 7–9 days until monolayers were confluent. Cells were replenished with DMEM medium, and BifA or BifA variants, at a final concentration of 10 μg/ml, was added to the wells. The plates were cultured in a controlled environmental chamber at 37°C in 5% CO2. Time-lapse images were acquired at an interval of 30 s for 300 min through an EC Plan-Neofluar 20×/0.50 M27 lens on an Axiom Observer.Z1/7 microscope, using the Applied Precision motorized stage (Carl Zeiss, Japan). ZEN software was used for image processing.
Stable isotope labelling of amino acids in cell culture (SILAC) was used to identify BifA interacting host proteins in HEK293T cells. Cells were labeled with heavy isotopes (Arg13C6, Lys13C6) or light isotopes (Arg12C6, Lys12C6) in Dulbecco’s modified Eagle medium (DMEM) with 10% FBS (Pierce, Rockford, IL, USA) at 37°C in 5% CO2 as previously described [42]. The cells were passaged for 6 generations to ensure adequate labeling of proteins. The heavy and light labeled cells were seeded in 10 cm cell culture dishes and transfected with 24 μg of pAcGFP-BifA or pAcGFP using Lipofectamine 2000 (Thermo, Waltham, MA, USA) respectively. Transfected cells were lysed in 500 μl cold Mammalian Protein Extraction Reagent (Thermo, Waltham, MA, USA), containing a protease inhibitor cocktail (Thermo, Waltham, MA, USA) and centrifuged at 14000 g for 10 min at 4°C. Protein concentrations were measured using the BCA Protein Assay Kit (Pierce, Rockford, IL, USA) according to the manufacturer’s directions. We mixed equal quantities of heavy and light lysates and pre-cleared them on Protein G agarose (Santa Cruz, Santa Cruz, CA, USA) with 100 μg mouse IgG (CMCTAG, Milwaukee, WI, USA) for 1h at 4°C with gentle agitation. Pre-cleared lysates were centrifuged and the supernatants transferred to new tubes. Mouse anti-GFP antibody (CMCTAG, Milwaukee, WI, USA) was added to the cold lysates and incubated at 4°C for 1 h, then 40 μl protein G agarose was added and incubated at 4°C on a rotating device overnight. Beads were washed five times with 1 ml of cold PBS. After the final wash, the bound proteins were eluted with 50 μl of elution buffer (50mM Tris-HCl, 1% SDS) and samples were boiled for 5 min. The eluted proteins were digested as previously described [35]. The peptides were separated by reverse-phase liquid chromatography using a nano-LC system (DIONEX Thermo Scientific) and analyzed by tandem mass spectrometry using an LTQ-Orbitrap mass spectrometer (Thermo Scientific) with a nanoelectrospray ion source.
HEK293T cells were seeded in 10 cm cell culture dishes, which were each transfected with 24 μg of pAcGFP-BifA and pCMV-HA-Moesin, or pAcGFP and pCMV-HA plasmids. In addition, truncated fragments of moesin were obtained by PCR and cloned into pCMV-HA plasmids. The resulting plasmids, pCMV-HA-Moesin FERM domain (1-470aa) and pCMV-HA-Moesin C-ERMAD domain (470-577aa) were co-transfected with pAcGFP-BifA plasmid respectively. Lysates were harvested 48 h later with lysis buffer and cleared by centrifugation as described above. Twenty microliters of lysate was saved for analysis of the expression efficiency and the remainder was immunoprecipitated with Protein G agarose bound to either anti-HA or anti-GFP specific antibody (CMCGAT, Milwaukee, WI, USA). Immunoprecipitated beads were resuspended and boiled for 5 min in 1× Laemmli sample buffer and then used for Western blot analysis.
Boiled cell lysates were subjected to SDS-PAGE, followed by transfer to a PVDF membrane (Roch, Basel, Switzerland) using a semi-dry transfer apparatus (GE Healthcare). Membranes were blocked in 5% non-fat milk powder in TBS containing 0.01% Tween 20 (TBST). Primary antibodies were used and diluted as follows: 1:500 anti-BifA rabbit polyclonal antibody; 1:1000 anti-Moesin (Abcam, Cambridge, MA, USA); 1:1000 anti- phospho T558-Moesin (Abcam, Cambridge, MA, USA); 1:2000 anti-HA (CMCGAT, Milwaukee, WI, USA); 1:2000 anti-GFP (CMCGAT, Milwaukee, WI, USA); 1:2000 Anti-ZO-1 tight junction protein antibody (Abcam, Cambridge, MA, USA) and 1:2000 anti-GAPDH (CMCGAT, Milwaukee, WI, USA). Membranes were incubated with primary antibody diluted in TBST containing 1% BSA overnight at 4°C and then washed for 30 min in TBST. This was followed by incubation with 1:5000 HRP goat anti-rabbit or goat anti-mouse IgG antibody (ABGENT, San Diego, CA, USA). Membranes were washed for 3 × 15 min in TBST before adding ECL reagent (Thermo, Waltham, MA, USA). Chemiluminescence was detected on a ChemiDoc system (Bio-Rad, Hercules, CA, USA).
The direct binding potential of BifA and moesin was analyzed using the Biacore X100 instrument (GE Healthcare). Recombinant His-BifA protein and its variants were separately immobilized onto Biacore NTA sensorchips (GE Healthcare). The recombinant moesin protein, moesin T558A or moesin T558D were injected individually and the binding interactions were recorded. Results were analyzed using the Biacore X100 Evaluation Software (GE Healthcare).
The hBMEC were seeded into 6-well plates (Corning Incorporated, Corning, NY, USA). For serum starvation, once cells reached 70% confluency, the medium was replaced by low serum medium (1% FBS) for 24h, and then replaced with DMEM without FBS and incubated for 16–20 h. After serum starvation, BifA or BifA variants at a concentration of 10 μg/ml in DMEM were added to cells. At different time points (0–120 min), cells were harvested for protein extraction. Cells were lysed with M-PER Mammalian Protein Extraction Reagent (Thermo, Waltham, MA, USA) on ice in the presence of Halt Protease and Phosphatase Inhibitor Cocktail, EDTA-free (100×). In some cases, cells were pretreated with 10 μM of a PKC inhibitor of ERM protein phosphorylation, NSC305787 [20] (MedChemExpress, Monmouth Junction, NJ, USA) for 30 min prior to addition of BifA. Phosphorylation levels of extracted proteins were detected by Western blot with anti-phospho T558-Moesin antibody (Abcam, Cambridge, MA, USA).
Phos Binding Reagent Acrylamide (PBR-A) (APExBIO, TX, USA) was also used to detect moesin phosphorylation. Briefly, extracted proteins were electrophoresed in a 6% acrylamide gel containing 50 mM PBR-A and 10 mM Mn2+. After electrophoresis, the gel was soaked in a general transfer buffer containing 10 mmol/L EDTA for a minimum of 10 minutes with gentle agitation, followed by gentle agitation in buffer without EDTA for 10 minutes. The gel was then transferred to PVDF membranes for Western-blotting with a rabbit anti-Moesin antibody (Abcam, Cambridge, MA, USA)
The amounts of GTP-RhoA or GTP-Rac1 in cell lysates were measured by a pull-down method based on specific binding to Rhotekin- RBD coated beads for 1h at 4°C under gentle rotation, followed by western blot with an antibody specific for the GTP-bound form of RhoA or Rac1(RhoA/Rac1/Cdc42 Activation Assay Combo Biochem Kit, Cytoskeleton, Inc., USA). Total RhoA or Rac1 in cell lysates were detected by anti-RhoA and anti-Rac1 antibodies respectively.
An shRNA targeting the moesin gene was ligated into the pLVX-shRNA1 vector (Clontech, Mountain View, CA, USA). The lentivirus was packaged using commercial reagents (Applied Biological Materials, Nanjing, JS, China). Lentiviral particles (MOI = 1) were added to the hBMEC and seeded in 6-well plates, with media change after 24 h of incubation. After transduction, cellular protein was extracted for Western blot detection, while the RNA was extracted at different time points for Quantitative Real-Time PCR detection. Moesin transcripts levels were determined using the ABI Prism 7300 and Sequence Detection System software (Applied Biosystems, Foster City, California, USA). The results were obtained using the mathematical model Ratio = 2-ΔΔct [43]. 72 h post transduction, puromycin was used for selecting positive colonies at final concentration 2 μg/ml. The sequences of the oligos used are found in S2 Table.
|
10.1371/journal.pntd.0001657 | Drug Susceptibility in Leishmania Isolates Following Miltefosine Treatment in Cases of Visceral Leishmaniasis and Post Kala-Azar Dermal Leishmaniasis | With widespread resistance to antimonials in Visceral Leishmaniasis (VL) in the Indian subcontinent, Miltefosine (MIL) has been introduced as the first line therapy. Surveillance of MIL susceptibility in natural populations of Leishmania donovani is vital to preserve it and support the VL elimination program.
We measured in vitro susceptibility towards MIL and paromomycin (PMM) in L. donovani isolated from VL and PKDL, pre- and post-treatment cases, using an amastigote-macrophage model. MIL susceptibility of post-treatment isolates from cured VL cases (n = 13, mean IC50±SD = 2.43±1.44 µM), was comparable (p>0.05) whereas that from relapses (n = 3, mean IC50 = 4.72±1.99 µM) was significantly higher (p = 0.04) to that of the pre-treatment group (n = 6, mean IC50 = 1.86±0.75 µM). In PKDL, post-treatment isolates (n = 3, mean IC50 = 16.13±2.64 µM) exhibited significantly lower susceptibility (p = 0.03) than pre-treatment isolates (n = 5, mean IC50 = 8.63±0.94 µM). Overall, PKDL isolates (n = 8, mean IC50 = 11.45±4.19 µM) exhibited significantly higher tolerance (p<0.0001) to MIL than VL isolates (n = 22, mean IC50 = 2.58±1.58 µM). Point mutations in the miltefosine transporter (LdMT) and its beta subunit (LdRos3) genes previously reported in parasites with experimentally induced MIL resistance were not present in the clinical isolates. Further, the mRNA expression profile of these genes was comparable in the pre- and post-treatment isolates. Parasite isolates from VL and PKDL cases were uniformly susceptible to PMM with respective mean IC50 = 7.05±2.24 µM and 6.18±1.51 µM.
The in vitro susceptibility of VL isolates remained unchanged at the end of MIL treatment; however, isolates from relapsed VL and PKDL cases had lower susceptibility than the pre-treatment isolates. PKDL isolates were more tolerant towards MIL in comparison with VL isolates. All parasite isolates were uniformly susceptible to PMM. Mutations in the LdMT and LdRos3 genes as well as changes in the expression of these genes previously correlated with experimental resistance to MIL could not be verified for the field isolates.
| Resistance to antimonials has emerged as a major hurdle to the treatment and control of VL and led to the introduction of Miltefosine as first line treatment in the Indian subcontinent. MIL is an oral drug with a long half-life, and it is feared that resistance may emerge rapidly, threatening control efforts under the VL elimination program. There is an urgent need for monitoring treatment efficacy and emergence of drug resistance in the field. In a set of VL/PKDL cases recruited for MIL treatment, we observed comparable drug susceptibility in pre- and post-treatment isolates from cured VL patients while MIL susceptibility was significantly reduced in isolates from VL relapse and PKDL cases. The PKDL isolates showed higher tolerance to MIL as compared to VL isolates. Both VL and PKDL isolates were uniformly susceptible to PMM. MIL transporter genes LdMT/LdRos3 were previously reported as potential resistance markers in strains in which MIL resistance was experimentally induced. The point mutations and the down-regulated expression of these transporters observed in vitro could, however, not be verified in natural populations of parasites. LdMT/LdRos3 genes therefore, do not appear to be suitable markers so far for monitoring drug susceptibility in clinical leishmanial isolates.
| Visceral Leishmaniasis (VL) is a potentially fatal protozoan infection with members of the Leishmania donovani complex as the causative species. This poverty related disease is endemic in 70 countries with a total of 200 million people at risk and an estimated 100,000 new infections annually concerning all age groups [1], [2]. More than 90% of the estimated VL cases occur in India, Bangladesh, Nepal, Sudan and Brazil [3] with India alone sharing almost 50% of the world's total disease burden. Post-kala-azar dermal leishmaniasis (PKDL) is a dermal sequel of VL that develops in 5–15% of the cured VL patients in India and in 60% cured VL patients in Sudan and is considered to constitute a major parasite reservoir in these regions. In the present situation, chemotherapy is the key strategy for VL control due to the absence of vaccines and the limited impact of vector control [4]–[6]. The situation is particularly grave in Bihar, India, where more than 60% of VL patients do not respond to traditional first line antimonial therapy. The use of Amphotericin B and its liposomal formulations, although highly effective even in antimony unresponsive patients, has limitations because of its renal toxicity, high costs and inconvenience due to slow I.V. based administration [7], [8].
The first oral antileishmanial drug Miltefosine (MIL), an alkylphosphocholine, has proved to be highly effective against VL with cure rates of 94%, including cases unresponsive to antimony [2], [9]. It was therefore, proposed as the first line VL therapy and remains the mainstay in the Kala-azar elimination program, that aims to reduce the incidence of VL to 0.0001% in the endemic areas of the Indian subcontinent by the year 2015 [10], [11]. Efficacy of MIL has also been established for the treatment of PKDL in India [12], [13]. In absence of directly observed therapy (DOT), widespread misuse of this self-administered drug could contribute to the rapid emergence of MIL resistance in the field. Moreover, phase IV trial of MIL in India suggested doubling of the relapse and failure rate compared to phase III trials [8], [14]. Treatment failures (almost all relapses) were recently also observed in Nepal [15]. Alarmingly, VL and PKDL treatment failure and relapse cases have already surfaced in a significant fraction of MIL treated patients in India (unpublished data). Unresponsiveness to MIL has been reported for cutaneous cases due to L. braziliensis and Leishmania (Viannia) guyanensis in South America [16], [17]. Paromomycin (PMM) is an aminoglycoside antibiotic exhibiting high efficacy towards VL [18]. Data from Phase IV trials confirm the safety and efficacy of PMM to treat VL [19].
The development and spread of drug resistance has made surveillance of drug susceptibility a high priority. In the present investigation, through long-term monitoring of MIL treated VL/PKDL patients, post-treatment stage parasite isolates were obtained either as residual parasites soon after completion of treatment or from cases that relapsed, and were compared to a set of pre-treatment isolates. The MIL susceptibility of these isolates was assessed with an in vitro intracellular amastigote assay and correlated with the clinical outcome. In addition, intrinsic sensitivity towards PMM was evaluated in the same set of clinical isolates in order to obtain baseline susceptibility prior to its future use in therapy.
An impaired functioning of P-Type ATPase transporters, the LdMT–LdRos3-dependent flippase machinery, resulting in a significant decrease of intracellular MIL concentration was observed in experimentally induced MIL resistant leishmanial parasites. [20], [21]. The resistant phenotype was related to the occurrence of two missense and a nonsense point mutations in the LdMT gene, and a nonsense mutation in the start codon of the LdRos3 gene. In our study, we have evaluated the presence of these mutations and the mRNA expression levels of LdMT and LdRos3 in our set of clinical isolates to explore their role as molecular markers for monitoring MIL susceptibility in the field.
Clinical isolates of L. donovani were prepared from splenic aspirates of VL patients reporting to KAMRC, Muzzafarpur, Bihar or from dermal lesions of PKDL patients reporting to Department of Dermatology, Safdarjung Hospital (SJH), New Delhi under the guidelines of the Ethical Committee of the respective Institute. All patients came from zones of high endemicity in Bihar, India. VL patients received MIL treatment for 28 days (50 mg capsule twice) while PKDL patients received MIL for 60 days (50 mg capsule, thrice daily). Splenic smears from all VL patients were examined microscopically at the pre and post-treatment stages. Patients with negative or +1 smear grade were not treated further. However, patients with smears grading ≥2 [22] were treated with amphotericin B deoxycholate.
Parasites isolated before onset of treatment were assigned XXX/0 codes. MIL treatment led to complete subsidence of VL symptoms, interpreted as clinical cure, although residual parasites could be cultured from splenic aspirates in a substantial number of patients at the end of 1 month of treatment. These post-treatment isolates were assigned with XXX/1 codes. All cases were followed up for one year. VL and PKDL cases that relapsed after an initial cure were treated with Amphotericin B and cured. Parasites were isolated from each of the relapse cases at the time of reported relapse (after four, six and seven months of MIL treatment completion in VL and after 12, 18 and 32 months of MIL treatment in PKDL and were designated as XXX/month in which relapse occurred.
Parasites were routinely grown in Medium 199 (Sigma) with 10% heat-inactivated fetal bovine serum (HI FBS, Gibco, USA) at 25°C.
Parasite DNA, isolated using Qiagen Kit, was subjected to ITS-1 PCR-RFLP analysis for species characterization, as described earlier [23]. The ITS1 PCR product was digested with 1 U HaeIII enzyme (Genei, Bangalore, India) at 37°C for 2 hours, followed by analysis on 2% agarose gel. All isolates were characterized as L. donovani on the basis of their RFLP pattern.
The drug susceptibility of L. donovani parasites was assessed as intracellular amastigotes, as previously described [24]. Briefly, J774A.1 cells (1×105 cells/ml) were infected with stationary phase promastigotes at ratio of 10 ∶ 1 (parasite ∶ macrophage), plated into 16 well chambered Labtek tissue culture slides and incubated for 4 h at 37°C in 5% CO2. Excess, non-adhered promastigotes were removed by washing and macrophages incubated for 18–24 h. Infected cells were re-incubated for 48 h, with MIL (1, 5, 10, 20 and 30 µM) (Paladin) or PMM (1, 5, 10, 20, 30 and 40 µM) (Gland Pharma). Macrophages were then examined for intracellular amastigotes after staining with Diff-Quik solutions. The number of L. donovani amastigotes was counted in 100 macrophages, at 100× magnification. The survival rate of parasites relative to untreated macrophages was calculated and IC50 were determined by sigmoid regression analysis. The assays were performed in duplicate and repeated at least twice.
Fourteen parasite isolates were tested for the presence of previously described point mutations in the LdMT gene, at positions T421N, L856P and W210*, and the LdRos3 gene, at position M1*(Figure S1) [20], [21]. Primers were designed targeting these mutations and their adjacent regions using Primer 3 software [25]. PCR reactions were performed with 35 ng of genomic DNA in a 50 µl reaction volume containing 10× PCR buffer (incl. 15 mM MgCl2), 5 U/µl of Taq polymerase (Roche), 2.5 mM dNTPs (NEB), and 10 µM of each of the locus specific primers (Table S1). Thermocycling conditions were as follows: initial denaturation at 95°C for 5 min, followed by 35 cycles of denaturation at 95°C for 30 sec, annealing with the specific primer pair at the specific TA°C for 30 sec and extension at 72°C for 1 min, and a final extension step at 72°C for 6 min. PCR products with fragment sizes between 149 and 277 bp were purified (QIAamp DNA Mini Kit, Qiagen) and sent for commercial sequencing using forward and reverse primers (SMB Services in Molecular Biology, Berlin Germany). Sequences were checked, trimmed and aligned using Chromas Pro v 1.32 [26].
Total RNA (5 µg), isolated from stationary phase promastigotes at day six using Trizol Reagent (Invitrogen, USA) was reverse transcribed at 42°C with M-MLV Reverse transcriptase (Invitrogen, USA) after deoxyribonuclease I treatment. All real-time PCR reactions were performed in duplicate in 25 µl volumes using SYBR Green as described before [27]. The 2−ΔΔCT method was used to calculate relative changes in gene expression determined from real-time quantitative PCR experiments. The data was presented as the fold change in the target gene expression in L.donovani parasites normalized to the internal control gene (GAPDH) and relative to the LdAG83 reference strain of L. donovani.
The study was approved by the Ethical Committee of the Institute of Medical Sciences, Banaras Hindu University, Varanasi and Safdarjung Hospital, New Delhi, India. Written informed consent was obtained from patients and from guardians in case of children <18 years.
MIL susceptibility was determined at intracellular amastigote stage for thirty clinical isolates, including eleven pre-treatment isolates (six from VL and five from PKDL cases), thirteen post-treatment isolates (obtained at the end of treatment from VL patients which depicted a clinical cure, although parasitology was still positive) and six relapse isolates (three VL and three PKDL). The clinical profile of patients and in vitro susceptibility of the isolates to MIL are summarized in Table 1.
The six VL pre-treatment isolates showed a sensitivity range of 0.95±0.06 to 2.91±0.24 µM towards MIL with the mean IC50±SD being 1.86±0.75 µM (Figure 1). The post-treatment VL isolates had a mean IC50 of 2.43±1.44 µM (range 1.02±0.17 to 5.20±0.80 µM) which was not significantly different in comparison with the pre-treatment group (p>0.05). The mean IC50 of the three VL isolates from relapse cases (4.72±1.99 µM) was significantly higher (p = 0.04) than that of pre-treatment VL cases (1.86±0.75 µM).
The three PKDL relapse isolates showed a significantly higher (p = 0.03), mean IC50 of 16.13±2.64 µM in comparison with the pre-treatment PKDL isolates (mean IC50 = 8.63±0.94 µM). The mean susceptibility of all VL isolates (2.58±1.58 µM, n = 22) when compared with PKDL isolates (11.45±4.19 µM, n = 8) revealed that the latter were significantly (p<0.0001) more tolerant to MIL (Figure 1).
We evaluated PMM susceptibility of 22 VL isolates, of which 16 were exposed to MIL treatment. The IC50 ranged from 3.41±0.29 to 10.70±1.12 µM with mean IC50 = 7.05±2.24 µM (Table 1). Furthermore, the PMM sensitivity was similar (p>0.05) in parasites non-exposed (mean IC50 = 7.73±2.25 µM) or exposed (mean IC50 = 6.79±2.25 µM) to MIL (Figure 2).
The inherent PMM susceptibility of 8 PKDL isolates ranged from 4.92±0.34 to 8.62±1.82 µM (Table 1). Like in VL, the PMM IC50 of PKDL isolates was similar (p>0.05) in parasites non-exposed (n = 5, mean IC50 = 6.12±1.40 µM) or exposed (n = 3, mean IC50 = 6.29±2.02 µM) to MIL (Figure 2). There was no correlation observed between MIL and PMM susceptibility in VL (r = 0.10) or PKDL isolates (r = −0.02).
None of the four reported SNPs in the LdMT and LdRos3 genes, which were suggested to be responsible for the resistant phenotype in a strain from Ethiopia, MHOM/ET/1967/HU3_MIL-R, with experimentally induced MIL resistance [21], could be detected in the clinical isolates investigated (Figure S1). Furthermore, no SNPs were detected when the three LdMT and the LdRos3 gene fragments were sequenced in 15 clinical isolates from Nepalese VL cases that relapsed after MIL treatment (data not shown). Notably, the LdRos3 gene fragment could not be amplified for the strains BHU800/1 and BHU1062/4 although no sequence polymorphisms were identified in the primer annealing sites.
mRNA expression level of LdMT and LdRos3 was analyzed in 19 VL and two PKDL isolates using real-time PCR in comparison to the reference strain L. donovani LdAG83. The expression was found comparable in all the groups including the relapse cases of VL and PKDL (Figure 3).
The introduction of MIL therapy as treatment for VL has pioneered the era of effective oral therapy for this potentially fatal disease. However, anthroponotic VL transmission in the Indian subcontinent and the long half life of MIL (150–200 h) poses the risk of development of resistance in natural population of parasites. It has been reported earlier that MIL resistant parasites can be easily generated in vitro [28]. The present study reveals for the first time, the intrinsic in vitro sensitivity of Indian L. donovani isolates from a set of VL and PKDL patients treated with MIL (including both responders and relapse cases). The data provides information on the extent of MIL-tolerance in natural populations following MIL treatment highlighting the need for adequate monitoring of drug susceptibility to preserve this valuable drug.
The drug susceptibility of the currently prevailing clinical isolates was similar to the L. donovani parasites from the era of pre- MIL treatment reported earlier [27]. At the end of MIL treatment all cases showed clinical cure although some of them were parasitologically positive and residual parasites could be cultured from splenic aspirates of such patients (Table 1). The drug susceptibility of L. donovani parasites isolated at the end of therapy was comparable to that of pre-treatment isolates. On the contrary, parasites obtained from the cases that relapsed exhibited significantly reduced susceptibility to MIL, although the IC50 values were below the expected serum threshold levels [29], implicating the possible involvement of host factors in rendering tolerance to drug. Indeed, reports on VL relapse in HIV co-infected patients treated with MIL suggest that host immunity plays a role in the elimination of parasites from VL patients [30], [31]. A small number of MIL treated cases showed relapse, parasite isolates from these were monitored for in vitro drug susceptibility. Although we did not find any clinically resistant strains, the observation of strains with higher MIL tolerance (up to eight times compared to the sensitive ones) emphasizes the need for close monitoring of cases under MIL treatment. The study also investigated for the first time, the intrinsic susceptibility of PKDL isolates towards MIL. The in vitro susceptibility of PKDL isolates was significantly higher in pre-treatment isolates than in isolates originating from relapse patients, which were exposed to MIL for long duration (over two months). The IC50 of PKDL isolates was significantly higher (∼4 fold) compared to VL isolates, a trend similar to that reported earlier for SAG susceptibility in isolates from high endemic regions [32]. This reduced drug susceptibility of PKDL isolates may be due to longer treatment regime in PKDL and prolonged exposure of parasites to the drug. PMM exhibited similar in vitro susceptibility in pre- and post-treatment isolates suggesting its potential in future VL and PKDL therapy.
The inactivation of the genes essential for MIL uptake has been proposed as the simplest mechanism of resistance towards the drug and L. donovani MIL transporter LdMT and its subunit LdRos3 have been reported as markers of experimental MIL resistance [28]. Experimentally induced MIL resistant L.donovani showed down regulated expression of these transporters [27].
In the current study, a selection of strains from Indian L. donovani with variable response to MIL treatment has been tested for four point mutations that were suggested to underlie the development of MIL resistance. No nucleotide exchanges were however, detected in the LdMT and LdRos3 gene fragments sequenced for the set of clinical isolates studied herein, or for clinical isolates from Nepalese MIL relapse cases (data not shown). Screening of whole-genome data revealed that both genes are highly conserved for the examined strains of L. donovani from the North of the Indian subcontinent, regardless whether they were isolated from cases responding or not responding to MIL treatment [33]. Comparison of the expression of LdMT and LdRos3 genes revealed a similar expression profile in the different groups of isolates studied herein. Further studies in truly resistant parasites from MIL treated cases, when available, are necessary to explore the possible utility of these genes as markers for monitoring drug susceptibility in clinical isolates. In conclusion, the causative forces leading to MIL resistance cannot be explained by the genomic data available up to date and it is very likely that multi-factorial events may be responsible for the tolerance to chemotherapeutics in L. donovani [33].
The study employed an amastigote-macrophage model for monitoring drug susceptibility towards MIL as this stage mimics the host milieu. However, amastigote assays are tedious, time consuming and technically demanding. Hence, drug sensitivity assays based on promastigotes, if found relevant, would be better as simplified biological tool that can be used in clinical settings. The current data recommends for keeping miltefosine susceptibility under close surveillance in the field. The risk of relapse after MIL therapy presses the need for maintenance regimen such as DOT for this oral drug and exploring new drug combinations for regions endemic for VL. Regional policies concerning judicious use of the drug and monitoring the treatment outcome should be implemented and supervised by the health authorities in the endemic areas to minimize the risk of emergence of MIL resistant strains. The development of markers to identify drug unresponsiveness at an early stage constitutes an essential step towards the elimination of this poverty driven disease.
|
10.1371/journal.ppat.1005741 | An In Vivo Selection Identifies Listeria monocytogenes Genes Required to Sense the Intracellular Environment and Activate Virulence Factor Expression | Listeria monocytogenes is an environmental saprophyte and facultative intracellular bacterial pathogen with a well-defined life-cycle that involves escape from a phagosome, rapid cytosolic growth, and ActA-dependent cell-to-cell spread, all of which are dependent on the master transcriptional regulator PrfA. The environmental cues that lead to temporal and spatial control of L. monocytogenes virulence gene expression are poorly understood. In this study, we took advantage of the robust up-regulation of ActA that occurs intracellularly and expressed Cre recombinase from the actA promoter and 5’ untranslated region in a strain in which loxP sites flanked essential genes, so that activation of actA led to bacterial death. Upon screening for transposon mutants that survived intracellularly, six genes were identified as necessary for ActA expression. Strikingly, most of the genes, including gshF, spxA1, yjbH, and ohrA, are predicted to play important roles in bacterial redox regulation. The mutants identified in the genetic selection fell into three broad categories: (1) those that failed to reach the cytosolic compartment; (2) mutants that entered the cytosol, but failed to activate the master virulence regulator PrfA; and (3) mutants that entered the cytosol and activated transcription of actA, but failed to synthesize it. The identification of mutants defective in vacuolar escape suggests that up-regulation of ActA occurs in the host cytosol and not the vacuole. Moreover, these results provide evidence for two non-redundant cytosolic cues; the first results in allosteric activation of PrfA via increased glutathione levels and transcriptional activation of actA while the second results in translational activation of actA and requires yjbH. Although the precise host cues have not yet been identified, we suggest that intracellular redox stress occurs as a consequence of both host and pathogen remodeling their metabolism upon infection.
| Upon recognition of the host, bacterial pathogens activate a genetic virulence program to establish their replicative niche. In this study, we selected for mutants in the model intracellular pathogen Listeria monocytogenes that did not up-regulate virulence factors during infection. The screen identified genes involved in sensing the host cell and suggests a model in which expression of virulence factors is spatially and temporally compartmentalized via regulation of transcription and translation. Specifically, results described here indicate two non-redundant host cytosolic cues are sensed by the bacterium in order to activate its virulence program. Future research will illuminate the exact molecular identity of these cytosolic signals. However, the majority of the genes identified are part of the bacterial redox stress response, suggesting that redox changes represent one of the biological cues sensed by L. monocytogenes to regulate its virulence program.
| Intracellular pathogens such as Plasmodium spp., Mycobacterium tuberculosis, Salmonella enterica, Trypanosoma cruzi, and Leishmania spp. are responsible for an overwhelming amount of morbidity and mortality worldwide. Successful dissemination of many of these pathogens requires complex life cycles that involve survival and replication in environmental or vector niches. To propagate within their hosts, these pathogens establish a variety of unique intracellular niches that are essential for their pathogenesis [1]. Although there is considerable understanding of how intracellular pathogens manipulate host cell biology to promote their pathogenesis, less is known about the precise mechanisms by which these pathogens sense their host cell. Such an understanding may lead to targets for therapeutic intervention. In this study we used Listeria monocytogenes as a model system for understanding virulence gene regulation of a facultative intracellular bacterium that transitions from extracellular to intracellular growth.
L. monocytogenes is a ubiquitous environmental saprophyte capable of causing severe disease as a foodborne pathogen [2]. L. monocytogenes is also a model system for studying bacterial adaptation to the host [3]. The bacterial virulence program is coordinated with a life cycle that begins upon entry into a mammalian cell either by phagocytosis or bacteria-mediated internalization. To commence intracellular growth, L. monocytogenes must first escape from the hostile phagosomal environment by the expression and secretion of a cholesterol-dependent cytolysin, listeriolysin O (LLO) that mediates destruction of the phagosome [4]. Upon entry into the cytosol, L. monocytogenes grows rapidly and expresses an essential determinant of pathogenesis, ActA, an abundant surface protein that mediates host actin polymerization [5,6]. Appropriate regulation of LLO and ActA is critical for L. monocytogenes pathogenesis and transcriptionally coordinated by the master virulence regulator PrfA [7].
PrfA is a cAMP receptor protein (Crp) family transcriptional regulator that is absolutely essential for L. monocytogenes virulence gene expression and pathogenesis [8]. PrfA-mediated gene expression is regulated by PrfA abundance, affinity for target promoters, and activation via cofactor binding [9]. PrfA levels are controlled by three promoters. The most proximal promoter contains a site of negative regulation, while the most distal is a PrfA-dependent read-through transcript that is essential for appropriately high levels of intracellular gene expression [10–12]. PrfA binds a palindromic DNA sequence (PrfA-box) and deviations from a consensus sequence result in lower affinity DNA-PrfA interactions [13]. The affinity of PrfA for DNA determines the degree of transcriptional activation prior to PrfA allosteric activation [14]. For example, the gene encoding LLO (hly) has a high affinity PrfA-box and consequently is expressed even during growth in broth when PrfA is not activated. In contrast, the actA promoter contains a lower affinity PrfA box and is not expressed during growth in broth [15,16]. Upon entry into the host cell cytosol, PrfA is over-expressed and is activated by a two-step process: first, binding of PrfA to DNA requires reduction of the four PrfA cysteine residues while full transcriptional activation of PrfA requires allosteric binding to glutathione [17]. The requirement for glutathione can be bypassed by mutations that lock PrfA in its active conformation (PrfA*) [18]. Strains with PrfA* mutations constitutively express PrfA-activated genes and consequently have growth defects extracellularly, demonstrating the importance of regulating virulence gene expression [19,20]. However, even PrfA* strains grown in broth fail to synthesize the amount of ActA observed intracellularly, which is likely attributable to translational control localized to the 5’ untranslated region (5’ UTR) [21]. Despite these findings of exquisite gene regulation, little is known about trans-acting factors that affect expression of PrfA or PrfA-activated genes.
In a previous study, a genetic system was designed to select for L. monocytogenes mutants that failed to express ActA intracellularly [17]. This screen led to the identification of L. monocytogenes glutathione synthase (GshF) and glutathione, a tripeptide antioxidant, as the allosteric activator of PrfA. In this study we sought to further understand the host cues that are recognized by intracellular pathogens during infection. We returned to the forward genetic selection and exhaustively screened for additional mutants that failed to express sufficient ActA intracellularly. This selection identified genes required at each stage of the intracellular lifecycle, including: vacuolar escape, PrfA activation, and cell-to-cell spread. These data suggest a model of compartmentalized gene expression, furthering our understanding of the L. monocytogenes virulence program.
The goal of this study was to identify genes involved in regulation of a principle virulence determinant in L. monocytogenes, ActA. A bacterial strain was constructed that failed to replicate upon activation of the actA gene, which is specifically up-regulated during cytosolic growth and is essential for pathogenesis. This ‘suicide’ strain harbored loxP sites in the chromosome flanking the origin of replication (ori) and several essential genes. Codon-optimized cre recombinase was expressed from the actA promoter (Fig 1A). The suicide strain grew like wild type in rich media but was unrecoverable after infection of bone marrow-derived macrophages (BMMs). A himar1 transposon library was then constructed in the suicide strain background and used to infect BMMs. When bacteria were isolated at five hours post-infection (p.i.) nearly all mutants harbored transposon insertions in cre, the actA promoter driving cre expression (actA1p), loxP sites, and gshF, encoding glutathione synthase. To identify additional genes required during infection, colonies were isolated at three and four hours p.i, generating a library of 1,090 transposon mutants from an initial inoculum of >1 million bacteria. Colony PCR excluded strains with transposon insertions in cre and gshF, resulting in a collection of ~700 strains (Fig 1A).
Transposon mutants in the suicide background were screened individually for survival in BMMs, narrowing the list to 300 mutants. Six transposon insertions were identified in hly and nine insertions in prfA, emphasizing that cytosolic access and PrfA are absolutely required for actA activation and subsequent cre expression. Saturation of the screen was further demonstrated after identification of 11 insertions in the actA promoter driving cre and 31 insertions in the loxP sites (which are each only 34 nucleotides). The remaining transposon mutations were transduced into a wild type background and analyzed in a plaque assay, a highly sensitive measure of cell-to-cell spread, which is completely dependent on actA expression [22]. Using a threshold of 85%, 12 mutants were identified that formed plaques significantly smaller than wild type in L2 murine fibroblasts (Fig 2A and Table 1). With one exception, the transposon insertions were in open reading frames and likely resulted in loss-of-function mutations. The transposon in the promoter of lmo2191 (spxA1), a gene predicted to be essential in L. monocytogenes [23], resulted in a 10-fold decrease in spxA1 expression when the bacteria were grown in broth, essentially resulting in a knock-down strain (S1 Fig). Attempts to make an in-frame deletion of spxA1 using conventional methods were unsuccessful, consistent with a previous report [23].
As the goal of this selection was to identify mutations that affect ActA expression in vivo, we measured ActA abundance during infection of BMMs. Four hours post-infection, cells were lysed and ActA and the constitutively expressed P60 protein were analyzed by immunoblot. Nine strains were found to express less ActA than wild type after normalizing to P60 abundance (Fig 2B). The work-flow of this selection used cre expression from the actA promoter and plaque area as a criterion for inclusion in the core set of twelve mutants analyzed here. It was therefore unexpected that three mutants (lmo0441::Tn, lmo0443::Tn, and citC::Tn) did not display a defect in ActA abundance during intracellular growth. We hypothesize that these mutations may disrupt elements of bacterial physiology critical to appropriate Cre activity or normal growth.
The twelve mutants isolated by the genetic selection were identified based on in vitro assays for virulence. While these assays are correlated to in vivo outcomes, the importance of these genes to L. monocytogenes pathogenesis was confirmed in a murine model of infection. Intravenous infection of mice revealed that four of the mutants displayed no virulence defect (lmo0441::Tn, rsbX::Tn, lmo2107::Tn, and gtcA::Tn) while the remaining eight mutants were significantly attenuated (Fig 2C). It was surprising that four mutants exhibited impaired plaque formation yet were fully virulent; it is possible that these four mutants are impaired in other aspects of pathogenesis not reflected by changes in CFU during these infection conditions. To determine if the plaque defects in these mutants were due to cell-specific defects evident only in the L2 murine fibroblasts used for plaque assays, cell-to-cell spread defects were also analyzed in TIB-73 cells, a murine hepatocyte cell line (Table 1). We observed consistent phenotypes between the plaque defects in TIB-73 cells and L2 cells with the exception of citC::Tn, P-spxA1::Tn, and ohrA::Tn. However, these mutants were significantly attenuated during infection and thus it was unclear why they did not display a plaque defect in TIB-73 cells.
The specificity of the transposon insertion in seven of the eight attenuated strains was confirmed by expressing the disrupted gene in trans and complementing the plaque defect (S2 Fig). Attempts to complement the pplA::Tn plaque defect were unsuccessful. However, pplA mutants are difficult to complement and the mutant we identified exhibited phenotypes consistent with published ΔpplA defects [24]. Other reports have identified genes necessary for virulence of L. monocytogenes by comparing changes in gene expression in vivo [25–27]. In our analysis, only gshF was differentially transcribed between host cells and rich media (Fig 2D). It remains to be investigated if the activity of these genes is regulated post-transcriptionally in response to the host.
In this study we focused on the following genes that were required for actA expression and pathogenesis (Fig 1B). yjbH (lmo0964) encodes a putative thioredoxin similar to YjbH in Bacillus subtilis (57% amino acid similarity) [28]. A transposon in L. monocytogenes yjbH was previously identified in a screen for mutants defective in LLO production in vitro and was found to be attenuated in a competitive infection model [29]. spxA1 (lmo2191) encodes an ArsC family transcriptional regulator similar to the disulfide stress regulator Spx conserved in Firmicutes (83% amino acid identity to B. subtilis Spx) [30]. The difference in nomenclature is due to the presence of a paralogous gene in L. monocytogenes (lmo2426 or spxA2) that is 59% identical to B. subtilis Spx while B. subtilis encodes only a single spx. In B. subtilis and Staphylococcus aureus YjbH post-translationally regulates Spx [28,31], although it is not known if this function is conserved in L. monocytogenes. lmo2199 encodes a hypothetical protein with a peroxiredoxin domain and is part of the organic hydroperoxide resistance (Ohr) protein subfamily. It is co-transcribed with lmo2200, encoding a MarR family transcriptional regulator which was not required for virulence, suggesting that Lmo2200 may act as a transcriptional repressor [26]. In B. subtilis homologs of Lmo2199 and Lmo2200 are named OhrA (63% amino acid similarity) and OhrR (68%), respectively, and we have adopted this nomenclature for consistency [32]. arpJ (lmo2250) encodes a predicted amino acid ABC transporter permease that was originally identified in a screen for genes with increased intracellular expression [25]. However, the data presented here did not show an increase in arpJ expression during infection of BMMs. This may be explained by the different growth media and cell types used in the two studies. It is also possible that arpJ is autoregulated, as the previous study analyzed arpJ expression in an arpJ transposon mutant. pplA (lmo2637) encodes a lipoprotein whose secretion is increased in a PrfA* mutant [33]. The signal sequence of this lipoprotein is processed into a secreted peptide, which is required for vacuolar escape from non-phagocytic cells [24]. Finally, gshF (lmo2770) encodes the only glutathione synthase in L. monocytogenes [34]. Glutathione has been demonstrated to be an allosteric activator of PrfA and therefore gshF mutants are severely attenuated in vivo due to insufficient virulence gene expression [17].
Given the role of glutathione in activating PrfA, we hypothesized that suppressor mutations of ΔgshF might illuminate alternative pathways for PrfA activation, potentially involving other genes identified. Accordingly, we screened for mutations that increased the virulence of a ΔgshF mutant. Mice were serially infected with a high-inoculum of ΔgshF, livers were harvested at 72 hours p.i., homogenized, and diluted to inoculate naive mice. After four successive infections bacteria isolated from infected livers were analyzed by plaque assay. This approach previously identified a mutation in prfA that constitutively activates the protein (G145S), known as PrfA*, completely bypassing the requirement for glutathione during infection [17]. The ΔgshF PrfA* suppressor forms 100% plaque; therefore, for these experiments we selected bacteria that formed intermediate-sized plaques, which were then subjected to genome sequencing. Two suppressor mutants were isolated and found to encode a G>A mutation 58 nucleotides 5’ of the prfA start codon (Fig 3A). This mutation lies within a previously identified site of negative regulation of prfA, the so-called “P2 promoter” (prfA2p, Fig 3A) and deletion of the -35 region of this promoter (ΔP2 mutant) results in a 10-20-fold up-regulation of the prfA1p-dependent prfA transcript [11]. We hypothesized that the prfA -58 G>A mutation also inactivated the P2 promoter and resulted in greater PrfA abundance. Indeed, the ΔP2 gshF::Tn double mutant and the ΔgshF prfA -58 G>A suppressor mutants all formed plaques approximately 60% the size of wild type (Fig 3B). These results did not directly implicate any of the other genes identified in our genetic selection, however these findings did highlight the impact of both PrfA abundance and activation during infection.
PrfA expression is controlled by a feed-forward loop in which activated PrfA drives its own transcription [12]. Strains expressing ΔP2 or PrfA* decouple PrfA abundance and activation whereby ΔP2 increases PrfA abundance but still relies on glutathione for PrfA activation; PrfA* increases both the amount and activity of PrfA, independent of glutathione. We next sought to determine if the other mutants identified in the screen affected PrfA abundance or activation by transducing each into L. monocytogenes ΔP2 and PrfA* backgrounds and measuring the plaque size in each background (Fig 3C). Based on these analyses, mutants fell into three categories. The first category (yjbH::Tn, P-spxA1::Tn, ohrA::Tn, and arpJ::Tn) was unaffected by alterations in PrfA expression or activity, indicating that these genes were required down-stream of PrfA. In the second category was gshF::Tn, which was partially rescued by ΔP2 and completely rescued by PrfA*, consistent with the demonstrated role for glutathione as the allosteric activator of PrfA. The third category describes pplA::Tn, which formed 100% plaques in both the ΔP2 and PrfA* backgrounds. These data suggested that the pplA mutant was capable of activating PrfA (because it was rescued by ΔP2) but was deficient in expression of PrfA-dependent genes required early during infection before cytosolic access and glutathione-mediated activation of PrfA.
A principle difference between early and late PrfA-dependent genes is that expression of early genes are less dependent on PrfA activation by glutathione [35]. The two early genes are hly (encoding LLO) and plcA, which share a high-affinity PrfA-box and are transcribed by unactivated PrfA [35,36]. The ΔP2 mutation results in increased transcription of early genes but does not affect late gene expression, whereas PrfA* increases transcription of both early and late genes. We hypothesized that strains rescued by ΔP2 are specifically deficient in early gene expression. Accordingly, we analyzed early gene expression (LLO production) in broth for each mutant. Several of the mutants were found to secrete less LLO than wild type (Fig 4A). To determine if the defect in LLO production led to impaired phagosomal escape and thus a plaque defect, these mutations were transduced into a Δhly mutant over-expressing hly from a constitutive HyPer promoter (pH-hly strain) [37,38]. In this background, efficiency of vacuolar escape should be equivalent in all strains, and indeed, equal LLO secretion was confirmed in broth. Constitutive expression of hly rescued the plaque defects of three mutants: P-spxA1::Tn, ohrA::Tn, and pplA::Tn (Fig 4B). Interestingly, there was discordance between LLO production in broth and the defect in plaque formation one might predict from an LLO deficiency. For this reason, measuring LLO production in broth may be revealing aspects of bacterial physiology unrelated to LLO production in vivo.
The above results suggested that mutations in P-spxA1, ohrA, and pplA resulted in aberrant LLO secretion and/or that these mutants were unable to survive in the harsh environment of the vacuole. Constitutive expression of hly would likely overcome either defect. We attempted to segregate these two possibilities by analyzing sensitivity to vacuolar conditions, including reactive oxygen species which L. monocytogenes must adapt to in order to survive [39,40]. The response of each mutant to peroxide, disulfide stress, and organic hydroperoxide was analyzed by measuring their sensitivity to hydrogen peroxide (H2O2), diamide, and cumene hydroperoxide (CHP), respectively. Knock-down of spxA1 and disruption of ohrA or gshF significantly increased the sensitivity of L. monocytogenes to both peroxide and disulfide stress (Fig 4C). In accordance with its annotation and the published role of ohrA in B. subtilis [32], the ohrA::Tn mutant was significantly more susceptible to CHP (Fig 4C). As these results suggested a role for redox control of virulence genes, we tested the hypothesis that host reactive oxygen or nitrogen species may be sensed by the bacteria during infection to activate actA. However, growth of the suicide mutant was not rescued in BMMs lacking inducible nitric oxide synthase (NOS2-/-) or NADPH oxidase (NOX2-/-) (S3 Fig). Therefore, L. monocytogenes may activate virulence genes in response to multiple redundant host cues or depend on yet unidentified host pathways.
Constitutive production of hly restored the majority of the plaque defect for P-spxA1::Tn and ohrA::Tn, however, it did not restore the plaque to 100% of the parent strain (Fig 4B). We hypothesized that these mutants might also be impaired in the ability to grow in the host cytosol, independently from virulence gene expression. All of the mutants identified in the screen grew similarly to wild type in BMMs with the exception of P-spxA1::Tn and ohrA::Tn (Fig 4D). In fact, P-spxA1::Tn and ohrA::Tn were also the only mutants that exhibited growth defects in rich media (Fig 4E). These pleiotropic growth defects and sensitivity to redox stress are likely why pH-hly was only partially able to complement the plaque defect of these mutants (Fig 4B).
Previous work clearly demonstrated that glutathione was essential for transcriptional activation of virulence genes [17]. In order to assess which factors might be independent of glutathione-dependent transcriptional activation, we combined each transposon with an in-frame ΔgshF mutation. The only mutation not epistatic to gshF was yjbH::Tn, which produced an additive plaque defect (Fig 5A). Further, yjbH::Tn was not rescued by constitutive activation of hly (Fig 4B) or prfA (Fig 3C). Together, these data suggested that yjbH was required for actA expression post-transcriptionally. Indeed, transcript levels of actA were identical in BMMs infected with wild type or the ΔyjbH mutant (Fig 5B). It is intriguing that arpJ::Tn was epistatic to gshF, yet not rescued by constitutive activation of PrfA, indicating that arpJ may contribute to glutathione-dependent transcriptional activation of actA through an unknown mechanism.
The actA gene is preceded by 149 nucleotides of untranslated mRNA (Fig 5C) which is important for sufficient ActA expression [21]. A strain was constructed in which ActA was expressed independent of PrfA by expressing the entire actA transcript (including the 5’ UTR) under the control of the constitutive HyPer promoter in a strain deleted for actA (pH-actA Strain, Fig 5D). ActA protein abundance was then analyzed by immunoblot. In this background, ActA abundance was equivalent among all strains when the bacteria were grown in broth (Fig 5E). However, during infection of BMMs, disruption of yjbH resulted in significant impairment in ActA abundance (Fig 5F), indicating a failure to translationally activate actA. Given that disrupting yjbH rescued the death of the suicide strain in which cre was expressed under actA1p and the 5' UTR, these data indicate a genetic interaction between yjbH and the 5’ UTR of actA. To further support this genetic interaction we engineered a fluorescent strain of L. monocytogenes in which rfp was expressed under the actA1p promoter and 5' UTR (actA1p-rfp, Fig 5G). During infection of BMMs the ΔyjbH actA1p-rfp strain exhibited significantly less fluorescence than wild type actA1p-rfp (Fig 5H). Unfortunately, we were unable to interrogate the effect of a yjbH mutation on ActA abundance in the absence of its 5’ UTR due to an inability to detect ActA when the 5’ UTR was deleted, consistent with this region being critical for ActA expression [21].
A drawback to pH-actA is that although ActA is over-expressed in broth, this strain still elaborates much less ActA in vivo and fails to form a plaque (Fig 5E and 5F). To analyze the role of translational activation during infection, the actA gene and 5’ UTR were moved to a neutral locus within the L. monocytogenes chromosome [41]. In this strain, actA was expressed only from the PrfA-dependent actA1p proximal promoter, eliminating read-through transcription from the distal actA2p promoter (Fig 5C). This strain was called actA1p and was only mildly impaired in plaque formation and virulence (Fig 5I and 5J). However, actA1p yjbH::Tn was unable to form a plaque (Fig 5I). The importance of actA translational activation was further underscored by a 3-log defect for actA1p yjbH::Tn in the livers of infected mice (Fig 5J). These data revealed a critical role for yjbH in actA activation that was less apparent in the wild type background due to redundant PrfA-dependent promoters.
In this study, rather than search for novel virulence factors or genes up-regulated in vivo, we screened for genes required for activation of an essential determinant of L. monocytogenes pathogenesis (ActA) that is up-regulated over 200-fold during intracellular growth. Mutants identified in the genetic selection fell into three broad categories: (1) those that failed to reach the cytosolic compartment; (2) mutants that entered the cytosol, but failed to activate the master virulence transcriptional regulator PrfA; and (3) mutants that entered the cytosol and activated transcription of actA, but failed to synthesize it (Fig 6). This approach highlighted how expression of virulence factors is spatially and temporally compartmentalized via regulation of transcription and translation during infection. One of the most striking findings of this study was that the majority of genes identified in the selection encode proteins predicted to control bacterial redox regulation, suggesting that redox changes represent one of the biological cues sensed by L. monocytogenes to regulate its virulence program. Redox stress during infection can arise from endogenous by-products of bacterial metabolism and exogenously derived factors generated by the host. However, it remains to be discovered whether the redox stress that may trigger virulence gene expression is produced by the host, the bacteria, or both.
YjbH, Spx, OhrA, and GshF have defined roles in maintaining redox homeostasis in the presence of disulfide and organic peroxide stresses in Firmicutes. In B. subtilis OhrA is a peroxiredoxin required during organic hydroperoxide stress [32]. In S. aureus and B. subtilis YjbH interacts with Spx to regulate the abundance and activity of Spx [28,31]. Specifically, YjbH-bound Spx is recognized by the ClpXP protease and is degraded so that Spx concentrations are low under steady-state conditions [42,43]. During disulfide stress the YjbH:Spx interaction is disrupted by intramolecular disulfide bonds in both proteins that result in reduced proteolysis of Spx. B. subtilis Spx represses transcription of 176 genes and activates transcription of 106 genes [44], the majority of which are required to adapt to redox stress, including genes for production of the low-molecular weight (LMW) thiol utilized by B. subtilis, bacillithiol [45]. L. monocytogenes spxA1 cannot be deleted and its regulon has not yet been characterized [23]. Similarly, in Streptococcus pneumoniae simultaneous deletion of both spxA1 and spxA2 paralogues is lethal [46], supporting the notion that the Spx regulon(s) may contain essential genes in some Firmicutes.
Mutants exhibiting the most severe virulence phenotypes contained insertions in gshF, which encodes the sole L. monocytogenes glutathione synthase [34]. Glutathione is a tripeptide LMW thiol antioxidant present at millimolar concentrations that contributes to maintaining a reducing environment in both bacterial and host cells [47]. Not surprisingly, L. monocytogenes ΔgshF mutants are more sensitive to redox stressors such as hydrogen peroxide and diamide and are 200-fold less virulent in mice, indicating that bacterially-derived glutathione is essential for pathogenesis [17]. However, ΔgshF mutants are fully virulent in L. monocytogenes harboring prfA* mutations that lock PrfA in its constitutively active conformation. Therefore, the primary role of GshF-derived glutathione during infection is to activate virulence gene expression via PrfA activation, although we cannot rule out a contribution of imported host-derived glutathione [17]. Indeed, host-derived glutathione activates virulence gene expression in Burkholderia pseudomallei [48]. In the case of L. monocytogenes, gshF is transcriptionally up-regulated 10-fold during intracellular growth, suggesting the existence of an unidentified cue, likely redox-related, that stimulates glutathione production.
The identification of many redox-related bacterial factors in this genetic selection led to our working model that specific redox changes during infection are sensed by the bacteria as a mechanism to identify their intracellular location and activate virulence genes appropriately. Redox stress during infection could arise from host-derived antimicrobial factors. For example, the host generates antibacterial factors that assault invading pathogens with redox stresses, including: reactive oxygen species (ROS), reactive electrophilic species (RES) such as methylglyoxal, and reactive nitrogen species (RNS) such as nitric oxide and peroxynitrite [40,49,50]. Interestingly, these redox stresses from the host are spatially compartmentalized. RNS and ROS are produced in the phagosome and once in the host cytosol, L. monocytogenes is confronted with RES and mitochondrial-derived ROS [40,51]. It is possible that the bacterial response to the redox stressors is also compartmentalized, requiring specific factors in the vacuole (such as spxA1 and ohrA) and host cytosol (such as yjbH).
Eliminating host nitric oxide synthase (NOS2) or NADPH oxidase did not rescue growth of the suicide mutant (S3 Fig). NOS2-generated nitric oxide is required for efficient L. monocytogenes cell-to-cell spread during infection, although this is due to the nitric oxide-mediated delay of phagolysosome maturation and not a direct effect on the bacteria [52]. Together, these data suggest that a combination of host factors are likely required to activate actA during infection.
Alternatively, the source of redox stress may come from bacterial metabolism via ROS generated from incomplete reduction of oxygen during aerobic respiration [53]. Carbon source and phosphate abundance also affect the production of ROS and methylglyoxal [54,55]. PrfA activity has been demonstrated to be sensitive to available carbon sources [2]. Growth on plant-derived beta-glucoside sugars in the environment, such as cellobiose, represses PrfA activation, whereas growth on host-derived sugars such as glucose-1-phosphate stimulates PrfA-dependent gene expression [9,56,57]. Therefore, entry of L. monocytogenes into the host cytosol results in a remodeling of carbon metabolism that may be linked to virulence gene regulation. Glycerol is the principle carbon source used by L. monocytogenes intracellularly and growth on glycerol is a well-described stimulant of methylglyoxal production [58–61]. In B. subtilis, methylglyoxal stress stimulates the Spx regulon and production of bacillithiol, a low molecular weight thiol used by B. subtilis to detoxify methylglyoxal [62]. Thus, the 10-fold increase in gshF transcript levels in L. monocytogenes may correspond to increased methylglyoxal production during infection, which would further link metabolism of an alternative carbon source to virulence. Coupling of metabolism to virulence gene regulation may allow the system to remain OFF in the environment while remaining poised to turn ON upon entering a host. Considering our finding of multiple redox factors that are required for proper virulence gene expression, we speculate that changes in carbon metabolism could alter the endogenous levels of ROS and RES produced, thus affecting PrfA activation and leading to the “sugar-mediated repression” observed previously [9].
Appropriate up-regulation of actA at the translational level is understood to require its 5’ UTR, although the mechanism remains unknown [21]. The data reported here further emphasize the sensitivity of actA translation to the environment in which L. monocytogenes is growing. In broth, the PrfA* strain elaborated 2.4% the amount of ActA protein as compared to constitutively expressed actA (Fig 5E), and increased 200-fold during infection (Fig 5F), despite the fact that transcript levels of actA are equivalent in both growth conditions [17]. These data emphasize the importance of the translational control of this virulence factor. Importantly, yjbH was required for the increased abundance of ActA protein during infection. In wild type L. monocytogenes, multiple PrfA-dependent promoters may compensate for loss of translational activation; however, when actA was isolated under its most proximal promoter, disruption of yjbH resulted in an attenuation of over 3-logs in the livers of infected animals (Fig 5J). It seems unlikely that the thioredoxin YjbH activates translation of actA via direct binding to the 5’ UTR. However, YjbH may indirectly activate translation via interaction with another factor(s) or modulation of a small-molecule signal produced by the host.
PrfA-dependent transcription and activation are regulated redundantly at multiple levels, including: a temperature-sensitive riboswitch [63], allosteric activation by glutathione [17], multiple read-through transcripts [10,64], positive and negative promoter elements [11,65], and yet to be fully characterized translational control. The complexity of actA activation is likely the result of selective pressure to respond appropriately to host-derived cues. This study investigated the virulence defects associated with failure to up-regulate virulence genes; however, over-production or inappropriate regulation of virulence factors extracellularly also results in a competitive disadvantage for L. monocytogenes [19,20]. How L. monocytogenes and other intracellular pathogens regulate virulence gene expression is central to understanding their pathogenesis. Results reported here suggest that redox cues are a mechanism by which intracellular pathogens recognize the host and represents an exciting new area of further investigation.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All protocols were reviewed and approved by the Animal Care and Use Committee at the University of California, Berkeley (AUP-2016-05-8811).
All L. monocytogenes strains are a derivative of wild type 10403S [67,68] and were cultivated in Brain Heart Infusion (BHI, Difco), shaking at 37°C unless otherwise stated. All E. coli strains were cultivated shaking in LB (Miller) at 37°C. Antibiotics (purchased from Sigma) were used at the following concentrations: carbenicillin (100 μg/mL), streptomycin (200 μg/mL), chloramphenicol (7.5 μg/mL for L. monocytogenes and 10 μg/mL for E. coli), erythromycin (1 μg/mL), and tetracycline (2 μg/mL). All E. coli strains are listed in Table 2 and all L. monocytogenes strains are listed in Table 3. Bacterial broth growth curves were performed as previously described [69]. The suicide strain was a gift from Peter Lauer and Bill Hanson (Aduro Biotech); details of its construction are reported elsewhere [17]. Briefly, loxP sites were inserted on either side of the origin of replication by allelic exchange into a ΔactAΔinlB strain of L monocytogenes. A transcriptional fusion of cre with actA that included the actA1p promoter, 5’ UTR, and ribosomal binding site of actA, was inserted adjacent to a loxP site.
Knock-in of pPL2 derivative plasmids was performed by standard methods [41]. Briefly, constructed pPL2 plasmids were transformed into chemically competent SM10 E. coli, selecting on chloramphenicol. Donor SM10 and recipient L. monocytogenes were mixed at a 1:1 ratio on a non-selective BHI plate at 37°C for 4–24 hours, then trans-conjugation was selected for by plating bacteria on BHI containing streptomycin plus either chloramphenicol (pPL2), erythromycin (pPL2e), or tetracycline (pPL2t). Single colonies were re-streaked for purifying selection onto BHI containing the same antibiotics as used after trans-conjugation.
In-frame deletions of genes was accomplished by allelic exchange using pKSV7-oriT and conventional methods [64]. Briefly, the constructed knock-out plasmid was transformed into SM10 E. coli, recovered on LB containing carbenicillin, and trans-conjugated into L. monocytogenes by mixing the donor SM10 and recipient L. monocytogenes at a 1:1 ratio on a non-selective BHI plate for 4–24 hours at 30°C, the permissive temperature for pKSV7-oriT to replicate in Gram-positive organisms. Trans-conjugation was selected on BHI containing both streptomycin and chloramphenicol at 30°C. After isolation of a single colony of L. monocytogenes containing pKSV7-oriT at 30°C, bacteria were grown at 42°C on BHI agar containing both streptomycin and chloramphenicol to select for chromosomal integration. Colonies were re-streaked onto selective media at 42°C two additional times for purifying selection and integrated pKSV7-oriT. This strain was then serially passaged at 30°C to enrich for excision and loss of pKSV7-oriT. Mutants that lost pKSV7-oriT were identified by sensitivity to chloramphenicol using indirect patch-plating methods. Finally, allelic exchange was confirmed by PCR and, when necessary, Sanger DNA sequencing.
Preparation of electro-competent L. monocytogenes and himar1 transposon mutagenesis were performed as previously described [29], generating a transposon mutant library that was not fully characterized previously [17]. Transposon junctions were mapped as previously described [71]. The position of each himar1 transposon refers to to the distance of the insertion site, 3’ of the first nucleotide of each gene. Transposons were mapped to the 10403S genome, however, for continuity of nomenclature the EGD-e loci names have been used. For reference: lmo0441 (LMRG_00133), lmo0443 (LMRG_00135), rsbX is lmo0896 (LMRG_02320), yjbH is lmo0964 (LMRG_02063), citC is lmo1566 (LMRG_01401), lmo2107 is (LMRG_01261), spxA1 is lmo2191 (LMRG_01641), ohrA is lmo2199 (LMRG_01633), arpJ is lmo2250 (LMRG_01581), gtcA is lmo2549 (LMRG_01698), pplA is lmo2637 (LMRG_02182), gshF is lmo2770 (LMRG_01925).
Transposons in the chromosome were introduced into different genetic backgrounds by generalized transduction using the phage U153, as previously described [29,75]. Briefly, a transducing lysate was generated by lysogenizing approximately 109 CFU of L. monocytogenes transposon donor with approximately 107 PFU of phage in 3–4 mL of 0.7% LB Agar containing MgSO4 and CaCl2 (10 mM each) on LB agar and incubated overnight at 30°C. Phage was soaked out of the agar by incubating with 5 mL of TM buffer (10 mM Tris, pH 7.5 and 10 mM MgSO4) for 8–24 hours and these recovered phage stocks were filter sterilized. With the newly generated transducing lysate, 108 L. monocytogenes recipients were lysogenized with 107 PFU of lysate, incubated at 30°C for 30 min in LB containing MgSO4 and CaCl2 (10 mM each), and then plated on selective BHI agar at 37°C. When transducing the himar1 transposon using erythromycin selection, colonies appeared after two days. These colonies were purified by re-streaking transductants for single colonies and verified by sequencing the transposon junction. U153 phage stocks were propagated using L. monocytogenes strain SLCC-5762.
Knock-in plasmids were constructed as previously described using primers listed in Table 4 and reagents are from New England Biolabs, unless otherwise specified [71]. Briefly, vectors for complementing yjbH and spxA1 were constructed by amplifying each gene along with its predicted native promoters using a reverse primer that appended a DNA sequence encoding a six histidine affinity tag at the C-terminus. These DNA fragments and pPL2 [41] were then digested with KpnI and BamHI and ligated using Quick Ligase, according to manufacturer’s instructions. The arpJ and ohrA complement vectors were constructed by amplifying their entire predicted operon and predicted native promoter (arpJ: LMRG_01581-LMRG_01580, ohrA: LMRG_01632-LMRG_01633) without addition of affinity tags. The DNA fragment was combined with linearized pPL2t harboring a transcriptional terminator [71] and assembled using In-Fusion Cloning (Clontech) or Gibson Assembly Ultra (Synthetic Genomics). The pPL2t.Phyper-actA vector was constructed by amplifying both 5’ UTR and CDS of actA (LMRG_02626), and combining the DNA fragment with linearized pPL2t harboring a modified Pspac-hy (Phyper) [38] sequence: “aattgtgagcgctcacaattttgcaaaaagttgttgactttatctacaaggtgtggcataatgtgtGTAATTGTGAGCGCTCACAATT”, inserted via gBLOCK (IDT), and a transcriptional terminator for assembly using In-Fusion Cloning (Clontech).
The pKSV7-oriT-ΔyjbH vector was constructed according to methods previously described [71]. Briefly, the vector was constructed by sequentially amplifying ~1 kb of homology flanking the yjbH coding region using primers in Table 3. These two fragments were joined by sequence overlap extension PCR, which included the coding region for the first six and last six amino acids of YjbH. The final PCR fragment and pKSV7-oriT were digested with KpnI and PstI (rSAP was also included for the vector) and ligated using Quick Ligase. The ligation product was transformed into XL1 Blue E. coli and transformants were screened by PCR for the presence of the insert, followed by Sanger sequencing confirmation.
The plaque assay was carried out by conventional methods [22,76]. Briefly, L2 fibroblasts (generated previously from L929 cells [77] and provided as a generous gift from Susan Weiss in 1988, as detailed in Sun et al. [22]) or TIB-73 hepatocytes (ATCC TIB-73) were maintained in high-glucose DMEM medium plus 10% FBS (Hyclone), 2 mM L-glutamine (Gibco), and 1 mM sodium pyruvate (Gibco). Cells were plated at 1.2 x 106 cells per well in a six-well dish and infected the next day at an MOI of 300 with L. monocytogenes grown overnight at 30°C, stationary. The infection was allowed to proceed for one hour before the wells were washed twice with PBS and 3 mL of medium plus 0.7% agarose and 10 μg/mL gentamicin was overlaid. At 48 hours post-infection the plaques were stained with 2 mL of medium plus 0.7% agarose, 10 μg/mL gentamicin, and 25 μL/mL neutral red (Sigma). The plaques were then imaged at 72 hours post-infection. Plaque area was quantified using ImageJ software [78]. Each experiment represented an average of the area of at least five plaques per strain as a proportion to wild type plaques in that experiment. Data are representative of at least three experiments.
Macrophage growth curves were performed as previously described [72,79]. Briefly, bone marrow-derived macrophages (BMMs) were derived from bone marrow of C57BL/6 mice purchased from The Jackson Laboratory and were cultivated/differentiated in high-glucose DMEM medium containing CSF (from mouse CSF-1-producing 3T3 cells), 20% FBS (Hyclone), 2 mM L-glutamine (Gibco), 1 mM sodium pyruvate (Gibco), and 14 mM 2-mercaptoethanol (BME, Gibco). BMMs were derived as previously described and plated in 60 mm non-TC treated dishes that contained 14 TC-treated coverslips at 3 x 106 cells per dish. These dishes were then infected at an MOI of 0.1 for 30 minutes, washed twice with PBS prior to replacing media, and gentamicin was added at 50 μg/mL one hour post-infection. Three coverslips were removed from each dish at 0.5, 2, 5, and 8 hours post-infection and added to 5 mL of sterile water. Coverslips were rigorously mixed prior to plating on LB agar. Each graph is representative of three experiments and each data point represents the average of three coverslips.
To analyze virulence, female CD-1 mice were infected intravenously (i.v.) via the tail vein using 200 μL of sterile PBS containing 105 CFU of each L. monocytogenes strain as previously described [80]. The infection was allowed to progress for 48 hours, at which point animals were euthanized and the spleens and livers were harvested. Organs were homogenized in 0.1% NP-40 and serial dilutions were plated on LB agar containing streptomycin. Graphs represent pooled data from at least two experiments of greater than four mice each. Groups were statistically compared using a heteroscedastic Student’s t-test.
In vivo suppressors were identified similarly to previously described methods [17]. Briefly, CD-1 mice were infected i.v. with 1 x 107 CFU of ΔgshF for 72 hours and the livers were harvested, homogenized, and 100 μL was inoculated into broth. Naïve mice were then infected with these liver homogenate cultures. After four successive infections bacteria isolated from infected livers were analyzed via plaque assay and two strains with intermediate plaque phenotype were selected for genome sequencing.
Genomic DNA was isolated from L. monocytogenes using the MasterPure Gram-Positive DNA Purification Kit (Epicentre) according to the manufacturer's instructions. Genome sequencing and DNA library preparation was performed as previously described [71] at the Vincent J. Coates Genomics Sequencing Laboratory at UC Berkeley. Data was assembled and aligned to the 10403S reference genome (GenBank: GCA_000168695.2) demonstrating >50x coverage. SNP/InDel/structural variation was determined as compared to the ΔgshF parent strain using CLC Genomics Workbench (CLC bio).
All immunoblotting was performed as previously described [17]. Briefly, for bacteria grown in broth, overnight cultures were diluted 1:10 into BHI, incubated for five hours at 37°C, shaking, then the bacteria were separated from the supernatant by centrifugation. For secreted proteins, the supernatant was treated with 10% v/v TCA for one hour on ice to precipitate all proteins. The protein pellet was washed twice with ice- cold acetone, followed by vacuum drying. The proteins were dissolved in LDS buffer (Invitrogen) containing 5% BME using a volume that normalized for OD600 of harvested bacteria, boiled for 20 minutes, and separated by SDS-PAGE. For surface associated proteins, bacteria were suspended in 150 μL of LDS buffer containing 5% BME, boiled for 20 minutes, and proteins separated by SDS-PAGE.
Immunoblots of bacteria grown intracellularly within infected BMMs used 12-well dishes with BMMs at a density of 106 cells per well and infected with an MOI of 10. One hour post-infection the cells were washed and media containing gentamicin (50 μg/mL) was added. Four hours post-infection the cells were washed twice with PBS and harvested in 150 μL LDS buffer containing 5% BME. The samples were then boiled and separated by SDS-PAGE. Primary antibodies were each used at a dilution of 1:5,000, including: rabbit polyclonal antibody against the N-terminus of ActA [81], rabbit polyclonal antibody against LLO, and a mouse monoclonal antibody against P60 (Adipogen). P60 is a constitutively expressed bacterial protein used as a loading control [82]. All immunoblots were visualized and quantified using Odyssey Imager and appropriate secondary antibodies from the manufacturer according the manufacturer’s instructions.
Transcript analysis in broth was performed as previously described [83]. Briefly, bacteria were grown overnight in BHI and subcultured 1:100 into 25 mL BHI. Bacteria were harvested at an OD600 = 1.0. Transcript analysis during infection was analyzed as previously described [17]. Briefly, BMMs were plated at a density of 3 x 107 cells in 150 mm TC-treated dishes and infected with an MOI of 10. One hour post-infection the cells were washed and media containing gentamicin (50 μg/mL) was added. Four hours post-infection the cells were washed with PBS and lysed in 5 mL of 0.1% NP-40. After collecting the lysate, the dishes were then washed in RNAprotect Bacteria Reagent (Qiagen), which was combined with the lysate. Bacteria were isolated by centrifugation. Bacteria harvested from either broth or BMMs were lysed in phenol:chloroform containing 1% SDS by vortexing with 0.1 mm diameter silica/zirconium beads (BioSpec Products Inc.). Nucleic acids were precipitated from the aqueous fraction overnight at -80°C in ethanol containing 150 mM sodium acetate (pH 5.2). Precipitated nucleic acids were washed with ethanol and treated with TURBO DNase per manufacturer’s specifications (Life Technologies Corporation). RNA was again precipitated overnight and then washed in ethanol. RT-PCR was performed with iScript Reverse Transcriptase (Bio-Rad) and quantitative PCR (qPCR) of resulting cDNA was performed with KAPA SYBR Fast (Kapa Biosystems). Primers used for qPCR are listed in Table 4.
Disk diffusions were performed similarly to previously described methods [84]. Briefly, approximately 3 x 107 CFU from overnight cultures of bacteria were immobilized using 4 mL of molten (55°C) top-agar (0.8% NaCl and 0.8% bacto-agar) spread evenly on tryptic soy agar plates. After the agar cooled, Whatman paper disks soaked in 5 μL of 5% hydrogen peroxide, 1 M diamide solution, or 80% cumene hydroperoxide solution were placed on top of the bacteria-agar. The zone of inhibition was measured after 18–20 hours of incubation at 37°C.
BMMs were differentiated and cultivated as described for BMM growth curves. Cells were plated at 5 x 105 cells per well in a 24-well dish in media without antibiotics. The following day BMMs were infected at an MOI of 5 with L. monocytogenes mutants that had been incubated at 30°C without shaking. After 30 minutes cells were washed once with PBS and fresh media containing gentamicin (50 μg/mL) was added. Six hours post infection media was removed from each well, the cells were washed with 1 mL of PBS, and 0.5 mL of PBS was replaced for each well. RFP fluorescence was measured using a plate reader (Infinite M1000 PRO, TECAN) with 555 nm excitation, 584 nm emission, and 5 nm band filters. Each well was interrogated 64 times on an 8 X 8 grid and the edge reads were excluded. Data were normalized by subtracting baseline fluorescence of wild type (without RFP) infected cells and plotting data as a percentage of wild type expressing actA1p-rfp. Each experiment represents three infected wells per L. monocytogenes genotype and data are representative of three pooled independent experiments.
|
10.1371/journal.pcbi.1004293 | Causal Modeling of Cancer-Stromal Communication Identifies PAPPA as a Novel Stroma-Secreted Factor Activating NFκB Signaling in Hepatocellular Carcinoma | Inter-cellular communication with stromal cells is vital for cancer cells. Molecules involved in the communication are potential drug targets. To identify them systematically, we applied a systems level analysis that combined reverse network engineering with causal effect estimation. Using only observational transcriptome profiles we searched for paracrine factors sending messages from activated hepatic stellate cells (HSC) to hepatocellular carcinoma (HCC) cells. We condensed these messages to predict ten proteins that, acting in concert, cause the majority of the gene expression changes observed in HCC cells. Among the 10 paracrine factors were both known and unknown cancer promoting stromal factors, the former including Placental Growth Factor (PGF) and Periostin (POSTN), while Pregnancy-Associated Plasma Protein A (PAPPA) was among the latter. Further support for the predicted effect of PAPPA on HCC cells came from both in vitro studies that showed PAPPA to contribute to the activation of NFκB signaling, and clinical data, which linked higher expression levels of PAPPA to advanced stage HCC. In summary, this study demonstrates the potential of causal modeling in combination with a condensation step borrowed from gene set analysis [Model-based Gene Set Analysis (MGSA)] in the identification of stromal signaling molecules influencing the cancer phenotype.
| All living cells rely on communication with other cells to ensure their function and survival. Molecular signals are sent among cells of the same cell type and from cells of one cell type to another. In cancer, not only the cancer cells themselves are responsible for the malignancy, but also stromal (non-cancerous) cells and the molecular signals they send to cancer cells are important factors that determine the severity and outcome of the disease. Therefore, the identification of stromal signals and their influence on cancer cells is important for the development of novel treatment strategies. With a computational systems biology model of stroma-cancer cell communication, we have compiled a set of ten proteins secreted by stromal cells that shape the cancer phenotype. Most importantly, our causal analysis uncovered Pregnancy-Associated Plasma Protein A (PAPPA) as a novel paracrine inducer of the pro-tumorigenic NFκB signaling pathway. In liver cancer patients, higher levels of PAPPA protein indicate a more progressed tumor stage, confirming its clinical relevance.
| Stromal tissue is a major component of solid tumors. It consists of extracellular matrix, connective tissue cells, inflammatory cells, and blood vessels. Stromal cells affect cancer development and progression by augmenting tumor cell proliferation, survival, motility and invasion [1,2,3]. Tumor and stromal cells can interact through both, direct cell-cell contact and secreted factors such as growth factors, cytokines, chemokines, and their cognate receptors [2,3].
Hepatocellular carcinoma (HCC) is one of the most prevalent and lethal malignant tumors worldwide. The major risk factor predisposing to HCC is hepatic cirrhosis. It arises through the activation of hepatic stellate cells (HSC), myofibroblast-like cells that are responsible for the excessive hepatic matrix deposition seen in chronically damaged livers [4,5]. Moreover, HSCs infiltrate the stroma of liver tumors localizing around tumor sinusoids, fibrous septa, and capsules [4,1]. Conditioned medium collected from activated HSCs induces growth, migration and invasion of HCC cells in vitro [6,7,8,9]. Furthermore, HSCs promote aggressive growth of HCC cells in experimental in vivo models [4,6,9,10] and their presence predicts poor clinical outcome in HCC patients [11]. These data indicate that HSCs affect HCCs. Yet, the molecular mechanisms of this crosstalk are largely unknown.
In functional assays, signaling pathways are analyzed through perturbation of the cellular systems. Unlike statistical associations in observational data, functional assays can directly distinguish between cause and effect. Their disadvantage is that they can be difficult to perform in high throughput.
Recently, Maathuis and colleagues introduced a novel method to extract causal information from observational gene expression data [12]. In their IDA algorithm they combine local reverse network engineering using the PC-algorithm [13] with causal effect estimation [14,15]. These virtual functional assays predict lists of genes that will change expression if the expression of a query gene was perturbed experimentally. The method was successfully applied to predict the expression profiles of yeast deletion strains from observational data of wild type yeast only [16].
Here, we adapt the IDA framework to the problem of identifying agents of inter-cellular communication. We combine a specific experimental design with tailored causal discovery and data integration algorithms. In brief, HSCs obtained from n = 15 human donors were cultivated to generate conditioned media for stimulation of the established HCC cell line Hep3B. Gene expression was then measured in both, HSCs as well as stimulated and un-stimulated HCC cells and a list of genes that change expression in HCCs upon stimulation was established. First, we aimed at identifying gene pairs (x, y) where the expression of gene x in HSCs affects the expression of gene y in HCC cells. Next, we searched for a small set of HSC expressed genes that, together accounted for the majority of stimulation sensitive genes in HCC cells. This yielded a set of 10 HSC genes predicted to jointly influence 120 of 227 HCC cell genes affected by supernatant stimulation.
To study cell communication directed from stroma to cancer cells, we treated the HCC cell line Hep3B with 15 media conditioned by 24-hour cultivation with HSCs that had been isolated from different human donors. This design allows us to study the messages sent from HSCs to HCC cells independently from feedback messages that might be sent in the opposite direction from HCC cells to HSCs. The lack of feedback in this design is an indispensable prerequisite for our analytic approach.
Genome-wide gene expression was measured in both, donor HSCs and HCC cells stimulated with conditioned media (CM), yielding 15 pairs of gene expression profiles. The gene expression profiles of four un-stimulated HCC cell cultures served as controls. We identified a list of 227 genes with more than two-fold expression changes between stimulated and un-stimulated cells at an estimated false discovery rate (FDR) of 0.001. Interestingly, 30 (13.2%) of the 227 genes were among the top 200 genes with the highest variance in expression across the 15 stimulation assays (Fig 1). These genes reflect biological variation both across HSC donors and cancer cells stimulated by the HSCs. The genes that drive HSC induced neoplastic progression, including proliferation and migration in HCCs, are most likely among them [17]. In fact, testing for overrepresented Gene Ontology terms [18] pointed to several hallmarks of cancer: negative regulation of apoptosis (anti-apoptosis, q < 10–4), angiogenesis (q < 10–4), inflammation (cellular response to lipopolysaccharide, q < 10–4), positive regulation of cell migration (q < 10–3), and growth factor activity (transforming growth factor beta receptor signaling pathway, q < 10–3)(S1 Fig).
Next, we searched for indications which pathways might be regulated by stromal signals in HCC cells. The CM sensitive genes were mapped onto the BioGRID interactome of established protein-protein and protein-gene interactions [19] and the largest regulated subnetwork was identified by the BioNet algorithm [20]. The regulated network comprises several interacting oncogenic signaling pathways including TGF-beta/SMAD3, NFκB, JAK1 and MAP kinase signaling components (Fig 2). Another branch of the subnetwork can be attributed to anti-apoptotic signals with the highly induced BIRC3 gene (ENSG00000023445) in its center. Amplification of the chromosomal region containing BIRC3 exons is frequently found in HCC and associated with chemotherapy resistance, metastasis and poor prognosis [21]. The strongest induced gene, RND1 (log2 fold change of 4.9; ENSG00000172602), a member of the Rho GTPase family [22], belongs to yet another branch of the subnetwork that comprises genes involved in regulating rearrangements of the actin cytoskeleton and, thus, changes in cell adhesion and motility in response to extracellular growth factors [23].
So far, we have only described the HSC-mediated changes in the HCC cell transcriptome. We have not yet identified the HSC secreted proteins that actually stimulate receptors or otherwise directly interact with HCCs. In a naïve analysis, we might find many genes in HSCs that correlate with some of the genes that are regulated in HCCs; however, most of them will not cause these changes.
In fact, if we counted the number of HCC genes a particular HSC gene correlates with (absolute Pearson correlation > 0.7), we would identify HSC-secreted POSTN (ENSG00000119655), PGF (ENSG00000119630), CSF1 (ENSG00000184371), NPC2 (ENSG119655) and FGF5 (ENSG00000138675). The top 10 list also includes HGF (ENSG00000019991) and is shown in S1 Table. Although this list points to potential stromal regulators, for some gene pairs correlation will be high due to a third factor that influences both of the correlated genes. To exclude the latter and to find true causal regulators instead, we use the “in silico perturbation framework” of the IDA algorithm [12] to filter for genes that are operative in stroma-to-tumor communication. Application of IDA comprises two steps. First, a partially directed network of regulatory interactions is constructed using the PC algorithm [13]. Second, causal effects are estimated from this network using Pearl’s Do-calculus [14]. To infer a potential effect of a stromal gene x on a cancer gene y, the Do-calculus needs the expression of y, x, and all genes x’ that generate spurious correlations between x and y (e.g. common regulators). Adjusting for the expression of the x’ (termed “parents of x”) differentiates between true causal effects and spurious correlations. If x does not have parents in the network (e.g. x10 in Fig 3), the estimated causal effect is identical to the correlation coefficient. However, if there are parents, causal effects are different from correlation coefficients. In these cases interpreting correlation coefficients is misleading. Since HSCs were never in contact with HCC cells, parent genes of x must be of HSC origin. Hence, it is sufficient to confine the reconstruction of a regulatory network to the HSC expression profiles only. An illustration of the HSC network is shown in Fig 3. To limit the computational burden resulting from genes that are not expressed in HSCs or that did not vary across HSCs from different donors, we only included the highest and most variably expressed genes (see Material and Methods) across the HSC samples in the analysis. The expression levels of HCC cell genes enter the model in the second step as y-genes, and the HSC network is used to derive causal effects of HSC on HCC genes (represented by green dashed arrows in Fig 3). For some genes, we have two expression values: one from the HSC sample that produced the CM, and one from the respective CM-stimulated HCC cell sample. For simplicity, we refer to these expression levels as the expression of the HSC gene and the HCC gene, respectively. For each of the 227 HSC-inducible HCC genes, we used IDA to screen for potential HSC genes that—when perturbed in expression—will have strong effects on the respective HCC gene. We limited our search for candidate HSC regulators to genes annotated as ‘secreted’, ‘extracellular’ or ‘intercellular’, but not ‘receptor’ by Gene Ontology and for which the gene product was detected in the conditioned media by HPLC/MS/MS. Gene products that are too small for detection, e.g. IGF1 (ENSG00000017427) and IGF2 (ENSG00000167244) were left in the analysis. This resulted in a final list of 186 HSC genes as candidate stromal regulators. The gene list with corresponding proteins can be found in S2 Table. Gene-pair-by-gene-pair, the HSC gene was “virtually repressed” by one standard unit and the expected change of the HCC gene was calculated. It is important to note that causal analysis will discover both direct and indirect effects of x on y, i.e. irrespective of potential mediators m, and discover effects of x and m if they are both secreted HSC genes. For example, in Fig 3, x10 has a causal effect on y3, although mediator node x11 also has a causal effect on y3. To be robust against small perturbations of the data, the "virtual repression" was run in a sub-sampling mode, repeating the experiment 100 times each on a different subset of the samples. Within each run, secreted HSC genes were ranked by the size of their estimated effects on the 227 target HCC genes. We kept causal effects only if they appeared in the top ranks across the majority of sub-sampling runs (see Material and Methods). This resulted in 96 HSC genes potentially regulating at least one of the 227 HCC genes. A flow-chart of our methodology is depicted in Fig 4.
Although all 186 HSC proteins have the potential to affect the expression of HCC genes, we postulate that a much smaller set of proteins is sufficient to activate HCCs. Thus we aimed at identifying a small set of HSC genes that jointly account for the wide spectrum of expression changes in HCC cells observed in response to stimulation with HSC-CMs. We have generated 227 lists of HSC regulators, one for each of the 227 CM sensitive HCC genes. Since many HSC genes were predicted to affect multiple HCC genes, these lists overlap. The lists can be reorganized by HSC genes instead of HCC genes. This resulted in 96 non-empty sets of HCC genes that are targeted by the same HSC gene.
Model based gene set analysis [24] (MGSA) is an algorithm that aims at partially covering an input list of genes with as little gene ontology categories as possible. It balances the coverage with the number of categories needed. We modified this algorithm in such a way that it covered the list of 227 CM sensitive HCC genes with the 96 sets of HSC targets. This strategy identified sparse lists of predicted targets that covered most of the observed targets. By definition, every list corresponded to one secreted HSC protein. This analysis brings HSC genes in competition to each other: an analysis based on frequencies (how many HCC genes does each HSC gene affect) discovers redundant HSC genes that target the same HCC genes. Our approach strives for a maximum coverage of the target genes with a minimum number of HSC secreted genes.
Both stability selection on the IDA algorithm and MGSA depend on the setting of a few parameters. Several studies have shown that hepatocellular growth factor (HGF) affects HCC cells [25], and is highly expressed in HSCs [25,26]. We exploited this knowledge and calibrated the parameters such that HGF appeared in the list of predicted HSC genes.
With these parameters, we identified 10 HSC secreted proteins. In addition to HGF the list included PGF, CXCL1, PAPPA, IGF2, IGFBP2, POSTN, NPC2, CTSB, and CSF1 (Table 1). With the exception of IGF2 all proteins were found in at least one of five CMs that were analyzed using LC/MS/MS. IGF2 is too small for successful detection [27]. Notably, the set of the most influential HSC regulators included several well-known tumor-promoting genes such as placental growth factor (PGF) [28], and the chemokine CXCL1, which promotes HCC angiogenesis and growth [29]. Periostin (POSTN) is a secreted cell adhesion protein whose expression levels are directly related to metastatic potential and poor prognosis of HCC [30]. High expression levels of the macrophage colony-stimulating factor 1 (CSF1) are another indicator of tumor progression and poor survival in HCC patients [31]. Over-expression of cathepsin B (CTSB), on the other hand, promotes HCC cell migration and invasion [32].
The role of Niemann-Pick Type C2 (NPC2) protein in cancer is just beginning to be understood. NPC2 regulates intracellular cholesterol homeostasis via direct binding with free cholesterol. Perturbations of cholesterol metabolism affect cancer progression [33]. Elevated serum levels of NPC2 have been observed in patients with lung cancer [34] and, more recently, HCC [35]. Modulation of cholesterol homeostasis by NPC2 also affects activation of mammalian target of rapamycin (mTOR) [36], a critical signaling cascade in several types of cancer including HCC [37].
Remarkably, we identified three genes of the insulin-like growth factor (IGF) axis. This signaling pathway regulates tumor progression in several types of tumors including HCC [38]. The key molecules in this pathway are the ligands IGF1 and IGF2, IGF-binding proteins (IGFBPs), and membrane-associated receptors (IGF-I receptor (IGF-IR), mannose-6-phosphate receptor/IGF-II receptor (IGF-IIR)). High expression levels of IGF2 are predictive of aggressive tumor growth and poor prognosis in HCC patients [39]. IGF2 binds to the receptor tyrosine kinases IGF1R (ENSG00000140443) and IGF2R (ENSG00000197081) on HCC cells and activates multiple intracellular signaling pathways, including the phosphatidylinositide-3′-kinase (PI3K)/Akt and MAP kinase signaling cascades [40]. IGFBPs bind to IGFs with higher affinity than IGF-receptors and, thereby, modulate local IGF concentrations and activities [40,41]. Unlike most IGFBP family members, which conduct antitumor activity, IGFBP2 promotes invasion, metastasis, and angiogenesis [41]. It is over-expressed in several tumor tissues including HCC [41,42].
The metalloprotease Pregnancy-Associated Plasma Protein A (PAPPA) is also a member of the IGF-axis. PAPPA is implicated in several biological functions [43], including the regulation of local IGF1 bioavailability through cleavage of IGFBPs [44]. Its expression in the liver under both, physiological and pathological conditions, including HCC development and progression, has not been elucidated yet. The few available studies on other tumor entities located PAPPA expression to cancer rather than stromal cells [45], and controversial roles of PAPPA regarding tumor progression have been reported in ovarian cancer [46]. Thus, we decided to focus our subsequent analysis on the role of PAPPA in HCC.
In principle, parameters in our analysis could be set to different values and lead to different results. We evaluated the influence of gene pre-filtering and parameter settings in our analyses and found that the results were stable within the computationally feasible settings. Gene pre-filtering was necessary because network estimation is computationally very demanding with many genes. We evaluated our criteria for gene selection in a leave-one-out cross-validation and found that the selected genes are stable (secreted HSC genes: 95.1% identical with standard deviation (SD) 0.7%, intracellular HSC genes: 86.6% identical with SD 1.3%, HCC genes: 97.2% identical with SD 1.4%). S3 Table shows an aggregation of results when varying parameters in the causal analysis and demonstrates that these results are also stable. Among others, PAPPA is always within the top 10 stromal regulators.
The list of CM sensitive HCC genes includes various members of the NFκB pathway (Fig 2; NFKB1 (ENSG00000109320), NFKB2 (ENSG00000077150), NFKBIZ (ENSG0000014480), NFKBIA (ENSG00000100906), RELB (ENSG00000104856)) and targets of the NFκB pathway previously collected by Compagno et al [47], such as BIRC3, EGR1 (ENSG00000120738), ICAM1 (ENSG00000090339), IL8 (ENSG00000169429), MAP3K8 (ENSG00000107968). Several of these genes were predicted to be targets of HSC secreted PAPPA by our causal analysis (ICAM1, MAP3K8, NFKBIA, see S4 Table for the full list). Also the other predicted target genes are known to be regulated by the transcription factor NFκB or to affect this signal transduction pathway [48,49,50,51,52,53]. To test whether PAPPA might be indeed responsible for activation and auto-regulation of the NFκB pathway, we assessed NFκB activity in stimulated HCC cells and observed a striking correlation of PAPPA levels in conditioned medium (CM) from the 15 different HSCs with NFκB activity induced in HCC cells upon incubation with these different CMs (Fig 5A). To verify a causal effect of PAPPA on NFκB activity in HCC, we stimulated Hep3B HCC cells with recombinant human PAPPA protein (rPAPPA). We applied rPAPPA (25 ng/ml) either alone or in CM of HSCs from two different donors containing endogenous PAPPA levels of 4.8 ng/ml and 6.2 ng/ml, respectively. In control medium, rPAPPA did not significantly affect IkB-α- and p65-phosphorylation, while together with CM both IkB- α- and p65-phosphorylation were higher than in CM-stimulated cells (Fig 5B).
Quantitative real time PCR analysis showed strong PAPPA mRNA expression in HSCs whereas no expression was detectable in 4 different human HCC cell lines including Hep3B (S2 Fig). Concordantly, PAPPA protein levels ranged from approximately 5 to 35 ng/ml in supernatants of HSC cultures, while no PAPPA protein was detectable in the supernatants of the 4 different HCC cell lines (Figs 6A and S3). In the 15 different HSCs, we observed a significant correlation between mRNA and protein levels of PAPPA (Fig 6B), indicating that secreted PAPPA levels are regulated at the transcriptional level. Next, we assessed PAPPA gene expression in HCC specimens from 52 patients and found a significant correlation with collagen type I (COL1A1; ENSG0000010882) mRNA expression (Fig 6C). This finding could be confirmed in the HCC cohort of The Cancer Genome Atlas (TCGA, http://cancergenome.nih.gov) (S4 Fig). HSCs infiltrate and form the HCC stroma and collagen type I is specifically expressed by HSCs in HCC tissue [4,54,5]. Together, these findings indicate that HSCs are the major source of PAPPA in HCC.
Histological staging of HCC is a prognostic factor of patient survival in HCC [54,55,56]. We found that PAPPA expression in human HCC specimens (n = 52) was significantly lower (p = 0.008, one-way ANOVA) in patients with low histological staging (stage I; n = 12) compared to patients with stage II (n = 19) and stage III (n = 21) disease (Fig 7). In an independent data set, the HCC cohort of TCGA, PAPPA expression was also significantly lower in stage I patients (n = 104) compared to stage II (n = 56) and stage III (n = 39) in a one-way ANOVA (p = 0.0126) (S5 Fig). Together, these findings indicate the clinical relevance of HSC secreted PAPPA for HCC progression.
Introductory statistical text books stress the difference between association and causation. For example, correlation between the expression levels of two genes does not imply that one gene regulates the other. They can as well be co-regulated by a third gene. The gold standard to infer causalities is experimental intervention. If a knock-down of the first gene changes the expression of the second, there is a functional relation between the two. In fact, the rationale of functional genetics is to understand the cell by breaking it. Functional assays that perturb biological networks experimentally shed light on cellular mechanisms.
Causal inference from observational data is a more advanced statistical discipline [13,14] that only recently found its way into bioinformatics and systems biology after a statistical breakthrough paper by Maathuis et al. (2009) [12]. To date it has been used for the analysis of yeast deletion strains [16], to predict genes regulating flowering time in Arabidopsis thaliana [57], and for the prediction of miRNA targets [58]. Here, we add another biological application to this list: The identification of secreted proteins that drive inter-cellular communication in human cancer.
State of the art statistical methodology does not allow for feedback mechanisms between the regulator and its target. This is an assumption that nature does not meet in many cases. In a tumor it is most likely that the communication between stromal and tumor cells is mutual. In our experimental setting however, feedback is blocked. Stromal and cancer cells grow in separate cultures. The stromal cells "talk" to the cancer cells via the CMs but there is no "reply".
Clearly, this does not give us a full picture of cellular communication; feedback mechanisms are blocked and so are signals mediated by cell-cell contacts. But it is this focus on unidirectional paracrine signaling that allows us to use causal modeling. The experimental design is tailored to the capabilities of the predictive model. In spite of these limitations our application to HCC demonstrates that the method can generate novel and potentially clinical relevant insights into the mechanisms of stroma-tumor communication. We unmasked PAPPA as a novel stroma secreted factor impacting the tumor phenotype.
Notably, our 10 HSC secreted regulators did not only include PAPPA but two more genes of the IGF-axis. The IGF-axis is one of the molecular networks involved in the formation, progression and metastatic spread of many cancer types, including HCC. IGF2 and IGFBP2 are known to critically affect HCC development and progression. Still, most studies focused on autocrine effects of these two secreted proteins in cancer cells, while our data suggest a paracrine effect whereby HSC derived IGF2 and IGFBP2 influence IGF-signaling in HCC cells.
The expression and function of PAPPA in normal and diseased liver were not known thus far. To date, PAPPA has been mainly used as a biomarker in prenatal screening for Down's syndrome [43]. More recently, PAPPA has been identified as a regulator of the bioavailability of IGFs through the cleavage of IGF binding proteins [43,59]. It has been suggested to exert a protumorigenic role in breast cancer, lung cancer, and malignant pleural mesothelioma [59]. In contrast, breast cancer cells have been reported to become more invasive after down-regulation of PAPPA [60]. Controversial roles of PAPPA have also been reported in ovarian cancer, with most ovarian cancer cell lines and primary tumors showing partial or complete loss of PAPPA expression [45]. Furthermore, PAPPA expression was shown to be consistently high in normal ovarian specimens, while it was suppressed by SV40 large T antigen [61].
In HCC, our data suggest PAPPA as a protumorigenic factor. We found significantly higher PAPPA expression levels in advanced stage tumors. On the mechanistic side, we found that PAPPA induces NFκB-activity in HCC cells. We observed a significant correlation between PAPPA levels in different conditioned media of HSCs and corresponding effects on NFκB activation in HCC cells in vitro.
Interestingly, studies in ovarian, breast and lung cancer as well as malignant pleural mesothelioma revealed the cancer rather than the stromal cells as the cellular source of PAPPA. Here, in contrast, PAPPA expression was only detected in HSCs, but not in HCC cells. This makes PAPPA a promising therapeutic target in HCC, as tumor stromal cells are genetically more stable than cancer cells, which renders them less likely to evade therapy. Moreover, it has to be considered that the IGF-axis also plays a critical role in HSC activation and fibrosis [62]. Although the function of PAPPA in HSCs is unknown, it may be speculated that PAPPA inhibition may suppress the fibrogenic phenotype of HSCs. Since HCC mostly develops in cirrhotic liver tissue [1,4], inhibition of PAPPA could not only affect HCC cells but also prevent the formation of a protumorigenic soil for cancer cells.
Due to its central role in cancer progression, a variety of reagents have been developed to modulate IGF signaling including neutralizing antibodies against IGFs and IGF-receptors as well as associated receptor kinase inhibitors in aim for cancer treatment [63]. The structural similarities of the insulin and IGF-IRs complicate the development of specific agents that block IGF-IR signaling without affecting insulin signaling. This is particularly true with regards to treatment of liver cancer due to the central role of the liver in glucose metabolism and homeostasis. In contrast to the persistent and versatile physiological functions of other components of the IGF1 axis, PAPPA could not be detected in normal human liver and primary human hepatocytes (S6 Fig). Therefore, PAPPA appears as a better therapeutic target for HCC with more tumor specificity and less risks of side effects as compared to other IGF1 axis components. Actually, genetic deletion of PAPPA extended lifespan of mice [59,64].
In conclusion, we have shown for the first time that causal modeling can be used to identify stromal signaling molecules that influence the cancer phenotype. Application of our modeling strategy unmasked PAPPA as a novel paracrine factor that shapes the tumor phenotype via activating the NFκB pathway.
Human liver tissues were obtained and experimental procedures were performed according to the guidelines of the charitable state controlled foundation HTCR (Human Tissue and Cell Research), with the informed patients’ consents, and approval by the local ethics committee of the Ludwig-Maximilians University of Munich (reference number 025–12). All experiments involving human tissues and cells have been carried out in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki).
The human HCC cell lines Hep3B (American Type Culture Collection (ATCC) number HB-8064), HepG2 (ATCC; HB-8065), PLC (ATCC; CRL-8024) and Huh-7 (Japan Collection of Research Bioresources (JCR) number B0403) were cultured as described [10,65].
Primary human hepatic stellate cells (HSCs) were isolated from 15 different human donors as described [10,66,67]. The isolation procedure and cell culture on uncoated tissue culture dishes led to the activation of HSCs as described [66,67]. For collection of conditioned medium (CM), HSCs were seeded into T75 flasks (2 × 106 cells). One day after seeding cells were washed twice with serum-free DMEM, and then incubated for another 24 h with serum-free DMEM (15 mL/T75). CM was centrifuged at 6,000 x g to remove cell debris, sterile filtered (0.45 μm pore size membrane filter), and stored in aliquots at −80°C until use. Serum-free DMEM incubated for 24 h in cell culture flasks without cells served as the control.
For stimulation with HSC conditioned media, HCC cells were seeded into T25 flask (106 cells). One day after seeding, cells were washed with serum-free DMEM, and then incubated for another 12 h with serum-free DMEM. Subsequently, the medium was changed and cells were incubated with 3 mL of HSC-CM or control medium (serum-free DMEM) for 4 h.
For individual experiments, CM was preincubated with recombinant PAPPA (R&D Systems, Wiesbaden, Germany).
HCC tissues were obtained from HCC patients undergoing surgical resection. Tissue samples were immediately snap-frozen and stored at -80°C until analysis.
Isolation of total cellular RNA from cultured cells and tissues and reverse transcription were performed as described [10,65]. 300 ng of RNA were hybridized to Affymetrix Human Gene ST 1.0 arrays following the standard Affymetrix protocol (Affymetrix, High Wycombe, UK). Hybridization and scanning were performed at an Affymetrix Service Provider and Core Facility, “KFB—Center of Excellence for Fluorescent Bioanalytics” (Regensburg, Germany; www.kfb-regensburg.de).
Quantitative real-time-PCR was performed applying LightCycler technology (Roche, Mannheim, Germany) and the following pairs of primers: human PAPPA (forward: 5'-AGC CAG CAG CAT CCC AGG TGT-3'; reverse: 5'-CGC CCG GAG CCA AAA AGT GGT)-3' and human collagen type I (forward: 5'- CGG CTC CTG CTC CTC TT -3'; reverse: 5'-GGG GCA GTT CTT GGT CTC -3'). Amplification of cDNA derived from 18s rRNA (forward: 5'-TCT GTG ATG CCC TTA GAT GTC C-3'; reverse: 5'-CCA TCC AAT CGG TAG TAG CG-3') was used for normalization.
Protein extraction and western blotting analysis were performed as described [65] applying antibodies against phospho-NF-κB p65 ((Ser536) rabbit mAb #3033) and phospho-IκBα ((Ser32); rabbit mAb #2859) both from Cell Signaling Technology (Danvers, MA, USA; all diluted 1:1,000). Furthermore, an antibody against actin (MAB1501 from Merck Millipore, Billerica, MA, USA; 1:1,000) was applied.
Activated NF-κB was quantified in nuclear extracts with the ELISA based kit TransAm from Active Motif (Rixensart, Belgium) according to the manufacturer's instructions, as described [66].
Normalization of raw intensity values from CEL files was performed using variance stabilization (VSN) [68]. Median polish and a custom chip description file based on ensembl gene identifiers [69] were used to summarize individual probes to obtain an expression level per gene. Raw intensities and normalized gene expression data are available publicly at the NCBI Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) under accession GSE62455. Differential gene expression between Hep3B cells treated with different CMs and untreated Hep3B controls was estimated using limma [70]. All analyses were performed within the statistical programming environment R.
Gene Set Analysis (GSA) was performed using hypergeometric tests implemented in the Bioconductor package HTSanalyzeR [71]. Genes meeting the FDR threshold of 0.001 and an absolute log2 fold change larger than one were selected for testing significant enrichment of Gene Ontology (GO) terms within the Biological Process (BP) branch. The Bioconductor package BioNet [20] was used to find the highest-scoring sub-network within the differentially expressed genes with FDR < 0.001 and an absolute log2 fold change larger than 0.7.
Aliquots of conditioned media (400 μL each) were used for protein precipitation with 4 volumes of ice-cold acetone. After 2 h incubation at -20°C, samples were centrifuged at 20,000 x g for 10 min. Pellets were air-dried and stored at -20°C until further use.
Combining the lists of proteins identified with gel-free and gel-based secretome analysis resulted in 305 proteins total.
Protein pellets were dissolved in 0.5 M triethylammonium bicarbonate (TEAB, Sigma Aldrich, St. Louis, MO, USA) and denatured at 60°C for 1 hour. The exact protein concentration was determined employing a Bradford assay, using a serial dilution of bovine serum albumin (BSA, Sigma Aldrich) from 31.25 to 2000 μg/mL in 0.5 M TEAB for calibration. Disulfide bonds were reduced at 60°C for 1 hour by addition of 4.55 mM tris(2-Carboxyethyl)phosphine hydrochloride solution (TCEP-HCl, Sigma Aldrich), followed by alkylation with 8.7 mM iodo acetamide (IAA, Sigma Aldrich) at 24°C for 30 min. Protein digestion was performed overnight at 37°C using trypsin (Promega, Madision, WI, USA) at a ratio of 1:50 to the protein concentration. Digests were dried in a SpeedVac before adjusting peptide concentration to 1 μg/μL in 0.05% trifluoracetic acid (TFA, Sigma Aldrich).
The HPLC instrument was an UltiMate 3000 Nano LC system from Dionex (Germering, Germany) and the mass spectrometer was an LTQ Orbitrap XL from Thermo Scientific (Waltham, MA, USA) equipped with a nano-electrospray ion source. The spray was generated with 10 μm id and 360 μm o.d. fused silica tips from New Objective (Woburn, MA, USA). Tryptic peptides were separated by nano-ion-pair reversed-phase (IP-RP)—HPLC at pH 2.0 on a 150 × 0.20 mm I.D. RP polymer monolith capillary column from Thermo Scientific using a 2-hour gradient of 0–40% acetonitrile in 0.05% aqueous trifluoroacetic acid at a flow-rate of 1 μL/min. The MS1 survey scans of the eluting peptides were executed in the LTQ Orbitrap XL with a resolution of 60,000, recording a window between 450.0 and 2000.0 m/z. The three most intense precursor ions were selected for fragmentation with collision-induced dissociation (CID). The normalized collision energy (NCE) was set at 35.0% for all scans. Data evaluation was performed with Proteome Discoverer (Thermo Scientific) and the open—source library OpenMS.
Protein pellets were dissolved in 10 μL of LDS-sample buffer and separated on Invitrogen NuPAGE BisTris SDS-gels (4–12%, MOPS-buffer system) with subsequent colloidal Coommassie staining. Lanes were cut into 30 slices of equal size and washed, carbamidomethylated and tryptically digested prior to nano-LC-QTOF-MS/MS analysis as published previously [72]. Tandem mass spectra were searched against the Uniprot database (version 57.15) using the Mascot 2.2 search algorithm (Matrix Science, London, UK) applying the two-peptide-rule.
To find HSC gene products that influence gene expression in HCC cells, we applied Intervention-calculus when the DAG (directed acyclic graph) is Absent (IDA) [12]. The algorithm consists of two parts: first, an equivalence class of DAGs is estimated from the observational expression data with the pc-algorithm [13], before causal effects are derived using the graph and intervention calculus [14].
Prior to modeling, gene selection was performed as follows: Gene products secreted from HSC cells were defined as all genes with the terms ‘extracellular’, ‘intercellular’ or ‘secret*’ in any Gene Ontology term or definition. This yielded 1919 genes. Next, genes coding for receptors were removed. The remaining genes were filtered based on expression level, excluding genes that had not been expressed at least in 1/15 CM-stimulated HSC samples at a level larger than the 40th percentile of expression values across all genes and HSC samples. Next, genes with low inter-quartile range, a robust estimate of the variance, across HSC samples were excluded (lowest 20%), yielding 1024 genes annotated to be secreted or present outside of the cell. Next, the overlap between these genes and the gene products detected by mass spectrometry in the HSC-CM (305 gene products) was generated, resulting in 153 gene products. Additionally, growth factors were retained even if they were not detected, as for example IGFs are too small to be monitored by mass spectrometry. This procedure led to a final number of 186 HSC-secreted proteins with a potential influence on HCC cell gene expression going into modeling. The list of HSC secreted gene products is provided in S2 Table.
From the remaining HSC genes, only the genes with highest expression levels (at least 3 samples above the 40th percentile) and with highest inter-quartile range (top 976, such that the total number of HSC genes was 2000) were selected. These genes were supposed to build the network that regulates the secreted genes. On the HCC sample side, genes were selected for differential expression based on significance (q < 0.001), and on log2 fold change (absolute log2 fold change > 1) to focus only on the strongest responses of the HCC cells. This resulted in 227 HCC genes. The filtering procedure is depicted in the left part of Fig 4. Gene expression values were centered and scaled to standard deviation equal to one to make causal effects comparable across genes. From the 2000 HSC genes (secreted and remaining genes), the equivalence class of DAGs was estimated and causal effects were derived from the secreted HSC genes on the selected HCC genes. IDA needs a single tuning parameter, α, which controls the neighborhood size of the graph. It was set to 0.2 as this resulted in the best balance between a not too sparse network and computational burden (higher α values lead to longer running times). To find effects insensitive to small disturbances of the data, IDA was run in a sub-sampling approach adopted from Meinshausen & Bühlmann [73]. For a total of 100 times, 12 out of the 15 samples were drawn, the CPDAG was estimated and causal effects were derived for each DAG in the equivalence class. As a lower bound, the minimum effect of the individual DAGs was retained. The effects were then ranked across all outcome genes (differentially expressed cancer genes) by effect size for each sub-sampling run and the relative frequency of an effect being among the top 30% of effects across all runs was recorded. All effects with a relative frequency equal or above 0.7 were retained for further analysis and the median effect across all sub-samples was recorded. The steps of the causal analysis are schematically shown in the right part of Fig 4.
To gain insights into the most important HSC derived regulators of gene expression in HCC, Model-based Gene Set Analysis (MGSA) [24] was employed with the modification that gene sets were redefined as all genes targeted by a specific regulator. For example, the gene set ‘CXCL1’ was comprised of all HCC genes on which CXCL1 exerted a predicted causal effect. MGSA was then used to find a sparse set of regulators explaining the observed differentially expressed genes (q < 0.001, absolute log2 fold change > 1). All predictor-target sets with a posterior probability > b were declared to be the most important regulators. The parameters within MGSA were left at default values, but the size of the gene sets (controlled by the relative frequency cutoff in stability selection) used as input of MGSA was calibrated such that HGF, a known true positive, was in the final list of secreted regulators. While this criterion did not give us unique parameter settings, the remaining genes in the lists resulting from different parameter settings that included HGF were almost identical (S3 Table).
Un-normalized RNA sequencing and clinical data of liver hepatocellular carcinoma (LIHC) patients was downloaded from The Cancer Genome Atlas (TCGA, http://cancergenome.nih.gov) and normalized using size factors calculated by the R package DESeq2 [74] (function ‘estimateSizeFactorsForMatrix’) and log2-transformed with a pseudo-count of 1 to avoid missing values for samples with zero counts. For the analysis of association of PAPPA expression levels with staging, patients staged with the 7th edition of the AJCC (American Joint Committee on Cancer) that were classified into stages I, II or IIIA were used (n = 199). Stages IIIB, IIIC, IV, and IVA were omitted because of low sample sizes (n<10). For the correlation of PAPPA levels with COL1A levels, all LIHC patients were used (n = 424).
|
10.1371/journal.pntd.0004490 | Changes in Proteome Profile of Peripheral Blood Mononuclear Cells in Chronic Chagas Disease | Trypanosoma cruzi (Tc) infection causes chagasic cardiomyopathy; however, why 30–40% of the patients develop clinical disease is not known. To discover the pathomechanisms in disease progression, we obtained the proteome signature of peripheral blood mononuclear cells (PBMCs) of normal healthy controls (N/H, n = 30) and subjects that were seropositive for Tc-specific antibodies, but were clinically asymptomatic (C/A, n = 25) or clinically symptomatic (C/S, n = 28) with cardiac involvement and left ventricular dysfunction. Protein samples were labeled with BODIPY FL-maleimide (dynamic range: > 4 orders of magnitude, detection limit: 5 f-mol) and resolved by two-dimensional gel electrophoresis (2D-GE). After normalizing the gel images, protein spots that exhibited differential abundance in any of the two groups were analyzed by mass spectrometry, and searched against UniProt human database for protein identification. We found 213 and 199 protein spots (fold change: |≥ 1.5|, p< 0.05) were differentially abundant in C/A and C/S individuals, respectively, with respect to N/H controls. Ingenuity Pathway Analysis (IPA) of PBMCs proteome dataset identified an increase in disorganization of cytoskeletal assembly and recruitment/activation and migration of immune cells in all chagasic subjects, though the invasion capacity of cells was decreased in C/S individuals. IPA predicted with high probability a decline in cell survival and free radical scavenging capacity in C/S (but not C/A) subjects. The MYC/SP1 transcription factors that regulate hypoxia and oxidative/inflammatory stress were predicted to be key targets in the context of control of Chagas disease severity. Further, MARS-modeling identified a panel of proteins that had >93% prediction success in classifying infected individuals with no disease and those with cardiac involvement and LV dysfunction. In conclusion, we have identified molecular pathways and a panel of proteins that could aid in detecting seropositive individuals at risk of developing cardiomyopathy.
| Chagasic cardiomyopathy is elicited by Trypanosoma cruzi infection. T. cruzi transmission is prevalent in Latin American countries, and its transmission is also noted in Mexico and Southern parts of the United States. In this manuscript, we have utilized blood samples from human subjects that were normal healthy or were infected with T. cruzi and exhibited variable symptoms of heart disease. We have employed a highly sensitive approach of protein labeling, developed a detailed proteomic map from all samples, performed comparative analysis of gel images, and identified a panel of proteins that were changed in abundance in clinically asymptomatic (C/A) and clinically symptomatic (C/S) chagasic individuals with respect to healthy controls. Functional annotation of these proteins suggested that pathologic mechanisms in disease progression would involve host’s inability to recruit immune cells, scavenge free radicals, and prevent cell death. We also describe a panel of proteins that can differentiate C/A from C/S subjects and will potentially be useful in identifying infected individuals at risk of developing clinical disease.
| Chagasic cardiomyopathy is caused by Trypanosoma cruzi. According to the World Health Organization report released in 2010, ~16 million individuals are infected with T. cruzi, and >25 million people are at risk of infection in Latin America and Mexico [1]. New challenges of increased transmission are faced due to lack of sustainability of the vector control programs [2,3], migration of infected individuals to non-endemic areas (e.g. US, Canada, Europe) [4,5], and transfer of infection through blood or organ donation [6,7]. The Centers for Disease Control reports that >300,000 individuals infected with T. cruzi are currently living in the United States [8]. Several years after the initial exposure to the parasite, ~30–40% of the infected individuals develop cardiomyopathy and may progress to heart failure (reviewed in [9]). No vaccine is available for the prevention of infection [10] and the available drugs, benznidazole and nifurtimox, have exhibited no significant effects in arresting the progression of chronic cardiomyopathy [11]. Importantly, tools to assess the effectiveness of new drugs against T. cruzi infection and Chagas disease are currently not available.
We have found that T. cruzi elicits oxidative stress of inflammatory and mitochondrial origin in immune and non-immune cells; and sustained oxidative stress plays a crucial role in eliciting left ventricular dysfunction during progressive Chagas disease [9,12,13]. Our studies showed that myocardial changes in oxidant/antioxidant balance and oxidative adducts were detectable in the peripheral blood of infected mice [14] and chagasic patients [15–17]. The level of oxidative stress markers (i.e. lipid hydroperoxides) and inflammation (i.e. myeloperoxidase) increased and the level of antioxidants (e.g. manganese superoxide dismutase) decreased in both heart and peripheral blood of infected rodents with progressive disease [14]. These studies, thus, support the notion that peripheral blood cells provide a suitable tissue for delineating the pathways that are deregulated during the chronic development of chagasic cardiomyopathy.
In this study, we have employed a quantitative saturation fluorescence labeling approach for the detection of the differential protein signature of peripheral blood mononuclear cells (PBMCs) in T. cruzi-infected subjects. All enrolled subjects were assessed by electrocardiography and transthoracic echocardiography and characterized for the severity of cardiac disturbances. We employed a thiol-labeling maleimide dye under saturating conditions that exhibits stable, specific, quantitative labeling of cysteine residues in conjunction with two-dimension electrophoresis and mass spectrometry for developing the PBMCs’ proteome of chagasic patients. Up to 92% of the human proteins contain at least one cysteine residue [18], and thus can be detected using the thiol-labeling maleimide dye. Our findings provide clues to the molecular pathways that may be disturbed with development of chronic Chagas disease. We discuss a panel of proteins that could potentially be useful in classifying the disease state and identifying asymptomatic individuals at risk of developing clinical disease.
All chemicals and reagents were of molecular grade (>99.5% purity). BD Vacutainer CPT Cell Preparation Tubes (heparinized) containing 8 ml whole blood samples were centrifuged following manufacturer’s instruction. The FICOLL Hypaque™ density gradient was employed to enrich the PBMC fraction, and the latter was pelleted by centrifugation at room temperature at 400 x g for 10 min. The PBMC pellets were suspended in 1 ml of hypotonic buffer to lyse contaminating red blood cells, and 9 ml of complete RPMI-1640 medium / 10% fetal bovine serum (Invitrogen) added. After centrifugation as above, final cell pellets consisting of 8-10-million PBMCs were stored at -80°C.
PBMC pellets from individual study subjects were lysed in 7 M urea, 2 M thiourea, 2% CHAPS, and 50 mM Tris (pH 7.5), containing benzonase nuclease (300-units/ml), as described previously [20,21]. Protein concentrations were determined by using a Pierce Modified Lowry Protein Assay Kit, and cysteine (cysteic acid) levels in all samples were determined by using an Amino Acid Analyzer (Model L8800, Hitachi High Technologies America, Pleasanton, CA) [20]. Samples were incubated for 1 h with 6 mM ascorbate (Asc) to ensure all cysteine residues were reduced and available for dye-binding, dialyzed against urea buffer to remove excess ascorbate, and then labeled with BODIPY FL N- (2-aminoethyl) maleimide (BD from Life Technologies, Grand Island, NY) at 60-fold excess to cysteine [21]. The mixtures were incubated for 2 h; the reactions were stopped with a 10-fold molar excess of 2-mercaptoethanol over dye. All incubations were carried out at room temperature in the dark in 200 μl reaction volume [20,21].
BD-labeled PBMC lysates (100 μg protein) were separated by 2-dimension electrophoresis (2DE), employing an IPGphor multiple sample isoelectric focusing (IEF) device (GE Healthcare) in the first dimension, and the Criterion Dodeca cell (Bio-Rad) in the second dimension, as we have described previously [22,23]. Briefly, samples were loaded on to 11 cm dehydrated precast immobilized pH gradient (IPG) strips (GE Healthcare), and strips were rehydrated overnight. IEF was performed at 20°C with the following parameters: 50 V, 11 h; 250 V, 1 h; 500 V, 1 h; 1,000 V, 1 h; 8,000 V, 2 h; 8,000 V, 48,000 V/h. The IPG strips were then incubated in 10 ml of equilibration buffer (6 M urea, 2% sodium dodecyl sulfate (SDS), 50 mM Tris-HCl, pH 8.8, 20% glycerol) for 30 min at 22°C, and electrophoresis was performed at 150 V for 2.25 h, 4°C using precast 8–16% polyacrylamide gels in Tris-glycine-SDS buffer (25 mM Tris-HCl, 192 mM glycine, 0.1% SDS, pH 8.3) [22,23].
Gels were fixed in 20% methanol / 7% acetic acid / 10% acetonitrile for 1 h and washed with 20% ethanol / 10% acetonitrile to reduce background. Gel images were acquired at 100 μm resolution using the Typhoon Trio Variable Mode Imager (GE Healthcare) to quantify BD-labeled proteins (Ex488 nm / Em520 nm). Up to 92% of the human proteins contain at least one cysteine residue [18]. The Totallab SameSpots software (formerly Nonlinear Dynamics Ltd. Newcastle, UK) selects one reference gel according to several criteria, including quality and number of spots with the intent on selecting the gel that best represents all the gels. The reference gel containing the most common features was selected from the pool of gels of the N/H samples, and all data were then derived by comparison to the N/H reference gel. To ensure that the maximum numbers of proteins were detected, the reference gel was also stained with SyproRuby (Life Technologies Grand Island, NY) that binds all proteins irrespective of presence or absence of cysteine amino acid, and gel image was acquired at Ex488nm/Em560nm. The exposure time for both dyes (BD and SyproRuby) was adjusted to achieve a value of ~55,000–63,000 pixel intensity (16-bit saturation) from the most intense protein spots on the gel [22,23].
In total, 83 BD-stained 2D gels representing 30, 25, and 28 samples from N/H, C/A, and C/S subjects, respectively, were scanned and analyzed with the Totallab SameSpots software. After manual and automated pixel-to-pixel alignment, the program performed automatic spot detection on all images. The SyproRuby stained reference gel was used to define spot boundaries; however, the gel images taken under the BD-specific filters were used to obtain the quantitative spot data. This strategy ensures that spot numbers and outlines were identical across all gels in the experiment, eliminating problems with unmatched spots as well as ensuring that the greatest number of protein spots and their spot volumes were accurately detected and quantified [23]. Protein spot abundance ratios were calculated from normalized spot volumes from affected samples versus the matched normal spot volumes (Δ protein abundance = Asc+chagasic/Asc+ N/H controls). Spot volumes were normalized for each sample using a software-calculated bias value assuming that the great majority of spot volumes did not change in abundance (log (abundance ratio) = 0). The scatter of the log (abundance ratios) for each spot in a gel (sample) is distributed around some mean value that represents the systematic factors that govern the experimental variation. Thus, a gain factor is calculated to adjust the mean spot ratios of a given gel to 0 (log (abundance ratio) = 0) and applied to each spot volume [23].
For the purpose of selecting differentially abundant protein spots for mass spectrometry, normalized spot volumes were subjected to statistical analysis using in-built tools in Totallab SameSpots software. Spot volumes were log2 transformed and spot-wise standard deviation, arithmetic mean, and coefficient of variation (CoV) values of the standard abundance values were calculated for each spot [24]. Student’s t-tests with Welch’s correction for unequal variances were used to test for differential protein expression between N/H controls and either C/A or C/S chagasic subjects. Benjamini-Hochberg multiple hypothesis testing correction was applied to account for the false discovery rate and significance was accepted at p<0.05. The protein spots identified to be differentially abundant (p< 0.05) in at least one of the groups were submitted for mass spectrometry identification.
Selected spots on the 2D gels that exhibited significant differential prevalence (p≤0.05) in at least one of the group were picked robotically (ProPick II, Digilab, Ann Arbor, MI), and trypsin digested as described by us [19,25]. In brief, gel spots were incubated at 37°C for 30 min in 50 mM NH4HCO3, dehydrated twice for 5 min each in 100-μl acetonitrile, dried, and proteins were digested in-gel at 37°C overnight with 10 μl of trypsin solution (1% trypsin in 25 mM ammonium bicarbonate). Peptide mixtures (1-μl) were directly spotted onto a MALDI-TOF MS/MS target plate with 1 μl of alpha-cyano-4-hydroxycinnamic acid matrix solution (5 mg/ml in 50% acetonitrile), and analyzed using a MALDI-TOF/TOF AB Sciex TOF/TOF 5800 Proteomics Analyzer (Framingham, MA). The Applied Biosystems software package included the 4000 Series Explorer (v.3.6 RC1) with Oracle Database Schema (v.3.19.0) and Data Version (3.80.0) to acquire and analyze MS and MS/MS spectral data. The instrument was operated in a positive ion reflectron mode with the focus mass set at 1700 Da (mass range: 850–3000 Da). For MS data, 1000–2000 laser shots were acquired and averaged from each protein spot. Automatic external calibration was performed by using a peptide mixture with the reference masses 904.468, 1296.685, 1570.677, and 2465.199. Following MALDI MS analysis, MALDI MS/MS was performed on several (5–10) abundant ions from each protein spot. A 1-kV positive ion MS/MS method was used to acquire data under post-source decay (PSD) conditions. The instrument precursor selection window was +/- 3 Da. Automatic external calibration was performed by using reference fragment masses 175.120, 480.257, 684.347, 1056.475, and 1441.635 (from precursor mass 1570.700) [19,25].
For protein identification, the MS and MS/MS spectral data were searched against the UniProt human protein database (last accessed: March 25, 2013; 87,656 sequences; 35,208,664 residues) by using a AB Sciex GPS Explorer (v.3.6) software in conjunction with MASCOT (v.2.2.07) as described previously [19]. The protein match probabilities were determined by using expectation values and/or MASCOT protein scores. The MS peak filtering included the following parameters: a mass range of 800 Da to 3000 Da, minimum S/N filter = 10, mass exclusion list tolerance = 0.5 Da, and mass exclusion list for some trypsin and keratin-containing compounds included masses (Da) 842.51, 870.45, 1045.56, 1179.60, 1277.71, 1475.79, and 2211.1. The MS/MS peak filtering included the following parameters: minimum S/N filter = 10, maximum missed cleavages = 1, fixed modification of carbamidomethyl (C), variable modifications due to oxidation (M), precursor tolerance = 0.2 Da, MS/MS fragment tolerance = 0.3 Da, mass = monoisotopic, and peptide charges = +1. The significance of a protein match, based on the peptide mass fingerprint (PMF) in the MS and the MS/MS data from several precursor ions, is presented as expectation values (p<0.05). To confirm the identified proteins were of human and not of parasite origin, we also performed a similar search against NCBI non-redundant protein database consisting of T. cruzi sequences.
In cases where abundance was ≥ |2| but protein IDs were ambiguous (protein scores <62), the digested proteins were submitted for analysis by LTQ OrbiTrap Velos (ThermoFisher, Waltham, MA).
We used the Ingenuity Pathways Analysis (IPA) web-based application (Ingenuity Systems, Redwood city, CA) to assess the biological meaning in the proteome datasets. IPA retrieves biological information from the literature—such as gene name, sub-cellular location, tissue specificity, function, and association with disease—and then integrates the identified proteins into networks and signaling pathways with biological meaning and significance [26]. An “e-value” was calculated by estimating the probability of a random set of proteins having a frequency of annotation for that term greater than the frequency obtained in the real set, and a significance threshold of 10−3 was used to identify significant molecular functions and biological processes [19]. With these parameters, we were able to highlight the most informative and significantly over-represented gene ontology terms in the dataset [19,27].
For MARS modeling, normalized spot volumes for all spots from 83 gels were exported from SameSpots in to Excel, and analyzed by using R and SPSS ver.20 software. For modeling the disease state specific response, a stringent cut-off was applied; differentially abundant protein spots were first screened by t test/Welch’s correction and then Benjamini-Hochberg test was employed at p<0.001 (≥І1.5І fold change). MARS was employed to model changes in multiple variables for distinguishing between infection and disease status [24]. We used 10-fold cross-validation and 80% (training)/20% (testing) approaches to predict the protein spots that can distinguish N/H from C/A and C/S subjects. The sensitivity and specificity of the identified models were validated by receiver operator characteristics (ROC) curves.
All protein extracts were analyzed for cysteine content by amino acid analysis and labeled with uncharged BODIPY FL-maleimide (BD, dye-to-protein thiol ratio > 60:1). The saturation fluorescence labeling with BD provided no non-specific labeling, had no effect on the isoelectric point and mobilities of the proteins, and provided a linear dynamic range of over four orders of magnitude in identifying the protein spots (detection limit: 5 f mol protein in a gel spot at a signal-to-noise ratio of 2:1), as we have also noted in a previous study [23].
PBMC lysates of the normal healthy (N/H) controls (n=30), and of seropositive, clinically asymptomatic (C/A, n=25) and seropositive, clinically symptomatic (C/S, n=28) individuals were resolved by 2D-GE. The representative 2D gel images for these groups are shown in Fig 1A–1C. All protein spots were within the relative molecular sizes 10 to 250 kDa.
All of the 2D gel images were assessed for quality control by SameSpots software, and then aligned both manually and automatically against the reference gel (Fig 2), chosen from the entire set of gel images by the software. The fluorescence intensity of the protein spots was normalized using a bias factor calculated assuming most spots did not change across the experiment. The log2 transformed abundance values for each protein spot on 2D gels were utilized to calculate the mean coefficient of variation (CoV) values (Fig 3) for the biological replicates. These data showed the mean CoV values were 49 ± 21.7%, 67 ± 26.4%, and 77 ± 41.3%, for N/H, C/A and C/S groups, respectively (Fig 3A–3C). Up to 75% of the spots in all groups did not exceed the CoV value of 80% indicating that most of the protein abundances are quite stable in the different groups. Protein spots exceeding a CoV of 100% were largely noted in chagasic subjects, indicating a changing and variable protein expression pattern with disease progression.
For the purpose of selecting protein spots for identification by mass spectrometry, the protein spot datasets were analyzed in pair-wise manner by t test with Welch's correction that accounts for unequal variances. This analysis yielded 315 (162 up-regulated, 153 down-regulated, p<0.05) and 348 (180 up-regulated, 168 down-regulated, p<0.05) differentially abundant protein spots in seropositive subjects with no disease and those with LV dysfunction, respectively. These datasets were then submitted to Benjamini-Hochberg multiple hypothesis testing correction to adjust the false discovery rate, and the differentially abundant protein spots (fold change: |≥1.5|, p<0.05 with B-H correction) were submitted for MALDI-TOF/TOF analysis. Homology searches were conducted against the UniProt’s human proteome database for protein identification [19]. A total of 213 protein spots (102 up-regulated, 111 down-regulated, fold change: |≥1.5|) in seropositive/clinically-asymptomatic subjects; and 199 protein spots (97 up-regulated, 102 down-regulated, fold change: |≥1.5|) in seropositive subjects with LV dysfunction were found to be differentially expressed with respect to normal controls, and identified by mass spectrometry (Table 2). These proteins were predicted to be localized in cytoplasm (67%), extracellular space (14%), nucleus (8%), or plasma membrane (9%) (Fig 4A). The changes in abundance frequency of the identified proteins ranged from > -3-fold to >9-fold in chagasic subjects (Fig 4B). A majority of the identified protein spots were differentially abundant in all chagasic subjects though the extent of change in expression was more pronounced in seropositive subjects with LV dysfunction. When we compared the differential abundance of proteins in seropositive C/A versus C/S subjects, we noted 20 and 10 protein spots that were uniquely changed in abundance in clinically-asymptomatic (Fig 4C) and clinically-symptomatic subjects (Fig 4D), respectively, and were relevant to disease state.
We performed IPA analysis to predict the molecular and biological relationship of the differential proteome datasets (Table 2). IPA recognizes all isoforms (e.g. gel-detected pI and size variants of actin, fibrinogen) as the same protein and collapsed the dataset to 82 and 78 differentially abundant proteins in seropositive subjects with no heart disease and those with LV dysfunction, respectively. IPA analysis of the differential proteome datasets predicted an increase in cytoskeletal disassembly and disorganization (z-score: -1.091 to -0.248, S1 Fig), immune cell aggregation (ALB↓, FGA↑, GSN↓, MPO↓, THBS1↑, z-score: 1.521, p value 1.48E-03) and recruitment/activation and migration of immune cells in chagasic (vs. normal) subjects (z-score: 0.501–1.698, p value: 1.94–5.29E-04, S2 Fig), though invasion capacity of cells was decreased in C/S subjects (S2 Fig panel B). Molecular and cellular function annotation of the proteome datasets by IPA predicted a balanced cell proliferation/cell death response in C/A subjects (S3 Fig panel A) while cell death along with inhibition of cell survival was dominantly predicted in PBMCs of C/S subjects (S3 Fig panel B, z-score: 0.858–2.406). IPA also implied a pronounced increase in production of free radicals associated with a decline in scavenging capacity with progressive disease in chagasic subjects (z-score: 1.019 to -1.455, S4 Fig).
The top up-stream molecules predicted to be deregulated and contributing to the differential proteome with disease progression in chagasic subjects included MYC, SP1, MYCN, and growth factor ANGPT2 (z-score -2.266 to -2.190) proteins (S5 Fig).
We performed MARS analysis to develop a classification model for predicting risk of disease development. MARS is a nonparametric regression procedure that creates models based on piecewise linear regressions. It searches through all predictors to find those most useful for predicting outcomes, and then creates optimal model by a series of regression splines called basis functions [28,29]. For this, MARS uses a two-stage process; first half of the process involves creating an overly large model by adding basis functions that represent either single variable transformations or multivariate interaction terms. In the second stage, MARS deletes basis functions in order of least contribution to the model until the optimum one is reached. End result is a classification model based on single variables and interaction terms which will optimally determine class identity [28,29].
Inputs to the model were log2 transformed values for protein spots that were differentially abundant in seropositive/no disease (84 spots, n = 25) and clinically-symptomatic (87 spots, n = 28) groups with respect to normal controls (n = 30) at p<0.001 with B-H correction. We assessed the model accuracy by looking at the prediction success rate and the ROC curves. To address the possible issue of over-fitting the data, we employed two approaches: 1) 10-fold cross validation (CV) allowing same number of maximum basis functions as were the differentially abundant protein spots at p<0.001 (with 1 max interaction term), and 2) testing/training approach in which 80% of the data was utilized for creating the model and the 20% of the remaining data was used to assess the fit of the model for testing dataset. The CV and 80/20 approaches identified 11 and 6 protein spots, respectively, with high importance (score >20, Fig 5A & 5B) for creating the MARS model, detecting differences between the controls and seropositive/no disease subjects. The prediction success showed the CV and 80/20 models fitted perfectly on the training dataset (AUC/ROC: 1.00) and by >93% on the testing dataset (AUC/ROC: 0.96 for CV and 0.933 for 80/20) (Fig 5C & 5D).
Likewise, the CV and 80/20 approaches identified 11 and 8 protein spots, respectively, with high importance (score >20, Fig 6A & 6B) for creating the MARS model distinguishing controls from clinically-symptomatic chagasic patients. The prediction success of the CV and 80/20 models were 100% for the training data (AUC/ROC: 1.00). When fitted on testing data, the CV model exhibited very high prediction success (AUC/ROC: 0.926, Fig 6) while the 80/20 model fitted perfectly on the training data (AUC/ROC: 1.00, Fig 6D). These analyses suggested that PBMC changes in the selected protein spots will have high specificity and sensitivity in predicting the disease state in chagasic subjects in comparison to normal/healthy controls.
This study was aimed at assessing the proteomic changes in PBMCs of chagasic subjects grouped as clinically asymptomatic (C/A, n = 25) and clinically symptomatic with heart involvement (C/S, n = 28) in comparison with healthy subjects (n = 30). 2DE/ MALDI-TOF MS analysis identified 213 and 199 protein spots that were differentially abundant in C/A and C/S subjects in comparison to normal/healthy controls (Table 2).
The major cell populations in PBMCs are lymphocytes (B, T and NK cells, ≥70%) and monocytes/macrophages (10–30%). Very few studies have, however, characterized the role of peripheral immune cells in parasite control vs. cardiac pathology in Chagas disease. For example, a recent study noted detection of no NK cells in early infection [30]. In late acute stage of infection, a selective increase in a distinct lineage of NK cells (CD16+CD56–), as well as a persistent expansion of B cells, possibly indicative of a relationship between B cell activation and a subset of NK cells was noted in humans [30,31]. Others have demonstrated a robust expansion of T cell response in patients with progressive chronic disease though their role in parasite control vs. pathology remains controversial [32–35]. A high frequency of T cells is found in peripheral blood of indeterminate (i.e. C/A) and cardiac (i.e. C/S) patients [35,36], and CD8+ granzyme+ T cells were the main cell type found in infiltrating infiltrate in the myocardium [37]. However, recent studies have suggested that CD8+T cells found in C/A subjects were parasite-antigen specific and functional, while CD8+T cells undergoing immunological exhaustion were noted in C/S patients and their lack of activity contributed to the establishment of pathology [38]. A correlation between the production of inflammatory cytokines (IFNγ > IL-10) by CD4+ T cells and monocytes of C/S patients, and the production of Th2 cytokine profile (IL-10 and IL-4) by the same cells of C/A patients is also shown [39,40]. These studies tend to conclude that functional capacity of T cells along with anti-inflammatory activation of monocytes determines the control of parasite and clinically asymptomatic state in chagasic individuals while functionally incapable T cells and consistent proinflammatory activation of monocytes contributes to chronic, clinically symptomatic disease.
IPA analysis of the proteome datasets in this study suggested that differential migration and/or invasion capacity of immune cells may also contribute to host’s ability to control T. cruzi and enter C/A vs C/S stage. An increase in cellular disassembly and disorganization associated with disruption of filaments that is central to remodeling of the cytoskeleton and modulation of cell shape for migration was observed in PBMCs of all chagasic patients (S1 Fig). Specifically, the expression profile of Ca2+-dependent phospholipid-binding members of the annexin family that possess phospholipase A2 inhibitory activity [41], vimentin and actin isoforms (ACTB, ACTG) that are the cytoskeletal component responsible for maintaining cell integrity and are mediators of internal cell motility [42] and filamin A (FLNA) that interacts with several molecules (e.g. integrins) to regulate the actin cytoskeleton organization [43] were all altered in PBMCs of chagasic subjects. However, the expression levels of small G proteins (Rab14, RAP1B) that regulate membrane trafficking across golgi and endosomal compartments [44,45] and of Rab13 that controls junctional development by directly binding to F actin and modifying actin cytoskeletal reorganization [46] and cell spreading via filamins [47] were increased and decreased in C/A and C/S subjects, respectively, and might have played an important role in determining the extent of immune cell migration in C/A versus C/S chagasic subjects. Consistent with this, all seropositive chagasic subjects exhibited an expression profile indicative of increase in migration of phagocytes and leukocytes (S3 Fig), though a small subset of molecules identified to be linked to invasion process (11 molecules, z score: -2.032, p value: 1.43E-03; ANXA1↓, ANXA2↓, FLNA↓, GSN↓, LTF↑, PKM↓, S100A6↑, SOD2↓, THBS1↑, VIM↓, YY1↑, S3 Fig panel B) were decreased in C/S subjects, thus suggesting that functional lymphocytes may be mobilized in periphery but not able to access and kill tissue parasites.
What might be the source of low-grade antigenic stimulus that results in persistence of immune cells and whether these surviving immune cells are functional in the context of parasite control is not entirely clear. Some investigators have argued that it is the long-term persistence of parasitic antigens that result in exhaustion of the functional T cell compartment [48,49]. The authors noted the frequency of parasite-specific functional CD4+ and CD8+ T cells decreased with more severe stages of clinical disease in human patients, and the T cells that persisted in chronically infected individuals were not metabolically or functionally active and exhibited the phenotypic characteristics of senescence [48,49]. Our data showed an increase in free radical synthesis and a decline in free radical catabolism and scavenging capacity in infected individuals that exhibited more pronounced disease state (S4 Fig, panel B). We and others have shown that oxidative stress is persistent in chronically-infected chagasic animals and patients [14,17,50,51], and oxidized cardiac proteins serve as neo-antigens and recognized by antibody response in chagasic mice and patients [25]. Thus, it is also possible that self-proteins that are oxidized due to persistence of oxidative stress serve as the source of antigenic stimulus for a low-grade but persistent activation of immune cells in chagasic host. The two hypotheses, i.e., parasite or self-antigens contributing to persistence of non-functional, senescent immune cells are not mutually exclusive and together explain why the persistent chronic inflammation is of pathological importance in Chagas disease.
The gene expression studies using global and custom arrays have shown the mitochondrial function-related gene expression is decreased in experimental models of T. cruzi infection and in the cardiac biopsies of chagasic patients [52–55]. A loss in the activity of mitochondrial respiratory complexes (I and III) was also noted in cardiac biopsies of chagasic rodents [14,56] and peripheral blood of human patients [17] that correlated with decreased coupled respiration and ATP generation [50,57]. In this study, PBMCs of chagasic patients showed protein expression pattern indicative of inhibition of glycolysis/gluconeogenesis (↓PKM, ↓GAPDH, ↓ENO1, ↓ADLOA, and ↓PGK1). The abundance of ATP5A1 that contributes to oxidative phosphorylation and ATP synthesis was counter-effected by abundance of MTCH1 that is localized to the mitochondrion inner membrane and induces Bax- and Bak- independent apoptosis [58,59] in chagasic PBMCs. Further, all isoforms of TUFM that participate in protein translation in mitochondria were decreased in chagasic PBMCs. Mutations in TUFM are shown to contribute to oxidative phosphorylation inefficiency and lactic acidosis in infantile encephalopathy [60]. These data provide a novel clue, and suggest that decreased translation and/or transport of mitochondria-targeted proteins affecting the functional assembly of electron transport chain complexes might play a major role in mitochondrial energy deficiency during progressive Chagas disease.
The top upstream regulators, MYC/MYCN and SP1 were predicted to be inhibited (z-score: < 2, p<0.001, all), and identified as common link contributing to expression profile of protein datasets related to metabolism, cell death/cell proliferation, ROS scavenging and cytoskeletal remodeling in chagasic subjects. MYC and MYCN are very strong proto-oncogenes that play a role in cell cycle, apoptosis and cellular transformation through diverse mechanisms. Recently, MYC has been reported to induce accumulation of DNA oxidative adducts and impair cell cycle regulatory capacity which potentially can increase the genomic instability and provide an environment conducive to growth of the cancer cells [61]. Others have shown MYC-dependent-ROS increase induced cell death [62]. Whether MYC-induced ROS contribute to tumorigenesis in human cells is not clearly demonstrated; however, in the context of chagasic subjects, our study suggests that the inhibition of MYC was likely an adaptive response to control pathological outcomes related to uncontrolled ROS production and immune cell proliferation. Indeed as early as 1992, a selective reduction of c-myc and c-fos mRNAs in association with the severe suppression of the IL-2 gene in lymphoid of mice infected by T. cruzi was noted [63]. Like MYC, SP1 transcription factor also modulates the expression of genes involved in cell division, apoptosis, and immune responses. Post-translational modifications of SP1 are suggested to alter its DNA binding and transactivation activity and thereby affect the transcriptional activity [64]. Up regulation of SP1 is shown to be tumorigenic and its reduction was found to be neuroprotective in in vitro and in vivo models of Huntington’s disease [65]. PARP-1, a member of the poly (ADP-ribose) polymerase family, produces poly(ADP-ribose) units (PAR) [66] and PAR modifications of SP1 suppressed its DNA-binding properties [67]. We have shown hyperactivation of PARP-1 stimulated by oxidative DNA damage in cardiomyocytes infected by T. cruzi [68]. How cross-talk of PARP-1 and SP1 determines the expression and transcriptional function of SP1 in the context of chronic chagasic cardiomyopathy remains to be elucidated in forthcoming studies.
In summary, this study demonstrates that unbiased proteomic analysis of PBMCs in a discovery mode is useful in enhancing our knowledge of the pathomechanisms that determine predisposition to and progression of clinically symptomatic Chagas disease. By employing a 2DE and MALDI-TOF/MS approach for developing the PBMC proteome signature of chagasic subjects, we have identified the possible pathologic mechanisms in disease progression would involve host’s inability to recruit immune cells, scavenge free radicals, and prevent cell death. MYC/SP1 transcription factors that regulate hypoxia and inflammatory stress were predicted to be key targets for controlling chagasic pathology. MARS-modeling identified a panel of protein spots that if monitored in infected individuals, will have >93% success in predicting risk of clinical disease development. Our results provide an impetus for further studies in a second independent cohort of patients for confirming the diagnostic potential of suggested panel of proteins.
|
10.1371/journal.pgen.1002263 | Distinct Cdk1 Requirements during Single-Strand Annealing, Noncrossover, and Crossover Recombination | Repair of DNA double-strand breaks (DSBs) by homologous recombination (HR) in haploid cells is generally restricted to S/G2 cell cycle phases, when DNA has been replicated and a sister chromatid is available as a repair template. This cell cycle specificity depends on cyclin-dependent protein kinases (Cdk1 in Saccharomyces cerevisiae), which initiate HR by promoting 5′–3′ nucleolytic degradation of the DSB ends. Whether Cdk1 regulates other HR steps is unknown. Here we show that yku70Δ cells, which accumulate single-stranded DNA (ssDNA) at the DSB ends independently of Cdk1 activity, are able to repair a DSB by single-strand annealing (SSA) in the G1 cell cycle phase, when Cdk1 activity is low. This ability to perform SSA depends on DSB resection, because both resection and SSA are enhanced by the lack of Rad9 in yku70Δ G1 cells. Furthermore, we found that interchromosomal noncrossover recombinants are generated in yku70Δ and yku70Δ rad9Δ G1 cells, indicating that DSB resection bypasses Cdk1 requirement also for carrying out these recombination events. By contrast, yku70Δ and yku70Δ rad9Δ cells are specifically defective in interchromosomal crossover recombination when Cdk1 activity is low. Thus, Cdk1 promotes DSB repair by single-strand annealing and noncrossover recombination by acting mostly at the resection level, whereas additional events require Cdk1-dependent regulation in order to generate crossover outcomes.
| Homologous recombination (HR) provides an important mechanism to eliminate deleterious lesions, such as DNA double-strand breaks (DSBs). DSB repair by HR uses homologous DNA sequences as a template to form recombinants that are either crossover or noncrossover with regard to flanking parental sequences. Furthermore, a DSB flanked by direct DNA repeats can be repaired by another HR pathway called single-strand annealing (SSA). HR is generally confined to the S and G2 phases of the cell cycle, when DNA has been replicated and a sister chromatid is available as repair template. This cell cycle specificity depends on the activity of cyclin-dependent kinases (Cdks), which regulate initiation of HR by promoting nucleolytic degradation (resection) of the DSB ends. Whether Cdks regulate other HR steps is unknown. Here, we show that Saccharomyces cerevisiae Cdk1 has a dual function in HR: it promotes SSA and noncrossover recombination by regulating primarily the resection step, whereas it plays additional functions in allowing recombination accompanied by crossovers. As crossovers during mitotic cell growth have the potential for deleterious genome rearrangements when the sister chromatid is not used as repair template, this additional function of Cdk1 in promoting crossovers can provide another safety mechanism to ensure genome stability.
| DNA double-strand breaks (DSBs) occur spontaneously during DNA replication and after exposure to certain genotoxic chemicals or ionizing radiation. Efficient repair of DSBs can be accomplished by nonhomologous end joining (NHEJ), which directly rejoins broken DNA ends, or by homologous recombination (HR), which utilizes a homologous DNA template to restore the genetic information lost at the break site (reviewed in [1]–[3]). Failure to repair DSBs can lead to genome instability and cell death.
HR is initiated by 5′-3′ nucleolytic degradation of the DSB ends to yield 3′-ended single-stranded DNA (ssDNA) tails. Replication protein A (RPA) binds to the ssDNA tails to remove their secondary DNA structures, but is then replaced by Rad51 aided by Rad52. Once formed, the Rad51 nucleofilaments search for homologous sequences and then promote invasion of the ssDNA into homologous donor double-stranded DNA to form a joint molecule with a displaced strand (D-loop) (reviewed in [1]–[3]). Following strand invasion, the 3′ end of the invading strand primes DNA synthesis using the donor sequence as a template, thus restoring those residues that were lost by resection [4].
According to the canonical double-strand break repair (DSBR) model [5], the displaced strand of the D-loop can anneal with the complementary sequence on the other side of the break (second end capture) to form a double Holliday junction (dHJ) intermediate. Random cleavage of the two HJs is expected to yield an equal number of noncrossover and crossover products. This DSBR model predicts that both crossover and noncrossover products derive from dHJ resolution. However, the finding that most DSB repair in somatic cells is not associated with crossovers [6] led to alternative models for noncrossover generation. In one of them, the action of helicases mediates the convergent branch migration of the two HJs, thus producing a hemicatenane structure that is decatenated to form exclusively noncrossover products [7]–[9]. A second mechanism, termed synthesis-dependent strand annealing (SDSA), leads to displacement of the invading strand that has been extended by DNA synthesis and that anneals with the complementary sequences exposed by 5′-3′ resection [10]–[12]. Because no HJ is formed, only noncrossover products are made. Interestingly, during meiotic recombination, where dHJ resolution into crossovers is essential to drive segregation of homologs to opposite poles, most crossovers are thought to arise via dHJ resolution, whereas noncrossovers form mostly by the SDSA pathway [13], [14].
When a DSB is flanked by direct repeats, its repair primarily occurs by single-strand annealing (SSA). Here, the resected DSB ends anneal with each other instead of invading a homologous DNA sequence (reviewed in [1]–[3]). Subsequent nucleolytic removal of the protruding single-stranded tails results in deletion of the intervening DNA sequence and one of the repeats. In principle, such a break can also be repaired by break-induced replication (BIR), where the repeat closer to the cut site can strand-invade the repeat that is further away and set up a recombination-dependent replication fork to copy all the distal sequences. However, SSA usually out-competes BIR, which is a kinetically slow process [15].
All the above HR pathways require 5′-3′ nucleolytic degradation of DNA ends and the strand-annealing activity of Rad52. In addition, DSBR, SDSA and BIR require the Rad51 protein, which is dispensable for SSA that does not involve strand invasion [16].
In Saccharomyces cerevisiae haploid cells, mitotic HR is generally restricted to the S and G2 phases of the cell cycle, when DNA has been replicated and a sister chromatid is available as an appropriate donor [17], [18]. This cell-cycle specificity depends on cyclin-dependent kinases (Cdks; Cdk1 in S. cerevisiae), which promote resection of the 5′ DSB ends to yield 3′-ended ssDNA tails that are necessary to initiate HR [17], [18]. End resection occurs through a biphasic mechanism: first the MRX complex and Sae2 clip 50–100 nucleotides from the 5′ DNA ends; then Exo1 or Sgs1-Top3-Rmi1 and Dna2 process the early intermediate to form extensive regions of ssDNA (reviewed in [19], [20]). The Sae2 protein has been shown to be a Cdk1 target in promoting ssDNA generation at DNA ends during both mitosis and meiosis [21], [22]. However, as Sae2 only resects a relatively small amount of DNA and other nucleases and helicases are required for efficient DSB resection, Cdk1 likely has additional targets in promoting this event.
Indeed, DSB end resection is also negatively regulated by the Yku heterodimer [23], [24] and by the checkpoint protein Rad9 [25], [26]. Interestingly, the ends of an endonuclease-induced DSB are resected in the G1 phase of the cell cycle (low Cdk1 activity) when Yku is lacking [24]. Moreover, RAD9 deletion allows DSB resection in G2 cells that display low Cdk1 activity due the overexpression of the Cdk1 inhibitor Sic1 [26]. These findings indicate that Cdk1 requirement for DSB resection is bypassed when the inhibitory function of either Yku or Rad9 is relieved.
Whether Cdk1 promotes other HR events is unknown. Some evidence suggests that HR steps other than DSB resection might be regulated by Cdk1 activity. For example, formation of Rad52 foci after ionizing radiation (IR) is less efficient in G1 than in G2, suggesting that Cdk1 might control Rad52 recruitment to DSBs [27]. Furthermore, Cdk1 targets the Srs2 helicase to dismantle D-loop structures, possibly by counteracting unscheduled Srs2 sumoylation [28]. Proteins implicated in late HR events have also been identified as potential Cdk substrates in other eukaryotes. In particular, human BRCA2 is phosphorylated by Cdks, and this phosphorylation has been proposed to negatively regulate Rad51 recombination activity [29]. Moreover, Cdk1-dependent phosphorylation of the fission yeast checkpoint protein Crb2 stimulates resolution of HR intermediates by the topoisomerase Top3 and the ReqQ helicase Rqh1 [30].
Here, we investigate the role of Cdk1 in homology-dependent repair of a DSB. We show that generation of 3′-ended ssDNA at the DSB ends bypasses Cdk1 requirement for the repair of a DSB by either SSA or noncrossover recombination, indicating that Cdk1 is dispensable for these repair events if DSB resection occurs. By contrast, resection is not sufficient to bypass Cdk1 requirement for generating crossover products. Thus, Cdk1 promotes SSA- and noncrossover-mediated recombination by regulating essentially the resection step, while Cdk1 controls further HR steps in order to allow crossover outcomes.
HR is inhibited in G1 when Cdk1 activity is low, whereas it occurs during S and G2/M cell cycle phases when Cdk1 activity is high [17], [18]. Although it is well known that Cdk1 promotes resection of DSB ends [17], [18], [21], it is still unclear if other HR steps are regulated by Cdk1. To investigate whether DSB resection is the only step controlled by Cdk1 in HR-mediated DSB repair, we asked if generation of ssDNA at the DSB ends is sufficient to allow HR when Cdk1 activity is low. As DSB resection in G1 is inhibited by the Yku heterodimer and YKU70 deletion allows ssDNA generation at DSB ends in G1 cells [24], we asked if yku70Δ cells are capable to carry out HR in G1.
Homology-dependent repair of a DSB made between tandem DNA repeats occurs primarily by SSA [15], which requires DSB resection and re-annealing of RPA-covered ssDNA by the Rad52 protein [1], [31]. This process does not involve strand invasion and is therefore independent of Rad51 [16]. We deleted YKU70 in a strain where tandem repeats of the LEU2 gene are 0.7 kb apart and one of them (leu2::cs) is adjacent to a recognition site for the HO endonuclease (Figure 1A) [32]. The strain also harbors a GAL-HO construct that provides regulated HO expression. Since homology is restricted to only one DSB end (Figure 1A), the HO-induced break cannot be repaired by gene conversion, making SSA the predominant repair mode. HO was expressed by galactose addition to α-factor-arrested cells that were kept arrested in G1 with α-factor for the subsequent 4 hours. Galactose was maintained in the medium in order to permanently express HO, which can recurrently cleave the HO sites eventually reconstituted by NHEJ-mediated DSB repair. Kinetics of DSB repair was evaluated by Southern blot analysis with a LEU2 probe that also allowed following 5′-end resection on each side of the break by monitoring the disappearance of the HO-cut DNA bands. The quality and persistence of the cell cycle arrest was assessed by FACS analysis (Figure 1B) and by measuring Cdk1 kinase activity (Figure 1F). Consistent with the requirement of Cdk1 activity for DSB resection and repair, both the 1.8 kb and 3.2 kb HO-cut band signals remained high throughout the experiment in wild type G1 cells (Figure 1C and 1D), where the 2.9 kb SSA repair product was only barely detectable (Figure 1C and 1E). By contrast, the SSA repair product accumulated in yku70Δ G1 cells (Figure 1C and 1E), where both the 1.8 kb and 3.2 kb HO-cut band signals decreased (Figure 1C and 1D). The ability of yku70Δ cells to carry out SSA does not require Cdk1. In fact, Cdk1 activity, which was present in exponentially growing wild type and yku70Δ cells, dropped to undetectable levels after G1 arrest (time 0) and remained undetectable in both cultures throughout the experiment (Figure 1F). Thus, the lack of Yku allows DSB repair by SSA in G1, suggesting that ssDNA generation is sufficient to bypass Cdk1 requirement for SSA.
SSA-based DNA repair requires degradation of the 5′ DSB ends to reach the complementary DNA sequences that can then anneal. If SSA in yku70Δ G1 cells depends on generation of 3′-ended ssDNA at DSB ends, then failure of resection to reach the homologous distal leu2 sequence should prevent SSA. Interestingly, Cdk1-independent resection takes place in yku70Δ cells, but it is confined to DNA regions closed to the DSB site [24], suggesting that other proteins limit extensive DSB resection in the absence of Yku. We therefore asked whether increasing the distance between the complementary leu2 sequences prevented DSB repair by SSA in yku70Δ G1 cells. To this end, we monitored SSA-mediated repair of an HO-induced DSB in a strain where the donor leu2 sequence was positioned 4.6 kb away from the HO recognition site at leu2::cs (Figure 2A) [32]. HO expression was induced in α-factor-arrested cells that were kept blocked in G1 with α-factor in the presence of galactose (Figure 2B). Consistent with previous findings [24], resection in yku70Δ G1 cells was restricted to DNA regions closed to the break site. In fact, the 2.5 kb HO-cut signal decreased more efficiently in yku70Δ than in wild type G1 cells, whereas similar amounts of the 12 kb HO-cut signal were detectable in both wild type and yku70Δ G1 cells (Figure 2C and 2D). Thus, 5′-3′ nucleolytic degradation in yku70Δ G1 cells failed to proceed beyond the distal leu2 hybridization region. The inability of resection to uncover the homologous distal leu2 sequence prevented DSB repair by SSA in yku70Δ G1 cells. In fact, the 8 kb SSA repair product was only barely detectable in both wild type and yku70Δ G1 cells throughout the experiment (Figure 2C and 2E). By contrast, when a similar experiment was performed in G2-arrested cells (Figure 2B), where the inhibitory function of Yku on DSB resection is relieved [33], [34], the 8 kb SSA repair product was clearly detectable in wild type and yku70Δ cells (Figure 2C and 2E), which both showed also a decrease of the 12 kb HO-cut signals compared to the same strains arrested in G1 (Figure 2C and 2D). Thus, the ability of yku70Δ G1 cells to repair a DSB by SSA depends on the extent of resection.
If ssDNA generation were the limiting step in SSA-mediated DSB repair in G1, then increasing the efficiency/extent of resection should enhance the ability of yku70Δ cells to carry out SSA in G1. The lack of the checkpoint protein Rad9 has been shown to allow DSB resection in G2 cells that displayed low Cdk1 activity due to high levels of the Cdk1 inhibitor Sic1 [26]. Thus, we asked whether the lack of Rad9 enhanced the efficiency of DSB resection in yku70Δ G1 cells. To compare resection efficiency independently of DSB repair, we monitored the appearance of the resection products at an HO-induced DSB generated at the MAT locus (Figure 3B) of G1-arrested (Figure 3A) cells, which were not able to repair this DSB because they lacked the homologous donor sequences HML and HMR [23]. As expected, wild type cells showed very low levels of the 3′-ended resection products (r1 to r5), which instead clearly accumulated in both yku70Δ and yku70Δ rad9Δ cells (Figure 3C and 3D). Moreover, the longest r4 and r5 resection products were detectable in yku70Δ rad9Δ cells 120 minutes earlier than in yku70Δ cells (Figure 3C and 3D), indicating that the lack of Rad9 enhances the resection efficiency of yku70Δ G1 cells. Interestingly, although RAD9 deletion was shown to allow MRX-dependent ssDNA generation in Sic1 overproducing G2 cells [26], rad9Δ G1 cells did not show increased efficiency of DSB resection compared to wild type cells (Figure 3C and 3D). Thus, Rad9 limits extensive resection in yku70Δ cells, but its lack is not sufficient, by itself, to escape the inhibitory effect of Yku on DSB resection in G1.
Because DSB resection in G1 was more efficient in yku70Δ rad9Δ cells than in yku70Δ cells, we asked whether the lack of Rad9 allows efficient SSA-mediated DSB repair in yku70Δ G1 cells carrying tandem repeats of the LEU2 gene 4.6 kb apart. Indeed, the amount of SSA repair products in G1 was much higher in yku70Δ rad9Δ cells than in wild type, yku70Δ or rad9Δ cells (Figure 4A–4C). Consistent with DSB resection being more extensive in yku70Δ rad9Δ than in yku70Δ G1-arrested cells (Figure 3), the decrease of the 12 kb HO-cut band signal was much more apparent in yku70Δ rad9Δ than in yku70Δ G1 cells, whereas the 2.5 kb HO-cut band signal decreased with similar kinetics in both G1 cell cultures (Figure 4B and 4D). Cdk1 kinase activity, which was present in all exponentially growing cells, was not required for accumulation of the repair products in yku70Δ rad9Δ cells, as it was undetectable in all G1-arrested cell cultures throughout the experiment (Figure 4E).
SSA requires the strand-annealing activity of the Rad52 protein, but it occurs independently of Rad51 [16]. Consistent with the SSA repair mode, formation of the repair products in G1-arrested yku70Δ rad9Δ cells was abolished by RAD52 deletion (Figure 4F), whereas it was unaffected by RAD51 deletion (Figure 4G). As a DSB flanked by direct repeats could be repaired, at least in principle, also by Rad51-dependent BIR [15], the finding that yku70Δ rad9Δ and yku70Δ rad9Δ rad51Δ G1 cells accumulated the 8 kb repair product with similar kinetics (Figure 4G) indicates that SSA is responsible for this repair event. Thus, we conclude that the lack of Rad9 increases the ability of yku70Δ cells to carry out DSB repair by SSA in G1, likely by enhancing the efficiency of DSB resection.
If competence for SSA-mediated DSB repair relies solely on 3′-ended ssDNA generation, then this repair process should take place with similar efficiency in G1- and G2-arrested yku70Δ rad9Δ cells. As this expectation is based on the assumption that G1- and G2-arrested yku70Δ rad9Δ cells resect DSB ends with similar efficiencies, we compared resection (Figure 5B and 5C) and SSA (Figure 5B and 5D) in yku70Δ rad9Δ cells arrested either in G1 or in G2 (Figure 5A) during break induction. Disappearance of the 2.5 kb and 12 kb HO-cut bands occurred with similar kinetics in G1- and G2-arrested yku70Δ rad9Δ cells (Figure 5B and 5C), which also accumulated similar amounts of the 8 kb SSA repair product (Figure 5B and 5D). As expected, Cdk1 kinase activity was undetectable in yku70Δ rad9Δ cells during the α-factor arrest, whereas it was high in nocodazole-arrested G2 cells (Figure 5E). Thus, DSB resection is the limiting step in DSB repair by SSA.
If SSA is generally restricted to G2 only because high Cdk1 activity allows DSB resection, then inactivation of Cdk1 in G2 should prevent SSA in wild type but not in yku70Δ rad9Δ cells, where DSB resection occurs independently of Cdk1. Thus, we compared DSB repair by SSA in G2-arrested wild type and yku70Δ rad9Δ cells expressing high levels of a stable version of the mitotic Clb-Cdk1 inhibitor Sic1 (Sic1ntΔ) [35]. Consistent with the hypothesis that Cdk1 promotes SSA by regulating the resection step, Sic1 overproduction inhibited SSA repair in G2-arrested wild type cells but not in yku70Δ rad9Δ cells. In fact, the 8 kb SSA repair product accumulated in yku70Δ rad9Δ GAL-SIC1ntΔ cells (Figure 5F and 5G), which showed a decrease of both the 2.5 kb and 12 kb HO-cut band signals (Figure 5F and 5H). By contrast, the same repair product was only barely detectable in G2-arrested GAL-SIC1ntΔ cells, where the HO-cut band signals remained high throughout the experiment (Figure 5F–5H).
When both ends of a DSB share homology with an intact DNA sequence, repair by Rad51-dependent recombination pathways leads to the formation of noncrossover or crossover products. We investigated whether generation of 3′-ended ssDNA can bypass Cdk1 requirement also in this process. To detect crossovers and noncrossovers at the molecular level, we used a haploid strain that bears two copies of the MATa sequence (Figure 6A) [28], [36]. One copy is located ectopically on chromosome V and carries the recognition site for the HO endonuclease, while the endogenous copy on chromosome III carries a single base pair mutation that prevents HO recognition (MATa-inc). Upon galactose addition, the HO-induced DSB can be repaired by Rad51-dependent HR using the uncleavable MATa-inc sequence as a donor. This repair event can occur either with or without an accompanying crossover (Figure 6A) with the proportion of crossovers being 5–6% among the overall repair events [28], [36]. We induced HO expression in α-factor-arrested cells that were kept arrested in G1 in the presence of galactose (Figure 6B). Galactose was maintained in the medium to cleave the HO sites that were eventually reconstituted by NHEJ-mediated DSB repair. The 3 kb MATa band resulting from recombination events that are not associated to crossovers re-accumulated in both yku70Δ and yku70Δ rad9Δ G1 cells, but not in wild type and rad9Δ G1 cells (Figure 6C and 6D). The repair efficiency in both yku70Δ and yku70Δ rad9Δ G1 cells was around 40% after 8 hours (Figure 6C and 6D), reaching 80–90% after 24 hours (data not shown). This finding indicates that the absence of Yku is sufficient for noncrossover HR events to take place despite of the low Cdk1 activity.
Interestingly, the 3.4 kb chromosomal band expected in the experiment above in case of crossover products was not detectable in any G1 cell culture (Figure 6C), suggesting a role for Cdk1 in promoting crossover outcomes that is different from its function in DSB resection. We then compared the products of interchromosomal recombination in G1- and G2-arrested wild type and yku70Δ rad9Δ cells (Figure 7A). As expected, Cdk1 kinase activity remained undetectable in all α-factor arrested cell cultures, whereas it was high in G2-arrested cells (Figure 7B). The overall DSB repair efficiency of G1-arrested yku70Δ rad9Δ cells was similar to that of G2-arrested wild type and yku70Δ rad9Δ cells (Figure 7C and 7D). However, while no crossover events were detectable in yku70Δ rad9Δ G1 cells, ∼4–5% of repair events were associated to crossovers in both wild type and yku70Δ rad9Δ G2 cells, as indicated by the appearance of the 3.4 kb crossover band (Figure 7C and 7E). Thus, yku70Δ rad9Δ G1 cells appear to be specifically defective in generating crossover products. This inability was not due to the absence of Yku and/or Rad9, because similar amounts of crossover products were detectable in wild type and yku70Δ rad9Δ G2-arrested cells (high Cdk1 activity) (Figure 7C and 7E). These results suggest that Cdk1 has a function in promoting crossover recombination that is independent of its role in DSB resection.
If the inability to perform crossover recombination in G1 were due to the lack of Cdk1 activation, then ectopic expression of active Cdk1 should allow crossover recombination in G1, whereas Cdk1 inhibition should prevent crossover formation in G2. We then constructed wild type and yku70Δ rad9Δ strains carrying the system in Figure 6A and expressing a stable version of the mitotic cyclin CLB2 under the control of the GAL promoter (GAL-CLB2dbΔ). This Clb2 variant forms active Clb2-Cdk1 complexes also during G1, because it lacks the destruction box, and therefore it is not subjected to B-type cyclin-specific proteolysis [37]. Strikingly, when both DSB formation and Clb2dbΔ overproduction were induced in G1-arrested cell cultures by galactose addition (Figure 8A), crossover products became detectable in both GAL-CLB2dbΔ and yku70Δ rad9Δ GAL-CLB2dbΔ cells, whereas they were not present in wild type and yku70Δ rad9Δ cells under the same conditions (Figure 8B and 8C).
To assess whether Cdk1 inhibition prevented crossover formation in G2, we compared the products of interchromosomal recombination in G2-arrested yku70Δ rad9Δ and yku70Δ rad9Δ GAL-SIC1ntΔ cells (Figure 8D), the latter expressing high levels of a stable version of the Cdk1 inhibitor Sic1 (Sic1ntΔ) [35]. When both DSB formation and Sic1ntΔ overproduction were induced in G2-arrested cell cultures by galactose addition, crossover products accumulated, as expected, in yku70Δ rad9Δ cells, but they were undetectable in yku70Δ rad9Δ GAL-SIC1ntΔ cells (Figure 8E and 8F). Thus, Sic1-mediated Cdk1 inhibition prevents generation of crossover products in G2, whereas ectopic Cdk1 activation leads to crossover recombination in G1, supporting the hypothesis that Cdk1 activity is required to promote crossover HR events even when DSB resection is allowed by the absence of Yku and Rad9.
HR is highly coordinated with the cell cycle: it takes place predominantly during the S and G2 phases, when the presence of a sister chromatid provides a donor template and high Cdk1 activity promotes DSB end resection to expose ssDNA that is necessary to initiate HR [17], [18], [21], [30]. To study whether Cdk1 plays additional role(s) in HR, we asked whether generation of ssDNA at the DSB ends is sufficient to bypass Cdk1 requirement for HR. Because the lack of either Yku or Rad9 allows Cdk1-independent generation of 3′-ended ssDNA at DSB ends [24], [26], we investigated whether cells lacking Yku and/or Rad9 could repair a DSB by HR when Cdk1 activity is low. We found that DSB repair by SSA can take place in G1-arrested yku70Δ cells. The ability of these cells to carry out SSA in G1 depends on Cdk1-mediated generation of 3′-ended ssDNA at the DSB ends. In fact, the lack of Rad9 increases efficiency of both resection and SSA in yku70Δ G1 cells. Furthermore, Cdk1 inhibition prevents SSA in G2 wild type cells, but not in yku70Δ rad9Δ G2 cells, where DSB resection occurs independently of Cdk1. We also found that G1-arrested yku70Δ and yku70Δ rad9Δ cells can undergo interchromosomal recombination events that are not accompanied by crossovers. Thus, Cdk1 requirement for carrying out SSA and noncrossover recombination is bypassed by DSB resection, indicating that Cdk1 promotes these HR events essentially by regulating the resection step.
Rad52 is essential for both SSA and noncrossover recombination events, while only the latter require the assembly of Rad51 nucleoprotein filaments, which promote homologous pairing and strand exchange (reviewed in [1]–[3]). As the function of Cdk1 in DSB repair by SSA and noncrossover recombination is primarily the regulation of the resection step, neither Rad51 nor Rad52 appear to require Cdk1 activity to exert their biochemical activities.
Interestingly, although RAD9 deletion was shown to allow MRX-dependent DSB resection in G2 cells that overproduced the Cdk1 inhibitor Sic1 [26], the lack of Rad9 did not increase DSB resection or HR-mediated DSB repair in G1 compared to wild type cells. Thus, although Rad9 provides a barrier to resection in yku70Δ G1 cells, its lack is not sufficient, by itself, to escape the inhibitory effect of Yku on DSB resection in G1. This finding is consistent with previous data showing that the resection block imposed by Yku is relieved in G2 [33], [34]. It also indicates that Rad9 prevents DSB resection in all cell cycle phases, but its inhibitory effect in G1 becomes apparent only in the absence of Yku.
Surprisingly, we found that G1-arrested yku70Δ rad9Δ cells are specifically impaired in the formation of crossovers by interchromosomal recombination. Expression of an activated form of Cdk1 allows crossover recombination in both wild type and yku70Δ rad9Δ G1 cells, whereas inhibition of Cdk1 activity in G2-arrested yku70Δ rad9Δ cells prevents crossover formation without affecting noncrossover outcomes. These findings are consistent with a role of Cdk1 in promoting crossover recombination that is independent of its function in DSB resection.
How does Cdk1 promote crossover outcomes? The choice between crossover and noncrossover is tightly regulated [38]. Meiotic recombination results frequently in crossovers [39], while DSB repair in mitotic cells is mostly not associated with crossovers [6]. An explanation of these differences could be that specific mechanisms limit crossovers during mitotic homologous recombination. Indeed, dissociation of the D-loop intermediates gives rise to noncrossover products, and this process is promoted by the helicases Srs2 and Mph1 [7], [28], [36], [40]. Furthermore, noncrossover outcomes can arise also from the dissolution of dHJ intermediates that requires the combined activity of the BLM/Sgs1 helicase, which drives migration of the constrained dHJs, and the Top3-Rmi1 complex, which decatenates the interlinked strands between the two HJs [7]–[9]. One possibility is that Cdk1 promotes crossover recombination by inhibiting proteins specifically involved in limiting crossover generation (i.e. Sgs1, Top3-Rmi1, Srs2 and Mph1). A similar mechanism seems to act during meiotic recombination, where proteins required for homologous chromosome synapsis have been proposed to antagonize the anti-crossover activity of Sgs1 [41]. However, none of the above anti-crossover proteins have been reported to undergo Cdk1-dependent inhibitory phosphorylation. On the other hand, crossovers arise from dHJ intermediate cleavage, which involves the resolvases Mus81-Mms4, Slx1-Slx4, Yen1 and Rad1-Rad10 (reviewed in [42]), suggesting that Cdk1 might promote crossover recombination by stimulating dHJ resolution. Consistent with this hypothesis, the Yen1 and Mms4 resolvases appear to be phosphorylated by Cdk1 [43], raising the possibility that they might represent Cdk1 targets in dHJ resolution. Further studies will be required to assess whether Cdk1-dependent phosphorylation of these proteins has a role in regulating crossover formation.
In conclusion, Cdk1 controls primarily DSB resection to allow SSA and noncrossover recombination, while crossover outcomes appear to require additional Cdk1-promoted events. As mitotic crossovers have the potential for deleterious genome rearrangements, their Cdk1-dependent regulation can provide an additional safety mechanism, ensuring that the rare mitotic recombination events accompanied by crossing over at least occur in S/G2, when a sister chromatid is available as appropriate donor.
Strain genotypes are listed in Table S1. Strains JKM139, YMV86 and YMV45 were kindly provided by J. Haber (Brandeis University, Waltham, USA). Strains YMV86 and YMV45 are isogenic to YFP17 (matΔ::hisG hmlΔ::ADE1 hmrΔ::ADE1 ade1 lys5 ura3-52 trp1 ho ade3::GAL-HO leu2::cs) except for the presence of a LEU2 fragment inserted, respectively, 0.7 kb or 4.6 kb centromere-distal to leu2::cs [32]. Strain tGI354 was kindly provided by G. Liberi (IFOM, Milano, Italy) and J. Haber [28]. To induce a persistent G1 arrest with α-factor, all strains used in this study carried the deletion of the BAR1 gene, which encodes a protease that degrades the α-factor. Deletions of the YKU70, RAD9, RAD51, RAD52 and BAR1 genes were generated by one-step PCR disruption method. YMV86, YMV45 and tGI354 derivatives strains carrying a fully functional CDC28-HA allele at the CDC28 chromosomal locus were generated by one-step PCR tagging method. A plasmid carrying the GAL-CLB2dbΔ allele was kindly provided by R. Visintin (IEO, Milan, Italy) and was used to integrate the GAL-CLB2dbΔ fusion at the URA3 locus in the tGI354 derivative strains. Strain YLL3019, carrying the GAL-SIC1ntΔ allele integrated at the URA3 locus, was obtained by transforming strain tGI354 rad9Δ yku70Δ with ApaI-digested plasmid pLD1, kindly provided by J. Diffley (Clare Hall Laboratories, South Mimms, United Kingdom). The GAL-SIC1ntΔ fusion was cloned into a TRP1-based integrative plasmid that was used to integrate the fusion at the TRP1 locus in the YMV45 derivative strains. Integration accuracy was verified by Southern blot analysis. Cells were grown in YEP medium (1% yeast extract, 2% bactopeptone) supplemented with 2% raffinose (YEP+raf) or 2% raffinose and 3% galactose (YEP+raf+gal).
For Cdk1 kinase assays, protein extracts were prepared as described previously [44]. HA-tagged Cdk1 was immunoprecipitated with anti-HA antibody from 150 µg of protein extracts and the kinase activity in the immunoprecipitates was measured on histone H1 [45].
DSB formation and repair in YMV86 and YMV45 strains were detected by Southern blot analysis using an Asp718-SalI fragment containing part of the LEU2 gene as a probe. DSB end resection at the MAT locus in JKM139 derivative strains was analyzed on alkaline agarose gels as described in [24], by using a single-stranded probe complementary to the unresected DSB strand. This probe was obtained by in vitro transcription using Promega Riboprobe System-T7 and plasmid pML514 as a template. Plasmid pML514 was constructed by inserting in the pGEM7Zf EcoRI site a 900-bp fragment containing part of the MATα locus (coordinates 200870 to 201587 on chromosome III). Quantitative analysis of DSB resection was performed by calculating the ratio of band intensities for ssDNA and total amount of DSB products. DSB repair in tGI354 strain was detected as described in [28]. To determine the amount of noncrossover and crossover products, the normalized intensity of the corresponding bands at different time points after DSB formation was divided by the normalized intensity of the uncut MATa band at time zero before HO induction (100%). The repair efficiency (NCO+CO) was normalized with respect to the efficiency of DSB formation by subtracting the value calculated 2 hours after HO induction (maximum efficiency of DSB formation) from the values calculated at the subsequent time points after galactose addition.
|
10.1371/journal.pcbi.1006821 | The life history of learning: Demographic structure changes cultural outcomes | Human populations show rich cultural diversity. Underpinning this diversity of tools, rituals, and cultural norms are complex interactions between cultural evolutionary and demographic processes. Most models of cultural change assume that individuals use the same learning modes and methods throughout their lives. However, empirical data on ‘learning life histories’—the balance of dominant modes of learning (for example, learning from parents, peers, or unrelated elders) throughout an individual’s lifetime—suggest that age structure may play a crucial role in determining learning modes and cultural evolutionary trajectories. Thus, studied in isolation, demographic and cultural evolutionary models show only part of the picture. This paper describes a mathematical and computational framework that combines demographic and cultural evolutionary methods. Using this general framework, we examine interactions between the ways in which culture is spread throughout an individual’s lifetime and cultural change across generations. We show that including demographic structure alongside cultural dynamics can help to explain domain-specific patterns of cultural evolution that are a persistent feature of cultural data, and can shed new light on rare but significant demographic events.
| Human populations show great cultural variety and complexity, which cultural evolutionary theory seeks to explain by applying ideas about evolution to the ways in which cultural traits change over time. We combined cultural evolutionary theory with information about how people learn over their lifetimes—changing their role models and teachers as they grow up. The result is a new theory of the interaction between life histories and learning that gives a more complete description of human cultural change. The results of our model show why different cultural traits might spread in one population compared to another and how cultural change might spark large-scale demographic changes.
| Cultural transmission can occur via multiple modes of learning; for example, an individual can learn from parents (termed vertical transmission), from non-parental adults (oblique transmission), or from peers (horizontal transmission) [1]. Studies of enculturation and socialization of children suggest that the primary modes of learning change over a lifetime and that children from many different societies learn from parents when young and from peers or other adults as they grow older [2]. Further, the extent to which other modes of learning supersede vertical transmission can vary among populations [3]. Differences in how and when people learn and teach [4] cultural traits, such as the use of a specific tool or the moves of a particular dance, can reflect these culturally preferred modes of learning. Coarse-scale differences in social learning may result from variation in subsistence strategies; for example, Hewlett et al. [5] investigated the different learning trajectories of the Aka hunter-gatherers in central Africa and their small-scale agriculturalist neighbors. They showed that the individuals from whom children learn differed by children’s age and by their group’s subsistence strategy. In the hunter-gatherer groups, children’s learning was predominantly vertical, especially before the age of 12, whereas children of small-scale agriculturalists learned primarily horizontally and obliquely, beginning at a much younger age. These learning life histories are illustrated in Fig 1.
Such population-specific learning patterns might be influenced by, and in turn influence, the human ecological niche. Through the hunting of game, the construction of large-scale settlements, the domestication of animals, and the spread of agriculture, humans have profoundly modified their ecological niche. This process is known as niche construction and encompasses alterations to the environment that can influence the selection pressures on future generations of humans and other species [6–9]. Similarly, some culturally transmitted behaviors may alter the evolutionary pressures on other cultural and/or genetic traits in a process termed cultural niche construction [9–12]. Social structure and social organization have a significant but understudied effect on the transmission of cultural information [13,14]. Similarly, by significantly altering associations between different population members, the cultural niche defined by a subsistence strategy appears to be associated with differences in the predominant mode of transmission (vertical, oblique, or horizontal) of cultural traits to the young [5]. In addition, the style of learning could, itself, be considered a cultural niche that determines the types of traits that are learned and the rate at which they spread.
In this paper, we develop age-structured models of cultural transmission to investigate the effects of different ‘learning life histories’ (sensu [3]) on cultural evolution, defining these learning life histories as ‘population-level life-stage differences in modes of learning’. We use observed learning life history differences between hunter-gatherers and small-scale agriculturalists to inform the parameters of the model. We then speculate about the potential importance of these different trajectories in transitions from foraging to farming and from predominantly vertical to predominantly horizontal learning. Our analysis explores how different modes of learning can affect the cultural evolution of a number of traits, including a fertility-enhancing trait (for example a farming practice that improves food stability), a fertility-decreasing trait (for example a small-family norm), and a trait controlling the mode of learning itself.
We extend an earlier age-structured model of cultural transmission [15] and then simplify that model for application to data on the early learning practices of the Aka hunter-gatherers and their small scale agriculturalist neighbors, the Bofi and agriculturalist Aka populations [5]. To that end, we characterize learning in terms of learning opportunity, replacing the absolute probability of choosing a role model from a particular age class, which is difficult to assess in reality, with the probability of spending time with individuals from that age class, which can be readily measured. Thus, we make the assumption that the amount of time spent with an individual is correlated with the amount learned from that individual. It is important to note, however, that the proportion of learning reported from parents in Aka society (~80% [16]) is considerably higher than the average proportion of time (~48%) that young children spend within arm’s reach of their parents and other adults, although it is more similar to the time they spend with adults and in mixed groups of adults and children (~85%) [5], implying that vertical learning might happen both when children are alone with their parents and when they are with mixed groups of parents and peers. We use an age-structured model of cultural evolution to investigate whether the distinct learning niches exemplified by these hunter-gatherer and agriculturalist groups lead to qualitatively different evolutionary dynamics.
Our model is based on the age-structured framework developed by Fogarty et al. [15], which included three age classes, with fertility and survival depending on the transmission of a cultural trait T. Here we consider five age classes. These five age classes loosely represent infancy, early childhood, late childhood, adulthood, and post-reproductive life. The post-reproductive age class represents the elderly or grandparents, who may act as reservoirs of cultural information that can be transmitted to younger age classes. We define the modes of learning in the population as vertical (learning from parents only), oblique (learning from any individual in an older age class), and horizontal (learning from members of one’s own age class only). Individuals may learn throughout their lives until they reach the post-reproductive age class (age class 5).
Suppose that the five discrete age classes, Ai, are of size ni (with total population size, n=∑i=15ni) with proportions ai=nin in each age class. The population age structure changes from generation τ to generation τ+1 in accordance with the adapted Leslie matrix, L, which describes life stages rather than age, specified in Eq (1) below.
(n1n2n3n4n5)τ+1=(f1f2f3f4f5s100000s200000s300000s4s5)︸L(n1n2n3n4n5)τ
(1)
In Eq (1), the number of individuals in a given age class in generation τ+1 is given by multiplying L by the vector containing the age class numbers in generation τ. Parameters fi represent fertilities, and the number of individuals in age class 1 at generation τ+1, for example, is given by n1,τ+1 = f1n1,τ+f2n2,τ+f3n3,τ+f4n4,τ+f5n5,τ. In all of the following analyses, however, we assume that only age class 4 can reproduce; that is, f1 = f2 = f3 = f5 = 0, and thus n1,τ+1 = f4n4,τ. Parameters si represents survival probabilities; for example, at time τ+1 the number of individuals in age class 2 is given by n2,τ+1 = s1n1,τ, i.e. the proportion of age class 1 individuals who survive to age class 2 multiplied by the number of age class 1 individuals at time τ. The recursions for other age classes follow similarly from Eq (1).
For this age-structured population, we consider a cultural trait T that has two variants, denoted by T and t. The frequency of T in age class i at time τ is xi,τ. Individuals learn throughout their lifetimes; initially they learn vertically from their parents and subsequently from parents, unrelated adults, or peers with likelihoods that may differ, for example, depending on subsistence strategy (see below). The probability that an individual has the cultural trait T after vertical learning is, therefore, the probability that its parent had T multiplied by a probability, pv, that the individual learns vertically from its parent. The parameter pv represents the effectiveness or fidelity of vertical learning from parent to child.
Although many cultural traits may be neutral (for example, certain decorative elements; see, e.g. [17]), some (such as norms about reproduction, marriage or childcare) may have a profound effect on demography. For example, a cultural trait that increases fertility may also increase access to high quality food or increase reproductive output in some other way. To model this, we assume that the reproductive age class (age class 4) has a baseline fertility b (associated with the cultural variant t), which increases to b+wf in T individuals, where wf is a fertility increase; that is, an increase in the number of offspring associated with T. The fertility f4 in Eq (1) is then given by
f4=(b+wf)x4,τ+b(1−x4,τ)
(2)
Since age class 4 is the only one that reproduces (i.e. f1 = f2 = f3 = f5 = 0), age class 4 represents the parents of the new individuals in age class 1. Therefore, the frequency of T in age class 1 at time τ is given by
x1,τ=(b+wf)x4,τ−1(b+wf)x4,τ−1+b(1−x4,τ−1)pv
(3)
From Eq (1), at time τ, individuals in age class i survive to age class i+1 at time τ+1 with probability si. Individuals may learn from their parents at the first learning opportunity, but may also learn from individuals of their own generation or an older generation at subsequent learning opportunities throughout their lives. Therefore, the proportion of age class 1 that has cultural variant T after the first learning event is given by Eq (3), and the frequencies of T in age classes i = 2,3,4 at time τ+1 after horizontal and oblique learning are given by
xi,τ+1=xi−1,τ+(1−xi−1,τ)(Vxparentpv+(1−V)ph∑y=iωny,τxy,τ∑z=iωnz,τ),
(4)
where V is the proportion of learning at age or stage i that is vertical, xparent is the proportion of the surviving parental population that has T, ph is the probability that a t individual acquires T horizontally or obliquely from contact with a T individual, pv is the probability that a t individual acquires T vertically, and ω represents the oldest age class from whom one can learn non-vertically. For the analyses presented here ω = 5, but in principle this need not equal the number of age classes in the model. For example, if a certain age class, i, prefers to learn from those slightly older and more experienced, ω might be i+1, restricting learning interactions to be between those in age class i and i+1 only. If the trait in question has an effect on survival as well as fertility, xparent must take account of different rates of survival in parents with T and parents without (see below). In the following analyses, the vertical and horizontal transmission rates pv and ph are set to be equal in the hunter-gatherer and agriculturalist models, except where otherwise stated.
We define life history learning strategies according to the amount of learning in each age class that is vertical, horizontal, or oblique. The age of the learner defines from whom the learning takes place: for example, Fig 1A shows a case where only vertical learning is used in the first three age classes and only horizontal or oblique learning is used in the final two age classes. As reference points, we include two extreme cases: an all-vertical learning strategy where individuals learn only from their parents throughout their lives, and an all-horizontal strategy where, after an initial round of vertical learning in infancy, individuals learn exclusively from same-age peers. We can then situate other learning strategies (where the time a child spends learning from parents or same-age peers, for example, has been observed and recorded for specific populations; Fig 1) between these extremes and examine the consequences of mixed strategies for a population’s cultural evolution and demography.
We assume that after age class 4, adults move to age class 5 where they do not learn. T individuals in the oldest age class, 5, at time τ+1 include both those maturing into group 5 from group 4 and survivors in group 5 from time τ; specifically,
x5,τ+1=x4,τs4n4,τs4n4,τ+s5n5,τ+x5,τs5n5,τs4n4,τ+s5n5,τ
(5)
The model described above can be extended to allow the proportion of time the young spend learning from other age classes to depend on the frequency of T in the reproductive population (i.e. age class 4), for example, through a cultural norm. In this way, T could affect reproductive output for T individuals, as well as the time that both T and t adults spend interacting with younger age classes through changes in diet or food stability for the former, and altered time budgets for the latter [18]. For example, in the Bofi, agricultural practices increase food stability but reduce the time that children spend with their parents while those parents are farming [5], which could increase the likelihood of oblique or horizontal learning. We use the model in Eq (6) below to investigate this case and to assess the implications for the evolution of traits such as farming. To these ends, let the cultural trait T affect fertility (as in Eq (2)) and simultaneously affect some population-wide norms related to child rearing. As the frequency of T in the adult population increases, individuals change their views or practices and spend less time teaching their offspring. Thus, an increase in frequency of T in the population leads to a decrease in the time spent with offspring regardless of the actual form of the trait carried by an individual’s parents, which is, however, taken into account when considering the probability of learning (Eq (7)).
To do this, we introduce vj,τ, which, at time τ, represents the proportion of time individuals from age class j spend learning T from their parents, who were in age class 4 when j individuals were born. vj,τ is given by
vj,τ=vb(1−x4,τ−jϵ),
(6)
where vb is the baseline amount of time that age class j individuals spend interacting with their parents who were in age class 4 at time τ−j, ϵ is a scaling factor that determines the strength of the effect of the cultural trait on time spent with offspring, and x4,τ−j is the frequency of T in the reproductive age class 4. Here vj,τ represents the time a child spends in the presence of its parent, not the rate of learning from them, although the former may be used as a proxy measure for the latter.
In Eq (6), as the frequency of T in age class j’s parental generation (x4,τ−j) increases, the amount of vertical contact decreases. We can then modify Eq (4), combining it with the formula for vertical learning (Eq (3)) to incorporate vj,τ. The equation for xj,τ+1 becomes:
xj,τ+1=xj−1,τ︸(A)+(1−xj−1,τ)︸(B)(vj,τ+1(b+wf)x4,τ−j(b+wf)x4,τ−j+b(1−x4,τ−j)︸(C)pv+(1−vj,τ+1)∑z=jωnz,τxz,τ∑z=iNnz,τ︸(D)ph),
(7A)
where the trait T increases fertility, or
xj,τ+1=xj−1,τ︸(A)+(1−xj−1,τ)︸(B)(vj,τ+1(s4+ws)jx4,τ−j(s4+ws)jx4,τ−j+s4j(1−x4,τ−j)︸(C)pv+(1−vj,τ+1)∑z=jωnz,τxz,τ∑z=iNnz,τ︸(D)ph),
(7B)
where the trait T increases survival and ws is a survival increase associated with T. Here, j is the focal age class (j = 2,3,4), and x4,τ−j is the frequency of T at the time of reproduction by the age class containing the parents of age class j. In Eq (7A and 7B), term (A) represents the individuals who survived to generation τ+1, from age class j–1, who are now in age class j and have already learned T (note that here T does not affect survival rates). (B) represents the members of age class j–1 who did not learn T in the previous time step, and so have another chance to learn. These individuals can learn vertically, with probability vj,τ+1, which is multiplied by (C), an expression adapted from Eq (4), which describes the frequency of T in the parental generation j generations ago (x4,τ−j) and ensures that only individuals with T parents can learn T vertically. Finally, individuals can learn horizontally or obliquely with a probability 1−vj,τ+1, and (D) represents the frequency of T, either in the individual’s own age class (horizontal learning) or in older (oblique learning) weighted by the population size in that age class. If the cultural trait affects survival and not fertility, the survival probabilities in the matrix in (1) will become si(1−xi)+xi(si+ws) where the survival probability of T individuals is increased relative to t individuals by an amount ws. Here we must also consider the differential survival of parents with T and with t when assessing vertical learning as in the case of fertility increases (Eq 7B).
In this formulation, children may learn from any individual they contact and the frequency of T in the population determines the percentage of time offspring spend with their parents, with others of the same age, or with other adults. This is likely to be a simplification of human learning processes; for example, Henrich and Broesch [19] reported that small-scale agriculturalist communities in Fiji show adaptive learning biases (e.g. prestige-biased social learning). We suggest that the evolution of a trait that allows children to learn from any individual, not just from their parents, may allow for and promote the evolution of such biases by widening the pool of potential role models.
Here, we assess the effects of different life histories of learning on cultural transmission and demography. We estimate these life histories of learning using the learning parameters suggested by Hewlett et al.[5] for hunter-gatherers and agriculturalists along with two reference strategies: one in which horizontal and oblique learning are used late in life and one in which they are used early (Fig 1). Using these different proportions of vertical and horizontal/oblique transmission at different life stages, we compare the effects of life histories of learning on cultural evolutionary processes. The model described in eqns (1–7) (see code) was iterated over a number of generations with the frequency of T in each age class (xi,τ), the number of individuals (n) in the population at time τ, and the proportions of the population in each age class (ai,τ) evolving simultaneously. The number of generations and other parameter values for each simulation are given in the corresponding figure caption. At the beginning of each simulation, the t form of trait T was close to fixation in the population, and T appeared at very low frequency (0.005 for analyses shown here). Fig 2 shows the final frequency (after 5,000 model iterations) of the T trait for each of these learning life histories in turn. The axes show the rates of horizontal/oblique and vertical learning of the cultural trait. As the rate of vertical learning increases, the final frequency of the trait depends on the rate of the dominant mode of learning to a greater degree. For example, if the rate of horizontal learning is greater than that of vertical learning, the final frequency of the trait depends to a large degree on the rate of horizontal learning.
In the model described by Eq (7), it is important to note that the trait being transmitted has a baseline advantage when it is transmitted vertically (or via ‘late horizontal’ life histories). We assume that after age class 4, adults move to age class 5 where they do not learn. The frequency of T is, therefore, highest in age classes 4 and 5, as individuals continue to learn throughout their lives up to that point. Relying only on parents as role models means that the frequency of T in the subpopulation from which an individual learns will be higher than the average in the population as a whole. Horizontal learners use the average rate of learning across all age classes and so adopt T at a slower rate. Fig 3 accounts for this phenomenon by showing the increase in the spread of the trait due to the trait’s fitness benefit compared to a neutral trait. In this way, we can show that interaction between the exact form of the fitness of a trait and the dominant mode of learning can be crucial.
For example, if the fitness benefit of a particular trait to an individual is an increase in fertility, a learning life history that relies heavily on vertical learning allows the cultural trait to gain a stronger advantage (Fig 3A). This increase occurs because a fertility benefit increases the number of offspring born to T individuals (knowledgeable individuals). By learning only from parents, these offspring increase the chance that they learn from a knowledgeable individual compared to sampling from the population as a whole. On the other hand, Fig 3B shows that when the trait increases survival in all age classes, the trait becomes overrepresented in the population as a whole, not just in parents. Therefore, sampling from the population increases the chance that an individual will choose a knowledgeable role model. Note that these results hold for initial age structures that are either even (as shown in Fig 3) or pyramidal (e.g. S1C and S1D Fig) but break down for extremely skewed initial structures (e.g. S1A, S1B, S1E and S1F Fig). This result underscores the importance of the life history of learning to both cultural and biological evolution: different types of traits are favored depending on the modes of their learning at different life stages, and these learning life histories can influence both cultural dynamics and population demography.
We can also investigate the evolution of these learning life histories in the case that the focal cultural trait (T) produces both a fertility increase and a change in the proportion of time children spend in the presence of their parents and peers. Thus T constructs a learning niche, altering the conditions of its own spread as it invades a population [5]. Such conditions might be characteristic of child-rearing practices in small-scale agriculturalist populations, and similar changes to time allocation may accompany the transition from subsistence, that is predominantly hunting and foraging, to predominantly farming.
Eq (7) describes a niche-constructing trait that alters the proportion of learning that is vertical relative to horizontal or oblique. Here we do not distinguish a priori between hunter-gatherer populations and agriculturalist populations. The trait is allowed to spread in the population by affecting both fertility and learning norms. We then examine the effect of the strength of changes in learning norms (namely, the strength of cultural niche construction) associated with the trait, which is determined by the parameter ϵ in Eq (6). First, we consider a case where the cultural trait increases fertility (i.e. is fitness enhancing) and also increases the amount of horizontal or oblique learning that occurs. This is roughly analogous to the suggested consequences of the successful invasion of agriculture [5]. Fig 4A shows that in the absence of niche construction (i.e. when ϵ = 0), the cultural trait can increase in frequency and persist in the population if the vertical learning rate is high or if both vertical and horizontal learning rates are high. Fig 4B shows that when cultural niche construction is strong (ϵ = 1), the cultural trait T supports its own transmission under some circumstances. Taking a closer look at the quadrants in Fig 4B (delineated by black dashed lines), it is clear that this phenomenon rests on the balance between vertical and horizontal transmission and on the efficacy of both modes of transmission. In the upper right quadrant of both panels, both vertical and horizontal learning are effective and the trait rises to very high frequencies at some points and fixes in the population regardless of the strength of cultural niche construction. However, if the rate of one mode of learning is higher than that of the other (ph>pv, upper left quadrant or pv>ph, lower right quadrant), increasing reliance on the higher-rate mode has the effect of increasing the parameter space over which the trait can spread. Fig 5 shows the same for a trait that decreases fertility.
We modeled the spread of a cultural trait that can affect demography (in particular, fertility and survival) as well as the spread of cultural norms. In turn, this trait affects age structure and population size, and it also influences the life history of learning; that is, when and from whom individuals learn. Agriculture is an important example of such a trait. Farming can increase rates of fertility through increased access to a stable supply of nutrition and as they spread, agricultural practices, such as tending crops, may affect the importance of vertical, horizontal, and oblique learning modes by altering the amount of time children spend in the company of their parents or in groups of same-age peers. As a case study of these differences, we refer to data from Hewlett et al. [5] who recorded the time children spent with members of different age classes in neighboring populations of foragers and agriculturalists.
Such cultural traits are likely to have played a central role in human history. Each new mode of production (for example, foraging, small-scale agriculture, intensive agriculture) might have led to important changes in lifestyle and, as a result, in the time children spend with parents, same-age peers, or unrelated adults. Our analysis addresses three important questions: 1. how do primary modes of learning (i.e., vertical, horizontal or oblique learning) affect the rate of spread and accumulation of cultural traits in populations with different learning life histories? 2. Do these changes mean that certain types of traits are more likely to spread in some societies than in others? Could hunter-gatherers, for example, accumulate or maintain some cultural traits that farmers would not, as a result of their dominant modes of learning? 3. What are the implications of these phenomena for demographic changes over time?
There is some evidence that modes of learning affect the diversity and composition of cultures. Guglielmino et al. [2] describe vertical learning and one-to-many transmission as being conservative, while horizontal and oblique learning are more likely to support the spread of innovations [1,20] However, age-dependent cultural transmission is arguably more common than the dominance of just one mode of learning throughout an individual’s life. For example, age-structured patterns of social learning have been observed for the Aka and Bofi in the Central African Republic [5] and for a number of horticultural and foraging societies in the Democratic Republic of Congo (DRC) [3].
Such age-structured learning strategies suggest that a model accounting for life stage is appropriate for human cultural transmission. In our age-structured model, we first show that different age-dependent learning strategies result in strikingly different patterns of cultural evolution. For example, Fig 2 shows that strong reliance on vertical transmission, as seen in hunter-gatherer populations, entails that the spread of a cultural trait relies on the rate of vertical transmission to a greater extent than on the rate of horizontal or oblique learning. It is noteworthy that in many previous verbal and mathematical models [1,21,22], vertical learning is regarded as conservative, with new traits failing to spread as widely or as rapidly as they would if they were transmitted horizontally or obliquely. However, our model highlights another aspect of reliance on vertical transmission that is difficult to resolve without knowledge of a population’s age structure, namely, those who learn primarily from parents are learning from the most knowledgeable subset of the population. In our model, as in most real-world populations, individuals learn throughout their lives, and as a result the proportion of the population that is well informed about useful cultural traits increases with age. The most knowledgeable age classes in our model are, therefore, age classes four and five. Relying on the knowledge possessed by younger age classes, as seen with horizontal transmission, reduces the overall probability of learning from a knowledgeable individual and curtails spread of culture relative to individuals who learn only from parents. In this context, optimal learning of important cultural traits would rely heavily on parental knowledge and thus on vertical transmission. These effects would be weakened by extremely fast environmental change that renders cultural information obsolete at a fast rate. This is not modeled here but see [18].
With the advent of farming, parents may spend less time with young children and more time engaged in agricultural activities, thus reducing reliance on vertical learning and increasing the importance of horizontal and oblique modes of learning. In this way, farming can be viewed as a trait that constructs a ‘learning niche’: that is, farming is a cultural trait that can change the way cultural traits are transmitted. We might also expect to see the evolution and spread of further traits or norms or even modes of transmission [23] that compensate for reduced time with offspring–for example many-to-one transmission. In the example shown in Figs 2 and 3, the proportion of time spent with each age group was estimated from the anthropological literature; in this niche-constructing example, however, we allow this proportion of time to vary with the frequency of the trait T. We can thus utilize this scenario to reflect the early evolution of a trait such as farming. The analysis showed that if the niche constructing effect is strong—the trait has a strong effect on learning norms and practices—it can facilitate its own spread and expand the parameter range over which it can be expected to increase in frequency, as well as substantially increasing the frequency of the trait at equilibrium under certain conditions (Fig 4). If the transmission becomes one-to-many, these effects would be more pronounced [1,22]. Although we expect the evolution of farming to increase fertility, the effects of cultural niche construction on learning norms in a population might also reduce fertility in some cases, which would thus act in opposition to natural selection. In other words, a niche-constructing trait can promote its own spread even if its effect on fertility is negative (Fig 5). For example, the spread of an education system, such as classroom-based learning, would likely change the learning niche by increasing oblique and horizontal learning, but has also been shown to decrease the desired number of children, and hence fertility [10,24].
In real-world populations, there are likely to be domain-specific differences in the transmission of information. For example, among undergraduate students at Stanford University, traits like religious beliefs and political inclinations were over 80% vertically transmitted but preferred forms of entertainment were over 60% horizontally or obliquely transmitted [25]. Further, in the Lese, Mamvu, Budu and Bila cultural groups in the DRC, sexual health practices were predominantly horizontally transmitted between adults [26]. Our model begins to address this interaction between learning strategies and knowledge domain by making explicit the effect of life history of learning on the effectiveness of cultural spread for traits with different types of fitness effects.
Finally, the model reveals rare but potentially important large-scale demographic differences between populations with different learning modes, especially when learning cultural traits that alter fertility or survival. Fig (6) shows an example of the spread of a fertility enhancing cultural trait in a small and precarious population. As discussed above, a population using predominantly vertical transmission can spread fertility-enhancing cultural traits more rapidly and more effectively than populations in which horizontal transmission is more important. Under certain circumstances, this advantage in terms of cultural transmission translates into a cultural demographic rescue for a population in which a particular mode of learning is important (Fig 6A) and the slower rate of spread results in demographic collapse in the other (Fig 6B). This shows not only that cultural traits can affect demography, but also the learning norms within a population and the way in which individuals choose to learn could have major effects on a population’s evolutionary and demographic trajectory.
The transition from foraging to farming, as described by Hewlett et al. [5], is accompanied by a change in learning mode; parents spend more time farming and their children spend more time with other children. By contrast, in foraging populations children accompany their parents as they gather food. This shift in the focus of learners from their parents to others in the population may allow individuals to actively choose a cultural role model and pave the way for emergence of cultural learning biases such as prestige bias or conformity bias, which may not be a prominent feature of societies that rely primarily on vertical learning, but are widespread in other societies [19]. Thus, cultural niche-constructing traits that affect the mode and rate of their own transmission may underpin the evolution of more varied and less obvious social learning biases. These, in turn, may have facilitated effective rapid cumulative cultural evolution and driven further changes in subsistence strategy, population size, or population age structure, profoundly affecting the cultural and biological evolutionary trajectories of human populations.
|
10.1371/journal.pgen.1000503 | Caenorhabditis elegans Genomic Response to Soil Bacteria Predicts Environment-Specific Genetic Effects on Life History Traits | With the post-genomic era came a dramatic increase in high-throughput technologies, of which transcriptional profiling by microarrays was one of the most popular. One application of this technology is to identify genes that are differentially expressed in response to different environmental conditions. These experiments are constructed under the assumption that the differentially expressed genes are functionally important in the environment where they are induced. However, whether differential expression is predictive of functional importance has yet to be tested. Here we have addressed this expectation by employing Caenorhabditis elegans as a model for the interaction of native soil nematode taxa and soil bacteria. Using transcriptional profiling, we identified candidate genes regulated in response to different bacteria isolated in association with grassland nematodes or from grassland soils. Many of the regulated candidate genes are predicted to affect metabolism and innate immunity suggesting similar genes could influence nematode community dynamics in natural systems. Using mutations that inactivate 21 of the identified genes, we showed that most contribute to lifespan and/or fitness in a given bacterial environment. Although these bacteria may not be natural food sources for C. elegans, we show that changes in food source, as can occur in environmental disturbance, can have a large effect on gene expression, with important consequences for fitness. Moreover, we used regression analysis to demonstrate that for many genes the degree of differential gene expression between two bacterial environments predicted the magnitude of the effect of the loss of gene function on life history traits in those environments.
| Transcriptional profiling is often used to identify genes that are differentially regulated in response to different environments. These experiments assume that genes differentially expressed in response to different environments are functionally important and, furthermore, that the degree of differential gene expression is predictive of the magnitude of functional importance. In genetic experiments, function is inferred from analyzing the phenotypes of removing, reducing or altering gene function. However, to date, there has not been a specific test of how well the degree of differential gene expression between two (or more) environments is predictive of gene function. Here we identified C. elegans genes that were differentially expressed in response to different bacterial environments and determined the phenotypic differences of life history traits between these environments using mutant strains that compromised gene function. We found that differential gene expression is indeed predictive of functional importance of the identified genes in different environments. This observation has important implications for interpreting the results of transcriptional profiling experiments of populations of organisms in their native environments, where in many cases the genetic tools to disrupt gene function have not yet been fully developed or interfering with gene functions in nature may not be feasible.
| We are interested in understanding the genetic mechanisms that underlie organismal responses to their environment, especially in light of human induced environmental change. To begin to address this challenge have chosen to model the interaction of soil nematodes and their environment in the laboratory using an ecological genomic approach. Nematodes play key roles in many ecosystems including nutrient cycling, turnover of microbial biomass [1] and decomposition of soil organic matter [2],[3]. In fact, bacterial-feeding soil nematodes are among the most abundant invertebrates in soils and are important members of grassland belowground food webs [4]. In addition, many bacterial-feeding taxa have been shown to be among the most responsive of nematode trophic guilds to various disturbance regimes [5]–[7]. These responses include shifts in the species composition of bacterial-feeding nematode assemblages, resulting in altered community structure and, presumably, function. For example the increased relative abundance of opportunistic Rhabditidae species is characteristic of a response to resource pulses caused by disturbance or changing land management practices [1],[8]. The effects of disturbance, can be direct, such as changes in chemical and physical properties of the soil that impact nematode movement or life history, or indirect, such as changes in other biotic components (e.g., food source) that affect the nematode community. Here we begin to address the genetic basis of one such indirect effect. Recent studies have demonstrated that the grassland soil bacterial (KLJ, JDC, MAH, unpublished) and bacterial-feeding nematode communities [5] on the Konza Prairie Biological Station responded to various disturbance treatments with species-specific responses, indicating that indirect causes through bottom-up effects of the responses of the bacterial-feeding nematode community are plausible. Other recent studies conducted in other ecosystem types have demonstrated microbial community responses to disturbances comparable to the bacterial community response we observed on Konza prairie. For example, tilling for agriculture [9],[10], burning of aboveground biomass [11]–[13] and the addition of nitrogen amendments [14]–[16] can cause changes in the relative abundance of bacterial species and microbial community diversity. Thus, the differential bacterial community response to perturbation, in conjunction with nematode food preference [17] and/or pathogenic interactions [18] with bacteria could drive the observed changes in nematode community structure. Therefore, we have focused on the genomic responses of microbial-feeding nematodes to the possible changes within the grassland microbial environment. We hypothesize that an examination of the genomic response of nematodes to different bacterial environments may reveal the genetic basis of the observed nematode community response.
We are interested in the genetic responses of native ecologically relevant nematodes that do not have well developed genetic and genomic resources and thus are not tractable for functional studies. Specifically, we have identified several members of the Rhabditididae (Mesorhabditis, Oscheius and Pellioditis) and Cephalobidae (Acrobeloides and Acroboles) families whose relative abundance is altered in response to nutrient additions. Thus, we have turned to a laboratory modeling approach using a related, genetically tractable, bacterial-feeding soil nematode to identify conserved candidate genes. Many groups have analyzed the transcriptional response of the genetically tractable nematode Caenorhabditis elegans to various, usually medically significant, bacteria [19]–[22], in order to model human innate immunity [18],[23]. While the high degree of evolutionary conservation allows C. elegans to be a good model for human-bacteria interactions, it may be an even better model for bacterial-feeding nematode responses to bacteria. Although C. elegans is not found in abundance in the grassland soils under study [5], it is related to many of the relevant nematode taxa of interest. Therefore, we exposed C. elegans to different grassland soil bacteria and used transcriptional profiling to identify differentially expressed genes. We determined the functional significance of a subset of the differentially expressed genes by measuring fitness and lifespan of mutant nematodes in the various bacterial environments. Our results demonstrate that the functions of many of the genes specifically induced in response to different bacteria contribute to nematode fitness and lifespan in those bacterial environments. Furthermore, for specific genes, the extent of differential gene expression between bacterial environments was correlated with the degree of the effect of mutations in those genes on life history traits in those environments. Thus we propose that examination of differential gene expression in different environments allows for prediction of degree of mutational effects of those genes in those environments. Thus, here we show the first evidence, to our knowledge, that there is indeed good predictive power for the effects of mutant phenotypes in an environment-specific manner, suggesting that the relative level of transcription can be informative about the relative contributions to function, at least for life history traits. Additionally, the examination of C. elegans gene function in new environments has uncovered new phenotypes for previously studied genes as well as genes that had not been shown to have obvious phenotypes under standard laboratory conditions, perhaps adding to our understanding of the C. elegans genome.
We are interested in understanding naturally occurring nematode-bacterial interactions of native soil nematodes. Since these nematodes do not have well characterized genomes or genetic tools, we have used C. elegans as a model to discover conserved genes involved in these interactions. For this purpose, we isolated bacteria from grassland prairie soils at the Konza Prairie Biological Station. Although C. elegans has not been found at the Konza prairie, related nematodes from the same family (Rhabditidae) are found there, thus it should be a suitable model nematode. Micrococcus luteus was the most abundant bacterial species in the nutrient amendment plots (supplemented annually with 10 g/m2 ammonium nitrate for 21 years) that was culturable on nematode growth media (NGM) plates (data not shown). Nematode growth media was used for bacterial isolation, as growth on NGM was a requirement of the experiment. Bacillus megaterium and Pseudomonas sp. were isolated in association with Rhabditid nematodes from Konza prairie soils (Oscheius sp. and Pellioditis sp. respectively) [5],[24]. Bacteria were isolated by extracting nematodes from the soil followed by thorough washing to remove bacteria weakly associated with the nematode cuticle. Nematodes were then placed on NGM plates and allowed to defecate surviving ingested bacteria. Although, this method of bacterial isolation makes it likely that bacteria came from nematode intestines, we cannot rule out that bacteria were associated with native nematode cuticles. Thus, these were termed nematode associated bacteria. The Pseudomonas sp. we isolated was most similar to Pseudomonas fluorescens with 98% sequence identity (Ribosomal Database Project) in the 16S rDNA sequence (See Methods).
Wild-type C. elegans (N2) was grown on the three prairie bacterial species as well as E. coli (OP50) which served as a control, as it is the typical laboratory diet for C. elegans [25]. The different bacteria served as food sources for C. elegans as well as the immediate environment during growth as the culture plates contained bacterial lawns. Therefore, the effects of the external features of the different bacterial lawns (e.g. oxygen concentration in the bacterial lawn, bacterial viscosity and potential bacterial secretions) on nematode physiology could not be distinguished from the effects of ingestion of the bacteria and will hereafter be collectively referred to as ‘bacterial environment’. To get a more accurate estimate of the effects of different bacterial environments on nematode fitness, we used life tables to estimate absolute fitness (λ), which accounts for age specific fecundity (mx) and survival (lx), as well as generation time (T) [26] and is subsequently more comprehensive than brood size alone (see Methods). The absolute fitness of wild-type animals differed significantly in the different bacterial environments. Animals displayed the highest fitness when grown on Pseudomonas sp. (λ = 3.99), which was significantly greater (P = 0.021) than when grown on E. coli (λ = 3.60), B. megaterium (λ = 2.81, P<0.0001) and M. luteus (λ = 2.63, P<0.0001). Fitness of wild-type animals on E. coli was also significantly higher than on either B. megaterium (P = 0.027) or M. luteus (P = 0.027; Figure 1A, Table 1). It is interesting to note that the only previous study to use life tables to calculate fitness in C. elegans [27] found highly similar values (λ = 3.85, with growth on E. coli OP50).
Lifespan is another important aspect of nematode demography. Lifespan is measured here as time to death for 50% of a population (TD50) [28],[29] using survivorship curves (Figure 1B and Figure S1A, S1B, S1C, S1D). Lifespan is a complex trait; with the pathogenicity of C. elegans food sources being a major component, as it has been suggested that bacterial colonization and resultant tissue damage is the major cause of nematode death even on the standard E. coli strain OP50 [30],[31]. Van Voorhies et al. showed that the substrate in which lifespan is measured is important with wild-type C. elegans lifespan in soil being much shorter than lifespan when grown on agar plates [32]. However, in order to simply investigate the effects of bacterial environment we have chosen to use the more controlled agar plate substrate for C. elegans growth. Wild-type animals had lower TD50 values (i.e. died more quickly) when grown on M. luteus (TD50 = 4.1) than during growth on E. coli (TD50 = 5.6), while growth on both Pseudomonas sp. (TD50 = 8.7) and B. megaterium (TD50 = 12.3) increased lifespan with all pair-wise comparisons of the four bacterial environments significant (P<0.0001) (Figure 1C, Table 1). The extended lifespan in the B. megaterium environment is not likely a consequence of starvation, as generation time (thus larval developmental rate) is not severely altered as would be expected of worms under caloric restriction [33],[34] (Table S1). Wild-type lifespan on E. coli OP50 was only 5.6 days, this is in line with some studies [35] and lower than in others [36] possibly illustrating lab to lab differences in OP50 strains.
To further characterize wild-type C. elegans response to the bacterial isolates we conducted food preference tests. In a previous study [17], it was found that food choice was comprised of more than chemotaxis and there was a dynamic of both food seeking and food leaving behavior whereby C. elegans seeks out higher quality food sources and leaves behind hard to eat bacterial types. Additionally, the same study reported that while C. elegans had little chemoattraction to various tested bacterial species, there was obvious food preference measured by a biased choice assay. As we were interested in food preference, we therefore chose to use the biased choice assay (Figure 2A) rather than chemotaxis assays [17] and we determined food preferences for all pair-wise combinations of bacterial isolates (Figure 2B). We observed a hierarchy of food preferences: Pseudomonas sp. was most preferred, closely followed by E. coli, both of which were preferred over B. megaterium, followed by M. luteus. Interestingly, this hierarchy mirrored the observed trend for fitness in the different bacterial environments (Figure 1A), with C. elegans preferring Pseudomonas sp. on which it was most fit, followed by E. coli, B. megaterium, and M. luteus, respectively. Thus C. elegans food preference appears to correlate with fitness, with bacterial environments on which worms were most fit being preferred.
Transcriptional responses of wild-type C. elegans adults were assayed after growth on each of the four bacteria: E. coli, M. luteus, Pseudomonas sp., or B. megaterium. While dauer formation is an important aspect of the C. elegans life cycle, we have not observed dauer formation in all the native nematodes species that we are modeling with C. elegans. Therefore, young adult animals were analyzed; this also reduced the possibility that age differences confounded gene expression responses to the different bacterial environments (Figure S2). A total of 372 genes were shown to be differentially expressed and statistically significant with multiple testing correction (q<0.01, [37]) across all pair-wise comparisons. Of these, 366 were differentially expressed greater than two-fold and six less than two-fold, illustrating the high power associated with six biological replicates. The 372 genes correspond to a total of 204 unique genes identified across all comparisons (e.g., some genes had significant interactions with multiple bacteria; Tables 2, Table S2). Microarray expression levels of ten genes were verified using quantitative Polymerase Chain Reaction (qPCR) and found to be comparable to the microarray results, indicating that on average the expression differences revealed in the microarray analyses were reliable (Table S3).
Gene Ontology (GO) terms for the identified genes were used to group genes by similar function (See Methods). Metabolism genes were highly represented (16.6%) as expected. Interestingly, genes previously implicated in innate immunity were found in all six comparisons (11.6%). Specifically, we found 20 defense genes upregulated in response to M. luteus, 12 in response to E. coli, 14 in response to B. megaterium and two in response to Pseudomonas sp. That we found defense genes upregulated in response to the latter two bacteria was unexpected, as they cause an increase in lifespan relative to E. coli (Figure 1B, Table 1). Surprisingly, 9.5% of identified genes were involved in cuticle biosynthesis or collagens and 9.0% were membrane associated. Other groups found to make up smaller proportions were developmental (7%), ribosomal (6.5%), proteases (5.5%), and gene expression (3.5%). Finally, genes of unknown function made up the largest portion (23.1% of the total, Figure 3, Table S2), also as expected since one aim of this work was to determine functions for such genes.
We also mapped the identified genes to the C. elegans co-expressed gene mountains as an additional approach to determine functional groups and found similar over-represented groups, as quantified by the representation factor (RF, Table S4, [38]. Thirty-two (RF = 3.4, P = 7.9e-05) of the identified genes mapped to mount 8, which is enriched with genes associated with mitosis as well as genes previously implicated in innate immunity. Mount 19, which is comprised predominately of genes involved in glycolysis, contained 14/204 (RF = 6.4, P = 4.2e-06) identified genes. Twenty-five of our identified genes were found in the mount 22, which represents genes involved in carbohydrate metabolism (RF = 14.3, P = 4.7e-20). Thus two different methods clustered our identified genes similarly, indicating an enrichment in genes for metabolic and defense mechanisms presumably for protection and nutrition.
We obtained all available viable non-sterile loss-of-function mutations for the 204 differentially expressed genes in our study, (21/204, or ∼10% of the total identified genes had available mutants) from the Caenorhabditis Genetics Center (CGC) and used them for biological validation of our microarray results (Table S5). We performed functional tests measuring multiple aspects of life history including brood size, generation time (Table S1, Figure S3), absolute fitness and lifespan (TD50) (Table 1, Figure S1) for all four bacterial environments. We found that many of the mutations had effects on life history traits and differed significantly from wild type in a given bacterial environment. While the majority of mutant strains tested had decreased fitness compared to wild type in each environment, surprisingly, a few mutant strains showed increased fitness when grown on E. coli, M. luteus and Pseudomonas sp. (Table 1 and Table 3). Interestingly, more mutant strains had increased fitness in the B. megaterium environment as compared to wild type than had reduced fitness.
A similar trend was found for brood size as most mutant strains had reduced numbers of progeny in response to growth on the E. coli, M. luteus, and Pseudomonas sp., while growth on B. megaterium resulted in equal numbers of mutant strains that significantly increased and decreased brood size (Table S1, Table 3). Similarly, generation times were slower for most mutants in the same three bacterial environments and only in the B. megaterium environment were there more mutants with faster generation times (Table 3). Surprisingly, lifespan showed a different trend. Growth of mutant strains on Pseudomonas sp., M. luteus and B. megaterium, primarily caused reductions in lifespan, while growth of mutant strains on E. coli resulted in the majority having significantly increased lifespan. Overall, many of the mutational affects on life history were environment specific, demonstrating that transcriptional profiling identified genes of functional importance in each bacterial environment.
We next tested whether differential gene expression between environments had predictive power for the mutational effects on life history traits (lifespan, fitness, generation time and brood size) in those environments. We tested the correlation between the change in relative gene expression and the phenotypic difference in life history traits of a strain containing a mutation in a given gene (Figure 4). We predicted that most genes that were up-regulated in an environment would positively regulate a particular life history trait, such that loss or reduction of that gene function would cause a reduction in fitness or lifespan (or brood size and a possible increase in generation time) in that environment. Therefore, our a priori expectation would be that data points would fall in the lower right and the upper left quadrants for lifespan, fitness and brood size, and the exact opposite for generation time. Indeed, we found that there is a correlation (r = −0.62) between mutant lifespan [Log2(fold change TD50)] and differential expression of genes in comparisons of bacterial environments (Table 4). The slope for the regression was negative, as expected given our prediction that up-regulated genes positively regulate lifespan, and the slope was found to significantly non-zero (p<0.0001; Table 4). It was striking that a correlation was observed for the 21 genes across all six comparisons (126 total comparisons), as most involved gene-by-bacterial comparisons for which we did not observe significant differential expression. In fact independent correlations of only those gene-by-bacteria comparisons that had significant differential expression (q<0.01) were more strongly correlated (r = −0.73), while those that did not have significant differential expression (q>0.01) were less correlated (r = −0.53, Table 4).
The same analysis performed on fitness (λ) revealed a reduced correlation between relative expression and fitness of mutants compared to that of lifespan. Overall, the correlation coefficient for the relationship between differential expression fitness was r = −0.31 (Table 4). The slope of the best-fit line was −0.077 and significantly non-zero (P<0.0004, Table 4). The correlation for the genes found to have significant differential expression (q<0.01) was again better than the complete data set (r = −0.44) and reduced in the gene by bacteria comparisons not found to be significantly differentially expressed (r = −0.26, Table 4).
Correlations were also performed on generation time and brood size, both constituents of fitness (λ). The correlation coefficient for generation time and fold change in gene expression (r = −0.057) indicates poor fit for the model, while the slope was significant (Figure S3, Table S4). Interestingly, there was a much stronger relationship between brood size and differential expression (r = 0.34), which was similar to that of fitness (Table S4) yet still reduced compared to lifespan. However, the regression best-fit line for brood size on differential expression had a non-significant slope (Figure S3, Table S4). Taken together, these tests suggest that while both brood size and generation time are important factors in the relationship between fitness and differential expression, brood size might have a larger contribution.
To illustrate the relationship between the level of gene expression and the degree of phenotypic effect, we consider two examples. The first is hsp-12.6 which encodes a heat-shock protein [39] and was found to be up-regulated 2.4-fold when wild-type worms were grown on E. coli as compared to growth on B. megaterium (Table S2). We found that hsp-12.6 mutants had a 15% proportional reduction in fitness as compared to wild type when the mutant was grown on E. coli (Figure S4A) which is significantly different (p<0.001) from that observed on B. megaterium. Not only is this difference significant, but the fitness of hsp-12.6 mutants was also significantly increased relative to wild type when grown on B. megaterium (Table 1). While the exact role hsp-12.6 plays in these particular interactions of C. elegans with bacteria remains to be elucidated, our results suggest that there was a cost associated with the expression of hsp-12.6 in an environment in which it was not needed and a detriment to loss of function in an environment in which it is needed.
Another example of the relationship between level of gene expression and degree of phenotypic effects was the effect of a rol-6 recessive loss-of-function mutation on lifespan. Transcriptional profiling showed that rol-6 was up-regulated 2.7-fold in the E. coli versus Pseudomonas sp. comparison, which is surprising as rol-6 encodes a cuticular collagen not previously implicated in response to bacteria. However we did observe that genes involved in cuticle formation and function, which includes rol-6, were among the most over-represented groups of genes identified, suggesting modulations in cuticle function could be common to C. elegans interactions with bacteria. Thus, our prediction was that loss of rol-6 function would decrease lifespan in an environment-specific manner, which was what we observed. When the rol-6 mutant strain was grown on E. coli there was a 45% proportional reduction in TD50 compared to the 11% reduction observed on Pseudomonas sp. and this difference was significant (p<0.0001, Figure S4B). Interestingly, rol-6 has been well characterized and extensively studied [40]–[43] for its role as a cuticular collagen and is needed for proper cuticle morphology; yet through use of alternate environments an additional role for rol-6 function in defense was uncovered.
We characterized the effect of exposing C. elegans to different soil bacterial food sources/environments to model naturally occurring interactions that may be driving changes in the bactivorous nematode community in response to land use change in grasslands [5]. Although these specific bacterial environments might not be encountered by C. elegans in the wild, C. elegans is likely to encounter various related bacterial species in its natural environment and is an excellent model for understanding the responses of bactivorous soil nematodes to their bacterial environment. Using transcriptional profiling we set out discover genes that function in environmental interactions, specifically interactions with bacteria. We identified 204 genes that were significantly differentially expressed when adult worms were grown on different bacterial food sources/environments isolated from grassland soils. Most of the identified genes were characterized to be involved in metabolism, defense, cuticle biosynthesis, or were of unknown function (Figure 3). In addition, 46 genes without annotation were identified, which can now be further investigated and functions determined, helping to further our understanding of the C. elegans genome.
A unique aspect of this work is that we calculated fitness using life tables. To our knowledge this is the first use of such analyses to biologically validate candidate genes identified by transcriptional profiling. In addition, we showed a strong correlation between the changes in relative gene expression in comparisons of environments using transcriptional profiling and the phenotypic differences of life history traits when that gene's function was compromised. The best relationship we observed was for lifespan on differential expression, which seems logical given the environments used were bacterial food sources. The correlation of fitness on differential expression was not as high, however. While lifespan is a complex trait controlled by many genes [44], perhaps fitness is an even more inclusive and complex trait that may be controlled by the interaction of many more genes. If this were the case, we would expect the relationship between single mutant effects on fitness in different environments and differential expression between those environments to be more complex. We observed a strong correlation between gene expression and life history trait despite the multiple factors that might complicate the relationship including, genes involved in negative regulation, redundancy, effects of genetic network structure to name a few. This demonstrates the predictive power of transcriptional profiling, at least when used to investigate responses to different external stimuli. Interestingly, other studies that have used gene inactivation to biologically validate environmentally induced differential expression, investigating the responses of C. elegans to cadmium exposure and D. melanogaster to alcohol exposure [45],[46], have also found that a large proportion of genes are functionally important suggesting this could be a common feature to transcriptional regulatory networks that are involved in response to external stimuli [47]. Furthermore, we suggest that not only can transcriptional profiling be used to identify relevant candidate genes, but also the direction and magnitude of expected mutant phenotypes of those genes in response to different environments, ultimately demonstrating their functional importance.
In addition, we used new environments to identify phenotypes for genes of unknown function as well as to show new aspects of phenotypes for previously well-characterized genes. For example, we observed differential expression of pab-2, which encodes a poly-A binding protein, in our microarray experiments and also demonstrated that pab-2 mutants had a significantly higher fitness (λ = 4.14) than wild type (λ = 3.60) when grown on E. coli (Table 1). This result is surprising as N2 was cultured on E. coli OP50 for decades (>1,000 generations) prior to being frozen [44]. It is likely that during this extended period of time that wild type worms became better adapted to life on E. coli, yet pab-2 mutants had a longer lifespan (TD50 = 6.6 vs. 5.6, Table 1), larger brood size (300.2 vs. 290.8) and a faster generation time (4.01 days) than wild type (4.4 days, Table S1) when grown on E. coli. Interestingly, this mutant does not follow the trend postulated by Hodgkin and Barnes [48] that there would be a trade-off between developmental rate and brood size because of differences in resource allocation, as there does not appear to be a trade-off between developmental rate and brood size for pab-2 mutant animals. Further investigation will be required to elucidate the mechanisms by which fitness is increased in pab-2 mutant animals.
When the identified differentially expressed genes were grouped by similar function we found significant enrichment for genes encoding cuticular collagens. This enrichment was also recently found for genes differentially expressed in response to other bacterial species [36]. Wong et al. (2007) found that many C. elegans cuticular collagens were part of a shared response, indicating that they were regulated in response to multiple pathogens (S. marcescens, E. faecalis, E. carotovora, and P. luminescens), suggesting this may be a common response to pathogens. When the functions of the cuticular collagen genes were compromised we found that many had significant effects on lifespan in an environment-specific manner, suggesting this is not merely the consequence of a general “sick” phenotype associated with abnormal cuticle. Furthermore, as these collagen genes were identified through their differential expression in adult worms, differences in juvenile molting are not the cause of their differential expression. Taken together these data suggest that cuticle function may be complex and cuticle structure may be much more dynamic than previously thought, perhaps changing in response to environmental perturbations. It has been show that some bacteria species secrete extracellular proteases and this is an effective nematocide and virulence factor aiding in the pathogen-associated killing of nematode species [49]–[51]. These proteases act to degrade the cuticle of nematodes ultimately leading to their death. It is possible that C. elegans differentially expresses cuticular collagens in response to bacterial protease secretions in an attempt to repair or avoid their pathogenic effects, however further experiments will be need to investigate this hypothesis.
It was somewhat curious that we observed genes involved in innate immunity to be induced in response to Pseudomonas sp. and B. megaterium (Table S2). The response to Pseudomonas sp. was curious because we found that wild type C. elegans was most fit on this bacteria (Figure 1A) and others have found that P. fluorescens, which is similar to our isolate, does not affect C. elegans lifespan [52]. The response to B. megaterium was curious because we showed our isolate increased lifespan (Figure 1B) and others found another isolate did not affect lifespan [53]. One possible explanation of these results is that the C. elegans genome is poised to respond to these bacteria and that the induction of these genes successfully protects worms from these bacteria.
While the responses of some genes can be easily reconciled, the roles of others are more difficult to understand. At first glance it may be difficult to reconcile how gld-1, which functions to limit the proliferation of the gonad germ cells [54],[55], could play a role in the interactions with the environment. However, as gld-1 regulates germ cell proliferation, it is well positioned to integrate signals from the intestine (and elsewhere) to control reproductive output. Specifically, we observed that gld-1 was up-regulated in the B. megaterium vs. Pseudomonas sp. environmental comparison, suggesting that germline proliferation was inhibited in the presence of B. megaterium. Furthermore, gld-1 mutants had a lower fitness in the B. megaterium vs. Pseudomonas sp. environment, suggesting the modulation of gld-1 expression is functionally important. This may be an example of an organism reallocating energy from reproduction to other functions in response to environmental stresses or changes. Thus modulation of gld-1 expression may allow for use of energy for functions other than reproduction including immune response. Interestingly, targets of the insulin signaling pathway (downstream of DAF-2/DAF-16), including innate immunity genes, have been shown to suppress gld-1 induced tumors [56],[57], indicating that immune response through insulin signaling could influence reproductive output. Recently, the activities of specific developmental signaling pathways involved in vulval development have also been shown to be modulated in response to environmental perturbations. In particular, the Notch pathway, which functions with gld-1 to control germ line proliferation, appears to be sensitive to perturbations in food availability [58] . This suggests that while development is robust in response to changes in the environment, slight modulations in processes that affect fitness traits do occur.
We also found that the hierarchy of food preference for the four bacterial isolates mirrored the trend observed for fitness of wild-type C. elegans in the different bacterial environments. This suggests that C. elegans prefers the environment in which it will be most fit. This observation is similar to that of Shtonda and Avery [17] except that their “food quality” measure only took into account developmental rate whereas the measures of fitness shown here also include age-specific fecundity and survival. It has also been suggested that bacterial size is an important determinant of food “quality” with bacteria of smaller diameter having higher quality [59], however we observed that the bacteria with the smallest diameter (M. luteus) resulted in C. elegans having the lowest fitness (Table S1) whereas B. megaterium, the largest bacteria tested (data not shown), did not have the dramatic effect previously shown [59]. Recently Rae et al., (2008) reported that Pristionchus pacificus, another bacterial-feeding nematode associated with scarab beetles, displays differential attractions and susceptibilities to the various bacteria isolated in association with it. The authors suggest that P. pacificus discriminates among bacteria in its environment to maximize reproductive success [60]. Interestingly, P. pacificus has also been found in Konza prairie soils (B. Darby and MAH, unpublished). Food preference could therefore contribute to the mechanism driving observed nematode community structure in grassland soils, as we have recently found that perturbations that mimic disturbances caused by land-use change not only alter soil nematode communities [5] but also the soil bacterial community (KLJ, JDC and MAH unpublished) on Konza prairie. We have also observed that Konza soil nematodes differ in their susceptibility to the different bacteria tested here in terms of infection/colonization (JDC and MAH unpublished data), thus pathogenicity may also contribute to soil nematode community structure. Taken together our data suggest that the expression of metabolism and defense functions may in part drive nematode community dynamics in grassland soil systems through interactions with their bacterial environment. The results from this study suggest that the application of transcriptional profiling to native grassland nematode populations will help identify the functionally important gene functions involved in these interactions.
The following loss-of-function mutants were used: cpi-1(ok1213), dpy-17(e1295), gei-7(ok531), mtl-2(gk125), dhs-28(ok450), Y57A10C.6(ok693), acdh-1(ok1489), rol-6(e187), ctl-1(ok1242), dpy-14(e188), fat-2(ok873), gld-1(op236), hsp-12.6(gk156), cey-2(ok902), cey-4(ok858), cyp-37A1(ok673), elo-5(gk182), pab-2(ok1851), sqt-2(sc108), F55F3.3(ok1758), C23H5.8(ok651). Growth and maintenance conditions were as described [25],[61]. Use of native soil bacteria was as for E. coli (OP50). Bacterial isolate 16S rDNA was sequenced to identify species and sequence is available at NCBI's GenBank Database (accession numbers): Micrococcus luteus (EU704697), Bacillus megaterium (EU704698), Pseudomonas sp. (EU704696).
Biased choice and lifespan assays were performed as previously described [17],[28],[52]. All pathogenicity tests were conducted in at least ten replicate experiments.
Demographic measures were collected on individual worms in the four bacterial environments. From life history measures including age specific reproduction and survival, life tables were used to calculate generation time, intrinsic growth rate and fitness calculated as lambda (λ). Mutant functional tests were performed by plating eggs onto the test bacteria and then placing progeny from this generation onto the test bacteria, one L4 hermaphrodite (P0 worm) per plate was incubated at 20°C with at least ten replicates per treatment per strain per environment. The original P0 worm was re-plated daily until death. Progeny per day was counted (age specific reproduction or mx). Survival of the P0 worm was monitored as well as the survival of all of the progeny from each reproductive period to determine age specific survival (lx). Using life table analysis, intrinsic growth rate (Ro) was calculated as the sum of lx times mx (∑lxmx). Generation time (T) was calculated by (∑lxmx)/(∑xlxmx where x = age class). Lambda was determined from Ro and T by calculating λ = e(lnRo/T) , and λ was used as a measure of absolute fitness [26]. Replicate populations and subsequent life table calculations were used as replicates for statistical tests.
Microarray hybridizations of Caenorhabditis elegans spotted oligonucleotide microarrays (Genome Sequencing Center at Washington University in St. Louis) were made using cDNA made from mRNA extracted from treated young adult C. elegans (N2). cDNA was made from extracted mRNA using Genisphere 3DNA Array350 kits according to manufacture recommendations (Genisphere Inc., Hatfield, Pennsylvania, USA). Microarray hybridizations were performed using a Tecan 400 Hybridization station (Tecan Inc., Zurich, Switzerland). Indirect labeling of cDNA was used to prevent hybridization bias associated with direct labeling procedures [62]. Hybridizations were carried out for 16 hours at 42° C according to manufacturer recommendations (Genisphere Inc. and Tecan Inc.). Hybridized arrays were scanned with an Axon GenePix 4000B (MDS Analytical Technologies, Toronto, Canada) and data was collected using GenePix 6.0 software (MDS Analytical Technologies). Gridding and preprocessing was done manually to remove bad spots and dye artifacts. Raw data files generated are MIAME compliant [63] and available at Gene Expression Omnibus (GEO) series accession number GSE15923.
All six pair-wise comparisons between treatment groups were made in a factorial design (Figure S1) to maximize the ability to detect differences between treatments [64]. Six biological replicates incorporating a dye swap for every other replication were performed to account for any potential dye bias associated with a particular fluorophore (i.e. Cy3 or Cy5, [62]). Data was analyzed as in Wolfinger et al. [65] using SAS statistical software (SAS Institute Inc., Cary, North Carolina, USA) using a two-step mixed model analysis of variance to account for all possible sources of variance. This two-step ANOVA was performed using the MIXED procedure in SAS, with the model for the first stage below and Y = background subtracted raw intensity from the raw data files generated by GenePix 6.0 (MDS Analytical Technologies).
Stage 1 model:Where residuals, termed Relative Fluorescence Intensities (RFI) from stage 1 serve as the input for stage 2.
Stage 2 model:
We used the false discovery rate (FDR) q to address the multiple testing problem [37]. q statistics were calculated in Q-VALUE and using the significance threshold q<0.01. We removed those genes that did not respond to bacterial environment in contrasts. Volcano plots were made in JMP 5.0 software (SAS Institute Inc., Cary, North Carolina, USA).An example of our SAS code can be found at (www.k-state.edu/hermanlab/SASCODE).
Identified genes were assigned GO terms and manually grouped by similar function (Figure 1C, Table S2) incorporating when possible new annotations found in recent literature.
Linear regressions were performed in GraphPad Prism 5 software using Log2(fold change expression) as the independent variable and Log2(fold change TD50), Log2(fold change lambda) Log2(fold change generation time) and Log2(fold change brood size) as the dependent variables. The linear regressions were performed on the 21 genes used in functional tests and for each gene, all 6 environmental comparisons for a total of 126 data points for each regression.
|
10.1371/journal.pgen.1004336 | A Novel C2H2 Transcription Factor that Regulates gliA Expression Interdependently with GliZ in Aspergillus fumigatus | Secondary metabolites are produced by numerous organisms and can either be beneficial, benign, or harmful to humans. Genes involved in the synthesis and transport of these secondary metabolites are frequently found in gene clusters, which are often coordinately regulated, being almost exclusively dependent on transcription factors that are located within the clusters themselves. Gliotoxin, which is produced by a variety of Aspergillus species, Trichoderma species, and Penicillium species, exhibits immunosuppressive properties and has therefore been the subject of research for many laboratories. There have been a few proteins shown to regulate the gliotoxin cluster, most notably GliZ, a Zn2Cys6 binuclear finger transcription factor that lies within the cluster, and LaeA, a putative methyltransferase that globally regulates secondary metabolism clusters within numerous fungal species. Using a high-copy inducer screen in A. fumigatus, our lab has identified a novel C2H2 transcription factor, which plays an important role in regulating the gliotoxin biosynthetic cluster. This transcription factor, named GipA, induces gliotoxin production when present in extra copies. Furthermore, loss of gipA reduces gliotoxin production significantly. Through protein binding microarray and mutagenesis, we have identified a DNA binding site recognized by GipA that is in extremely close proximity to a potential GliZ DNA binding site in the 5′ untranslated region of gliA, which encodes an efflux pump within the gliotoxin cluster. Not surprisingly, GliZ and GipA appear to work in an interdependent fashion to positively control gliA expression.
| Fungal infections cause severe problems for immune-compromised individuals. Current antifungal treatment is problematic, as some therapies are toxic to humans and others are not highly effective. These fungal infections also burden hospitals, as costs to treat and prevent such disease runs high. Aspergillus fumigatus is the most common cause of fungal infections worldwide. A. fumigatus produces a variety of toxins that aid the fungus in survival both in the environment and within host systems. Genes involved in producing such toxins are often found in clusters within the genome, being almost exclusively dependent on transcription factors located within the clusters. Gliotoxin, one such toxin, is known to negatively affect immune cell function. Although gliotoxin has been studied extensively, information is still lacking with regards to regulation of gliotoxin biosynthesis. Our lab has discovered a novel C2H2 transcription factor, GipA, which plays an important role in gliotoxin production. Not only does GipA enhance gliotoxin production when over-expressed, but loss of GipA causes a significant reduction in gliotoxin production. As this gene is not located within the gliotoxin cluster, understanding its mode of action and upstream partners could shed light on toxin production in general and lead to better, more effective antifungal therapies.
| Secondary metabolites are small, low-molecular mass molecules made by numerous organisms that are not essential for normal growth, but can play important roles in defense or signaling [1]–[4]. They can be benign in nature, such as pigments or molecules used in interspecies communication, but they can also be malignant, exhibiting antimicrobial or toxic activities to eliminate competing organisms [5], [6]. Some of these compounds have been exploited by scientists because of their potential benefit to humans. For example, penicillin, produced by P. chrysogenum, is used as an antibiotic and lovastatin, produced by A. terreus, reduces cholesterol [7]. While some of these secondary metabolites benefit humans, others cause harm. Aflatoxin, produced by A. flavus, is carcinogenic, and gliotoxin, produced by A. fumigatus, exhibits immunosuppressive properties [4], [8]–[10]. Genes involved in secondary metabolite production are frequently found as clusters, which are often located in subtelomeric positions [2]–[5], [11]. Furthermore, these clusters are often coordinately regulated, being almost completely dependent on induction by transcription factors located within the clusters themselves [3]–[5], [12], [13].
Zinc binuclear (Zn(II)2Cys6) transcription factors are uniquely found in fungi and represent the most common type of regulators located within these clusters [3], [4], [14], [15]. AflR, a Zn2Cys6 transcription factor, is located within the aflatoxin/sterigmatocystin gene cluster and is required for production of both metabolites [3], [4], [10]. Zn2Cys6 transcription factors, like aflR, generally recognize and bind as homodimers to palindromic sequence motifs, such as CGG(Nx)CCG [14]–[17]. Interestingly, although the palindromic sequences of these binding motifs can be similar or identical for multiple Zn2Cys6 transcription factors, the length and base composition of the linker sequence is highly variable. Therefore, this linker sequence greatly contributes to the specificity of binding for each individual transcription factor [15], [17], [18]. Although gene clusters are often coordinately regulated by the cluster-specific transcription factor, some members can be independently regulated. For example, gliT encodes an oxidoreductase of the gliotoxin biosynthetic cluster, which is required for self-protection against gliotoxin. Even though gliT expression is decreased when the Zn2Cys6 transcription factor, gliZ, is deleted, exogenous gliotoxin induces the expression of gliT, even in a ΔgliZ background [4], [19].
Aside from pathway-specific transcription factors that reside within the cluster, there are numerous other regulatory elements that affect the expression of secondary metabolite clusters. Nutritional and environmental factors, as well as developmental processes, have been shown to affect secondary metabolite production in multiple fungal species [3]–[5]. For instance, penicillin production in A. nidulans is repressed in the presence of glucose, a phenomenon termed carbon catabolite repression [8], [20]. Secondary metabolite repression also occurs in response to nitrogen source, which involves AreA, the global positive regulator of nitrogen metabolite repression. Indeed, in A. nidulans, sterigmatocystin production is repressed in the presence of ammonium and induced when nitrate is the sole nitrogen source [2], [6], [21]. This type of regulation is not seen in all fungi, as carbon and nitrogen source have the opposite effect in A. parasiticus, where growth in the presence of glucose or ammonium results in higher levels of aflatoxin production [2], [6], [21].
A. fumigatus, the leading cause of mold infections worldwide, is an opportunistic pathogen that causes severe problems in immune-compromised populations [9], [22]. These populations include: AIDS patients, cancer patients receiving chemotherapy, solid organ transplant/skin graft patients and victims of chronic granulomatous disease [12], [23]–[25]. One of the most studied secondary metabolites produced by A. fumigatus is gliotoxin, which is also produced by several other Aspergillus species, Trichoderma species, and Penicillium species [13], [26]–[28]. Gliotoxin is a member of the epidithiodioxopiperazine (ETP) class of toxins, which are characterized by a disulfide bridge across a piperazine ring [13], [23]–[27], [29]–[33]. The oxidized form of gliotoxin travels into host immune cells where it is able to affect cellular functions essential to the immune response. These include impediment of phagocytosis and NF-κB activation, as well as induction of apoptosis [23], [25], [26], [29], [32], [33]. As with other secondary metabolites, most of the genes responsible for the production and transport of gliotoxin exist within a gene cluster. The gliotoxin biosynthesis cluster was first identified based on its homology to the sirodesmin PL biosynthesis gene cluster in the ascomycete Leptosphaeria maculans [13], [27], [34], [35].
Within this cluster lies a Zn2Cys6 binuclear finger transcription factor, GliZ, thought to be responsible for general gliotoxin induction and regulation. Indeed, over-expression of gliZ leads to an increase in gliotoxin production and deletion of gliZ results in a loss in gliotoxin production [12], [26], [28]. A DNA binding site has been proposed for GliZ (TCGGN3CCGA), but has not been experimentally proven. This site is present within the promoter region of every gene within the gliotoxin cluster, except gliZ and gliA, which encodes an efflux pump within the cluster [14]. Gliotoxin itself positively regulates expression of the genes within the gliotoxin cluster, as deletion of gliP, the non-ribosomal peptide synthetase (NRPS) required for the first step in the synthesis of gliotoxin, virtually eliminates expression of the other genes in the cluster. This loss in gene expression can be reversed by the addition of exogenous gliotoxin to culture medium [19], [26], [30]. Interestingly, gliT, encoding an oxidoreductase required for resistance of the fungus to gliotoxin, is induced by exogenous gliotoxin even in the absence of gliZ [19]. This demonstrates regulation of genes within the cluster that are independent of the coordinate regulation by GliZ. LaeA, a global regulator of secondary metabolism, has also been shown to regulate the gliotoxin cluster, as gliotoxin is among the secondary metabolites that are lost with deletion of laeA [36]–[38]. Furthermore, loss of vel1 in T. virens (homologous to VeA in Aspergilli) results in a loss in gliotoxin production [39]. This is not surprising, since VeA, VelB, and LaeA form a heterotrimeric complex that regulates secondary metabolism in several fungal species [2], [25], [36], [37], [40]. RsmA, a bZIP transcription factor, positively regulates the gliotoxin cluster through LaeA and GliZ, as loss of either protein abolishes the inducing effects of RsmA over-expression [41]. Map kinase signaling is another element that is important for gliotoxin production, as a strain lacking mpkA, the map kinase in the cell wall integrity pathway, is significantly reduced in gliotoxin production [42]. Nutrient starvation, murine infection, and exposure of germlings to neutrophils have also been shown to up-regulate the gliotoxin biosynthesis cluster through microarray analysis [11], [26].
At this time, information is still lacking as to what proteins and transcriptional regulators are responsible for differential expression of the gliotoxin cluster. We have identified a novel C2H2 transcription factor, gipA, which plays a role in regulating the gliotoxin biosynthesis gene cluster. This protein was identified using a high-copy inducer screen and has not been discussed in detail before now. High-copy expression of gipA results in an increase in transcription of multiple genes within the gliotoxin cluster. Conversely, loss of gipA negatively affects expression of multiple genes within the gliotoxin cluster. In addition, this C2H2 transcription factor specifically regulates gliA expression through a binding site we have identified. Here, we propose a model for gliA expression that involves both GliZ and GipA. Our aim was to identify a novel protein that is involved in gliotoxin production, which we achieved through the discovery of GipA.
To identify genes that regulate the gliotoxin biosynthesis cluster, we performed a high-copy inducer screen. Our goal was to identify genes that, when present in extra copies, induce the gliotoxin biosynthesis cluster in repressing conditions. We used a LacZ reporter system, under the control of the gliA promoter (identified as GL in strain names). GliA encodes an efflux pump within the gliotoxin cluster that is thought to be involved in the transport of gliotoxin out of the fungal cell [13], [27]. Although GliA has not been shown to be essential to gliotoxin transport in A. fumigatus, expression of gliA in a mutant strain of L. maculans provided protection against gliotoxin, but not sirodesmin [27]. We chose gliA for several reasons. First, it has been shown that expression of gliA peaks when the amount of gliotoxin in surrounding medium is maximal [13], [27]. Second, experiments in our lab revealed that gliA is induced within 30 min of A. fumigatus germlings being exposed to human neutrophils, while nutritional induction of the gliotoxin cluster takes 24–48 hrs (data not shown). These data led us to conclude that gliA expression would be a good indicator of gliotoxin cluster expression, as well as gliotoxin production.
The first round of the high-copy inducer screen involved transforming an AMA1-NotI A. fumigatus genomic library [43] into Af293.1-GL (Fig. 1a). AMA1 stands for autonomous maintenance in Aspergillus and was first discovered and isolated from A. nidulans [44], [45]. Due to the presence of this element, the A. fumigatus library plasmids replicate autonomously upon transformation. This feature allows for easy recovery of the plasmids from the fungal genome by transformation of the genomic DNA into bacteria [46], [47]. Furthermore, this screen is a high-copy inducer screen, because there are on average 10 to 30 copies of any given AMA1-containing plasmid per genome [44], [45]. We grew transformants on gliotoxin-repressing medium with X-gal and screened for colonies that produced a blue pigment. This indicated that the AMA1-NotI plasmid within the genome was inducing gliotoxin cluster expression in repressing conditions. For the second round of the high-copy inducer screen, we transformed individual plasmids, isolated in the first round of the genetic screen, into Af293.1-GL (Fig. 1b). We transformed individual vectors to be sure that the effect we observed from the first round was the result of only one plasmid and not multiple plasmids. To measure LacZ levels in a more quantitative manner, we isolated total protein from transformants and measured β-galactosidase activity. We transformed a pDONR AMA empty vector (AMA.GL) and pDONR AMA-gliZ (AMA-gliZ.GL) as a negative control and positive control, respectively. The third round of the high-copy inducer screen entailed isolating the individual genes from each AMA1-NotI library plasmid and transforming them into Af293.1-GL (Fig. 1c). For this round, we only measured β-galactosidase activity.
The gene that induced lacZ the most was a C2H2 transcription factor (AFUA_6G01910), which we have designated gipA for gliotoxin inducing protein. High-copy expression of gipA in the Af293.1-GL background (AMA-gipA.GL) induces a 400-fold increase in β-galactosidase activity, compared to the empty vector control (AMA.GL) (Fig. 2). The level of β-galactosidase activity in our positive control, AMA-gliZ.GL, was almost 30-fold higher than the empty vector control, AMA.GL. These data indicate that GipA positively regulates gliA expression, similar to GliZ.
The predicted sequence of gipA contains one intron and two C2H2 regions at the 3′ end (Fig. 3a). The two C2H2 regions are canonical, the first being X2-C-X2-C-X12-H-X3-H and the second being X2-C-X2-C-X12-H-X5-C, which is a natural variant [48], [49]. The 5′ untranslated region (UTR) of gipA appears to be unusually long (877 bp) and contains three μORFs, which indicates that gipA may be under post-transcriptional control due to regulation of translational efficiency and mRNA stability. Northern analysis was consistent with cDNA predictions (Fig. 3b). Sequence similarity searches indicated that proteins homologous to GipA are present in only a limited number of species, the closest being a C2H2 transcription factor in N. fischeri, a close relative to A. fumigatus (Fig. 3c). The rest of the homologous proteins were from other Aspergillus species and a few Penicillium species, which suggests that GipA is not highly conserved at the primary sequence. For proteins, primary sequences evolve and change much more rapidly than do the tertiary structures. There have been numerous examples where two proteins have low sequence similarity, but once crystallized, exhibit almost identical folding patterns and subsequently share similar functions [50]. For example, Gcn4 of Saccharomyces cerevisiae and CpcA of A. niger share a 35% identity, yet they both function in amino acid biosynthesis. Furthermore, CpcA is able to complement a Δgcn4 mutant in S. cerevisiae [51]. When solely comparing the putative DNA binding domain of cpcA, the identity between CpcA and Gcn4 increases to 70% [51]. Therefore, the lack of homologous counterparts to GipA in other organisms does not necessarily mean that there is not a protein present in other fungi that functions similarly to GipA.
In a previous section, we showed that extra copies of gipA induce expression of lacZ, under the control of the gliA promoter, which suggests that GipA is inducing gliA. Since the gliotoxin biosynthesis cluster is coordinately regulated, extra copies of gipA should also induce other genes within the cluster. As expected, AMA-gipA.GL had a higher amount of gliA mRNA, than AMA.GL. Transcript levels of gliA in AMA-gipA.GL were 4.5-fold higher, compared to AMA.GL (Fig. 4a). The mRNA levels of the other gliotoxin-specific genes tested were also significantly higher in AMA-gipA.GL, compared to AMA.GL, as gliZ was induced 12-fold, gliP was induced 5-fold, and gliT was induced 2-fold (Fig. 4a). Gliotoxin production reflected what was seen with mRNA levels, as AMA-gipA.GL produced gliotoxin at a higher amount than AMA.GL (7-fold higher) (Fig. 4b). AMA-gliZ.GL was the positive control and showed the same pattern as AMA-gipA.GL, with respect to induction of the gliotoxin biosynthesis cluster.
To expand our view of GipA regulation, we performed a microarray analysis of AMA-gipA.GL vs. AMA.GL, grown in repressing conditions for 24 hrs. Of the 9,436 total genes analyzed, 443 genes were up-regulated >2-fold and 75 genes were down-regulated >2-fold in the AMA-gipA.GL strain (Fig. S1). There were several genes common to secondary metabolism clusters (e.g. transporters, oxidoreductases, methyltransferases, nonribosomal peptide synthetases and polyketide synthases) up-regulated (Table 1). Approximately 31 secondary metabolism clusters have been proposed using genomic mapping and microarray techniques [52], [53]. Of these 31 potential secondary metabolism clusters, 18 contained at least one gene that was up-regulated >2-fold in AMA-gipA.GL, compared to AMA.GL (Table 2 & Fig. S2). Based on microarray data done previously [52], loss of laeA, a global regulator of secondary metabolism, affected 13 of 22 identified secondary metabolite clusters. This suggests that GipA potentially induces numerous other secondary metabolism gene clusters in A. fumigatus, similar to LaeA. One caveat of this microarray, however, is that we measured the effects of high-copy expression of gipA, while the microarray done with LaeA compared a deletion strain to a wild-type and complemented strain. Therefore, based on these results, we cannot conclude that loss of gipA will have a significant effect on multiple secondary metabolite clusters, as is observed with loss of laeA.
Since GipA can induce gliotoxin production, we sought to discover if loss of gipA has any effect on the gliotoxin cluster. We replaced the coding region of gipA with pyrG and designated this strain as ΔgipA. We created a complemented strain, gipA(R), using a hygromycin resistance cassette (hygroR) as the selective marker. We also created a gliZ deletion strain as a control, since previous studies have shown that loss of gliZ results in a significant decrease in mRNA levels of gliotoxin-specific genes [12]. Loss of gipA caused a significant decrease in mRNA levels of gliA, gliZ, gliP, and gliT in non-repressing conditions, as most genes exhibited close to a 50% reduction (Fig. 5a). The gliotoxin biosynthesis cluster is not completely dependent on gipA, as there was still mRNA being made for the genes we tested. The gipA deletion mutant also produced significantly less gliotoxin than the 1160G control strain (50% reduction) (Fig. 5b). Gliotoxin-specific gene expression and gliotoxin production of gipA(R) were restored beyond wild-type levels, which demonstrates that the effect we observed with the gipA deletion was due to the absence of gipA. As expected, loss of gliZ caused almost a complete loss in gene expression for gliA, gliP, and gliT and abolished gliotoxin production. A gliZ/gipA double deletion mutant revealed a pattern of gene expression and gliotoxin production similar to the gliZ single deletion mutant (data not shown).
Since GipA is a C2H2 transcription factor, it is likely that GipA is directly binding to DNA. We sought to identify a consensus sequence and to discover if this sequence was present within the gliotoxin biosynthesis cluster. A protein-binding microarray analysis identified a consensus DNA binding sequence for GipA (5′-TNNVMGCCNC-3′) (Fig. 6a). This putative sequence is 10 nucleotides, which coincides with one complete turn of the DNA double helix. Interestingly, there are 7 high-content positions (5′-VMGCCNC-3′), which would correlate with two C2H2 zinc finger DNA binding domains [54], and then a “T” residue two base pairs upstream of the 7 high content core positions. The protein-binding microarray verified direct DNA binding of GipA, as the purified DNA-binding domain, and not whole cell extract, was analyzed in the microarray. We analyzed the genomic sequence of the gliotoxin biosynthesis cluster to locate any potential GipA binding sites. Indeed, we found variations of this consensus sequence scattered throughout the gliotoxin biosynthesis cluster. In fact, we identified a possible GipA binding site within the intergenic region of gliA (5′-TTGCCGCCAC-3′ 315 bp upstream of the start site), as well as all other gliotoxin-specific genes, except gliM (Fig. 6b). Not all putative GipA DNA binding sites throughout the intergenic regions in the gliotoxin biosynthetic cluster contain the “T” residue in the 1 position, which means that either these sites are possibly only weakly recognized by GipA, if at all, or the “T” residue is not part of the actual GipA DNA recognition element.
We created two mutant backgrounds: BSM1 (TTGCCGCCACCTGCCGCCAC) and BSM2 (TTGCCGCCAC→TTGGGTGAGC), by mutating the putative GipA DNA binding site on the gliA-lacZ construct we used in the original screen. We measured lacZ expression with β-galactosidase assays in the presence of different high-copy plasmids (strains used in this study are listed in Table S2). When normalized to AMA.GL, AMA-gipA.GL displayed increased LacZ levels (13-fold), as was to be expected from previous results (Fig. 6c). High-copy expression of gipA in the BSM1 background also induced lacZ significantly (53-fold), relative to AMA.BSM1 (Fig. 6c). The magnitude of induction was enhanced with the BSM1 mutation when compared to the wild-type binding site (13-fold vs. 53-fold, respectively). This enhanced expression was not due to higher overall lacZ levels, as shown when comparing empty vector controls in all three backgrounds (Fig. 6d). AMA-gipA.BSM2 only weakly induced lacZ (4-fold), compared to AMA.BSM2 (Fig. 6c). This suggests that mutation of a core sequence in the putative GipA DNA binding site significantly reduces the ability of GipA to induce lacZ. Mutation of the 5′ “T” residue did not reduce GipA-specific gliA induction, but rather increased the response of gliA to GipA high-copy expression.
Interestingly, expression of lacZ in the pDONR AMA-gliZ strains followed a similar pattern to that of pDONR AMA-gipA strains (Fig. 6c). Therefore, lacZ was induced in AMA-gliZ.GL (19-fold) and AMA-gliZ.BSM1 (93-fold), compared to AMA.GL and AMA.BSM1, respectively. The level of induction was enhanced by the BSM1 binding site, compared to the wild-type binding site (19-fold vs. 93-fold, respectively). Furthermore, lacZ was only weakly induced in AMA-gliZ.BSM2, relative to AMA.BSM2 (4-fold). Therefore, mutation of the putative GipA DNA binding site is affecting the ability of GliZ to induce gliA.
To determine if GipA is dependent on GliZ for expression of gliA, we created a series of mutants and measured mRNA levels of gliA and gliP. High-copy expression plasmids used in previous sections were transformed into a wild-type background (AMA.G, AMA-gliZ.G, and AMA-gipA.G) and a gliZ deletion background (AMA.Z, AMA-gliZ.Z, and AMA-gipA.Z) (strains used in this study are listed in Table S2). All strains were grown in non-repressing conditions to induce expression of the gliotoxin biosynthesis cluster. In the pyrG+ background, AMA-gliZ.G and AMA-gipA.G had increased mRNA levels for gliP, while only AMA-gipA.G had increased mRNA levels for gliA, compared to AMA.G (Fig. 7). These changes were not significant because growth in non-repressing conditions already induced these genes to high levels, so having extra copies of GliZ or GipA did not greatly contribute to gene expression. In AMA.Z, mRNA for both gliA and gliP was almost completely undetectable, as to be expected from previous experiments. High-copy expression of gliZ brought transcript levels of both genes back to AMA.G levels (Fig. 7). AMA-gipA.Z displayed a reduction in mRNA levels similar to AMA.Z. For gliA, the level of mRNA present in AMA-gipA.Z was slightly higher (close to 5-fold) than what was observed for AMA.Z. However, the level of gliP mRNA did not exceed that of the AMA.Z empty vector control. Therefore, GipA was not able to induce gliA or gliP in the absence of GliZ. This suggests that GliZ is required for GipA to induce both gliP and gliA.
To determine if GliZ is dependent on GipA for expression of gliA, we created a series of mutants and measured mRNA levels of gliA and gliP. High-copy expression plasmids used in previous sections were transformed into a wild-type background (AMA.G, AMA-gliZ.G, and AMA-gipA.G) and a gipA deletion background (AMA.A, AMA-gliZ.A, and AMA-gipA.A) (strains used in this study are listed in Table S2). All strains were grown in non-repressing conditions to induce expression of the gliotoxin biosynthesis cluster. As expected from previous experiments, mRNA levels in AMA.A of both gliA and gliP were reduced significantly, 50% and 80%, respectively. Levels of gliP mRNA in AMA-gliZ.A were comparable to those of AMA.G; however mRNA levels of gliA were not significantly higher than background levels (AMA.A) (Fig. 7). This indicates that GliZ is not dependent on GipA for induction of gliP, however induction of gliA by GliZ does appear to be dependent on GipA.
Although recent studies have revealed gliotoxin intermediates, which have led to a better understanding of the biosynthesis of gliotoxin, information on regulation of the genes involved in the biosynthesis pathway is lacking [28]. With the use of a high-copy inducer screen, our lab has uncovered a novel protein, GipA, which appears to be involved in the regulation of the gliotoxin cluster. GipA is a C2H2 transcription factor, which harbors an unusually long 5′ UTR. Furthermore, there are three μORFs within the 5′ UTR, which suggests that gipA may be under post-transcriptional control. There are two canonical C2H2 zinc finger DNA binding regions at the 3′ end of gipA, the first being X2-C-X2-C-X12-H-X3-H and the second being X2-C-X2-C-X12-H-X5-C, which is a natural variant [48], [49]. Two C2H2 zinc finger binding domains is consistent with a 6–8 base pair DNA recognition element [54]. The putative DNA binding site that we identified for GipA contains 7 high content positions (5′-TNNVMGCCNC-3′), in addition to a 5′ “T” residue two base pairs upstream of the core residues. Based on mutagenesis of the putative GipA DNA binding site in the gliA promoter, the 7 high content positions are likely part of an active GipA DNA binding site. While the “T” residue is not necessary for GipA-specific gliA induction, mutation of the “T” residue does enhance GipA-specific gliA expression. We have devised a possible model for the specific regulation of gliA that involves GliZ and GipA (Fig. 8). We propose that GliZ and GipA work together at the same binding site or in close proximity to induce gliA. In this model, GipA and GliZ are dependent on each other for inducing gene expression of gliA.
GipA appears to play an important role in gliotoxin biosynthesis, as high-copy expression of gipA causes increased gliotoxin production and loss of gipA causes a significant reduction in gliotoxin levels. Furthermore, GipA is possibly regulating other secondary metabolite clusters within A. fumigatus, similarly to LaeA, a global regulator of secondary metabolism [52]. In contrast to LaeA, though, GipA is a C2H2 transcription factor, which likely directly binds to DNA recognition sites. Other C2H2 transcription factors, such as PacC, NsdC, NsdD and CreA, have also been identified as regulators of secondary metabolism. PacC regulates a wide range of genes in response to ambient pH, including those involved in secondary metabolism. Penicillin production in A. nidulans is induced in alkaline conditions, while sterigmatocystin production is repressed [4], [55], [56]. NsdC and NsdD are both involved in sexual development in A. nidulans. However, recent work by Jeffrey Cary et al. [57] supported additional roles for NsdC and NsdD in asexual development and secondary metabolism in A. flavus. Although AflR expression was not affected, loss of either nsdD or nsdC resulted in decreased expression of other genes in the aflatoxin biosynthesis gene cluster [57]. Finally, CreA is a global regulator of carbon catabolite repression in fungi. In addition to carbon metabolism, CreA also affects secondary metabolism, as cephalosporin production is reduced in Acremonium chrysogenum in response to high glucose [56], [58]. It is possible that GipA serves to enhance expression of the gliotoxin cluster in certain environmental conditions. Both high-copy expression and loss of gipA affect gliZ expression, so the effects we see for the entire cluster could be the direct consequence of GipA binding to each promoter region, or an indirect consequence of GipA partially regulating gliZ, which in turn regulates other genes within the cluster. Loss of both gliZ and gipA does not give conclusive evidence for either possibility, as loss of gliZ alone completely abolishes gliotoxin production. None of the mutants tested displayed abnormal growth rates or attenuated virulence in a Drosophila melanogaster model system, indicating that high-copy expression of gipA or loss of gipA does not affect overall fitness of A. fumigatus (data not shown).
Our data show a dependency between GipA and GliZ with respect to gliA expression. Firstly, there is a putative GliZ DNA binding site embedded within a GipA DNA binding site in the gliA 5′ UTR. Although a GliZ binding site has not yet been experimentally determined, one has been predicted (TCGGN3CCGA). This sequence is present in the intergenic region of every gene within the gliotoxin cluster, except gliZ and gliA [14]. Recognition of these sequences by Zn2Cys6 binuclear finger transcription factors is often very specific and even a slight change to the length or base composition of the linker sequence can result in reduced binding in vivo [15], [17], [18]. Within the gliA promoter region, a sequence overlapping the GipA binding site has the CGG-CCG inverted repeats common to Zn2Cys6 binuclear finger transcription factor binding sites, but the linker sequence is longer (8 bps) than that of the predicted GliZ DNA binding site (3 bps). Mutation of the CCG repeat almost completely abolishes gliZ-mediated induction of gliA. This suggests that either the prediction for the GliZ DNA binding site is incorrect and is actually CGGN8CCG or that the GliZ DNA binding site within the gliA UTR is unique. Due to the position of the putative GliZ DNA binding site, mutation of this CCG also changes the core sequence of the putative GipA DNA binding site. Accordingly, gipA-mediated induction of gliA is almost completely abolished.
Dependent dual regulation of two transcription factors has been uncovered in other organisms. For example, in A. nidulans, FlbB and FlbD both bind in close proximity to the brlA promoter to regulate asexual development through brlA activation [59], [60]. Furthermore, FlbB does not bind to the brlA promoter in the absence of FlbD, indicating that these two transcription factors are dependent on each other for DNA binding and activation of brlA [60]. Although the mutational analysis of the gliA promoter raises the possibility of a GliZ-GipA dependency, it does not give conclusive evidence. One could argue that disruption of the GliZ DNA binding site causes loss of gliA induction that indirectly abolishes gipA-mediated induction of gliA. However, our hypothesis of a dependency between GipA and GliZ with respect to gliA expression is further supported by the fact that GliZ cannot induce gliA in the absence of GipA. As expected, loss of gliZ severely decreases mRNA levels of gliA, which cannot be rescued by high-copy expression of gipA. Interestingly, loss of gipA also significantly decreases mRNA levels of gliA, which cannot be rescued by high-copy expression of gliZ. Clearly, these two proteins work together to induce gliA beyond basal levels in non-repressing conditions. This pattern of dependency is unique to gliA expression, as gliP is induced by high-copy expression of gliZ, even in the absence of gipA.
Together, these data support a model in which GliZ and GipA are working together to regulate gliA (Fig. 8). There appears to be a dependency with regards to gliA expression that we demonstrated in our experiments. This model does not apply to every gene within the gliotoxin cluster, as no other genes have a GipA binding site embedded within a possible GliZ binding site, except gliZ, although these sequences are located farther upstream of the gliZ start site (Fig. 9). Possibly, GipA serves to aid GliZ in binding to the gliA promoter region, as this binding site is different from the others present in the gliotoxin cluster. Although there are possible GipA binding sites in all gliotoxin gene promoters, except gliM, we cannot say with certainty whether GipA is directly binding to these other promoter regions or if GipA is simply binding in the gliA promoter in conjunction with GliZ. This adds to mounting evidence that genes within a gene cluster are typically coordinately regulated, but can also be individually expressed in response to certain stimuli. It is possible that gliT and gliA are both independently regulated to protect the fungus from exogenous gliotoxin, although gliA was not induced in a ΔgliZ mutant in the presence of exogenous gliotoxin as gliT was [19]. Another possibility is that gliA is independently regulated to aid in the transport and/or expression of other secondary metabolism clusters in A. fumigatus. There has been evidence in other fungal species that crosstalk between these gene clusters exists [61].
All primers used in this study are listed in Table S1.
All strains used in this study and genotypes are listed in Table S2. We maintained Af293.1, Af293.1-GL, 1160, Af293.1-BSM1, and Af293.1-BSM2 on YAG medium supplemented with uridine and uracil (0.5% yeast extract, 1% glucose, trace elements and vitamin mix as modified [62], 10 mM MgCl2, 1.5% agar, 5 mM uridine, and 10 mM uracil,). We grew AMA.G, AMA.Z, AMA.A, AMA-gliZ.G, AMA-gliZ.Z, AMA-gliZ.A, AMA-gipA.G, AMA-gipA.Z, and AMA-gipA.A on YAG medium with 400 µg/ml hygromycin. We maintained all other strains on YAG medium. Unless otherwise noted, we grew all strains at 37°C for 48 hrs. For phenotypic growth assays of high-copy and deletion strains, we inoculated approximately 1000 spores of each strain onto MMVAT (1× MM salts [20 mM ammonium tartrate, 7 mM KCl, 2 mM MgSO4·7H2O], 1% glucose, 12 mM KPO4 pH 6.8, trace elements, vitamin mix as modified [62], and 1.25% agar), MMVAT with 10 µg/ml gliotoxin, and YAG. MMVAT plates were incubated for 72 hrs. We repeated plate growth assays twice for a total of three independent tests. We measured radial growth of each colony and scanned plates on the final day of growth.
We used fusion PCR (f-PCR) to create a gliAP-lacZ-gliAT construct. For the first reaction, we created three cassettes: gliA 5′P, lacZ, and gliA 3′T, using primer pairs GliA F1 and GliA 5′ R, lacZ F and lacZ R, and GliA 3′ F and GliA R, respectively. We obtained these three fragments by PCR using e2TAK DNA polymerase (Takara Bio Inc., Otsu, Shiga, Japan, Otsu, Shiga, Japan) following manufacturer recommendations. The gliA 5′P fragment had a 3′ extension identical to the first 15 base pairs of lacZ. The lacZ fragment had a 5′ extension identical to the last 15 base pairs of gliA 5′P and a 3′ extension identical to the first 15 base pairs of gliA 3′T. The gliA 3′T fragment had a 5′ extension identical to the last 15 base pairs of lacZ. We used Af293 genomic DNA as template for the gliA 5′P and gliA 3′T regions and λGT11 as template for lacZ.
For the second reaction, we fused the three fragments together using GliA F1 and GliA R as primers. We amplified a 50 µl reaction containing 50 fmol of each fragment, 0.3 µM of each primer, 500 µM of deoxynucleoside triphosphates, buffer 3 at a 1× concentration, and 1 µl of Expand Long DNA Template Mix (Roche Applied Science, Indianapolis, IN) per manufacturer's instructions (briefly, 94°C for 2 min, 10 cycles of 94°C for 10 sec, 62°C for 30 sec, 68°C for 4.5 min and 15 cycles of 94°C for 15 sec, 62°C for 30 sec, 68°C for 4.5 min, increasing the final extension time by 20 sec with each cycle).
We cloned the fusion product into pDONR HPH A [63] using a BP recombination reaction (Invitrogen, Grand Island, NY). We transformed the reaction mix into TOP10 cells (Invitrogen, Grand Island, NY) by electroporation, as recommended by the manufacturer. We grew the transformation mix on LB (1% tryptone, 0.5% yeast extract, 1× SOB salts [10 mM NaCl, 2.5 mM KCl], 1.5% agar) +50 µg/ml kanamycin at 37°C overnight. We picked colonies and transferred to 2 ml of LB liquid +50 µg/ml kanamycin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany). We digested plasmid DNA with specific enzymes to verify the correct insertion. We designated this vector as pDHGL.
We grew Af293.1 in MAG medium supplemented with uridine and uracil (2% malt extract, 0.2% peptone, 1% glucose, trace elements and vitamin mix as modified [62], 2% agar, 5 mM uridine, 10 mM uracil). We performed the transformation as previously described [64], keeping pDHGL as a circular vector. We grew transformants on MMVAT supplemented with 5 mM uridine, 10 mM uracil, 0.2 M sucrose, and 400 µg/ml hygromycin at 37°C for 3–5 days. We identified the presence of pDHGL by Southern hybridization [65] of the lacZ coding region (Fig. S3). We also streaked transformants onto MMVSN supplemented with uridine and uracil (as described above for MMVAT, except 1×MM salts contain 20 mM sodium nitrate instead of ammonium tartrate) and 40 µg/ml X-gal to grow at 37°C for 2 days. We screened for transformants that grew in the presence of hygromycin and developed a blue pigment on X-gal, signaling that lacZ expression was working properly. We designated this strain as Af293.1-GL.
For the first round of the high-copy inducer screen, we grew Af293.1-GL in MAG supplemented with uridine and uracil. We transformed the AMA1-Not1 A. fumigatus genomic library [43] into Af293.1-GL as described previously [64], but with changes. We combined each transformation mix with 50 mls of CM top agar (MMVAT, as described above, 0.1% yeast extract, 0.2% peptone, 0.1% tryptone, 1% CM supplement [27 mM adenine HCl, 33.5 mM methionine, 173 mM arginine, and 1.3 mM riboflavin], and 1% agar) supplemented with 1 M sucrose and spread the mixture over 10 plates (5 ml/plate). We grew transformants on CM supplemented with 0.2 M sucrose and 40 µg/ml X-gal at 37°C for 3–5 days. We screened for transformants that were both prototrophic for uridine and uracil and producing a blue pigment. We prepared genomic DNA from transformants [66] and transformed 1 µl of genomic DNA into TOP10 cells (Invitrogen, Grand Island, NY) by electroporation, as recommended by the manufacturer. We grew the transformation mix on LB +100 µg/ml ampicillin at 37°C overnight. We picked colonies and transferred to 2 ml of LB liquid +100 µg/ml ampicillin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany) and digested it with KpnI to identify individual plasmids.
For the second round of the high-copy inducer screen, we grew Af293.1-GL in MAG supplemented with uridine and uracil. We transformed each individual plasmid isolated from the first round of the genetic screen as described above, but with changes. We plated two amounts of protoplasts (20 µl and 100 µl) each in 4 mls of MMVAT top agar (as described above but with 1% agar) supplemented with 1 M sucrose. We grew transformants on MMVAT supplemented with 0.2 M sucrose and 40 µg/ml X-gal at 37°C for 3–4 days and put plates in the 4°C refrigerator to facilitate blue pigment production. Plasmids causing at least 80% of colonies to turn blue were sequenced using primers AMA-NotI F and AMA-NotI R. We also grew transformants in 10 mls CM at 37°C stationary overnight. We collected mycelia, froze in liquid nitrogen, and lyophilized overnight. We collected total protein and performed β-galactosidase assays to measure LacZ levels quantitatively (detailed below).
For the third round of the high-copy inducer screen, we PCR amplified individual genes from genomic library plasmids, flanked by native 5′ and 3′ non-coding regions. For gipA (Afu6g01910), we used the primer pair 6g01910 F and 6g01910 R. We amplified the fragments from Af293 genomic DNA using e2TAK DNA polymerase (Takara Bio Inc., Otsu, Shiga, Japan) following manufacturer recommendations. We cloned the PCR fragments into pDONR AMA [63] with a BP recombination reaction (Invitrogen, Grand Island, NY). We transformed the reaction mix into TOP10 cells (Invitrogen, Grand Island, NY) by electroporation, as recommended by the manufacturer. We grew the transformation mix on LB +100 µg/ml ampicillin at 37°C overnight. We picked colonies and transferred to 2 ml of LB liquid +100 µg/ml ampicillin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany). We digested plasmid DNA with specific enzymes to verify the correct insertion. We designated the gipA-containing vector as pDONR AMA-gipA. We created a control vector, pDONR AMA-gliZ, which contained the gliZ coding region flanked by promoter and terminator regions. We generated this vector as described above for pDONR AMA-gipA using primers GliZ attB 1 and GliZ attB 2.
We grew Af293.1GL in MAG medium, supplemented with uridine and uracil. We performed the transformation as previously described [64], with changes, using pDONR AMA and pDONR AMA-gliZ as controls. After the 3 hour incubation of protoplasts, we carried out all reactions at half the specified volume. We used 500 ng–750 ng of circular vector DNA for each reaction. We plated two amounts of protoplasts (20 µl and 50 µl) each in 4 mls of MMVAT top agar supplemented with 1 M sucrose. We grew transformants on MMVAT medium supplemented with 0.2 M sucrose at 37°C for 2–3 days. To measure LacZ levels quantitatively, we grew transformants in 10 mls CM at 37°C stationary overnight. We collected mycelia, froze in liquid nitrogen, and lyophilized overnight. We collected total protein and performed β-galactosidase assays (detailed below). We chose three strains to use for future analysis: AMA.GL, AMA-gliZ.GL, and AMA-gipA.GL.
We isolated gipA cDNA clones from a λ phage library constructed with the UniZAP vector and poly (A) + mRNA, as described by the manufacturer (Stratagene, La Jolla, California). We performed a primary screen of the λ phage library as recommended by manufacturer (Stratagene, La Jolla, California). We used the gipA coding region as a probe [65]. We excised plugs containing positive plaques and placed them in SM (.1 M NaCl, 10 mM MgSO4·7H2O, 50 mM Tris·HCl [pH 7.5], and .01% gelatin) at 4°C overnight. After determining pfu concentration of plaques, we performed a secondary screen, as performed for the primary screen. We PCR amplified cDNA from excised phagemids using e2TAK DNA polymerase (Takara Bio Inc., Otsu, Shiga, Japan) and following manufacturer recommendations. We used M13 F and M13 R as primers for the reaction. Based on the PCR band sizes, we picked colonies from selected plates and transferred to 2 ml of LB liquid +100 µg/ml ampicillin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany) and sequenced the cDNA insert using M13 F and M13 R primers. We used DNASTAR software (DNASTAR Inc.) to assemble and analyze the sequences obtained from the library screen.
For all Dot Blot assays, we grew stationary cultures in 25 mls of CM (repressing) (described above) or CD (non-repressing) (87.6 mM sucrose, 35.3 mM sodium nitrate, 5.8 mM K2HPO4, 2 mM MgSO4·7H2O, 6.7 mM KCl, and 0.025 mM ferrous ammonium sulfate) at a concentration of 5×106 spores/ml at 37°C for 48 hrs. We included 400 µg/ml hygromycin in dot blot assays involving AMA.G, AMA.Z, AMA.A, AMA-gliZ.G, AMA-gliZ.Z, AMA-gliZ.A, AMA-gipA.G, AMA-gipA.Z, and AMA-gipA.A. We prepared total RNA from freeze-dried mycelia using the TRIzol method [62]. We incubated total RNA with denaturing solution (50% formamide, 16%formaldehyde, 1× borate buffer [20× borate buffer: 0.4 M Boric Acid, 4 mM EDTA, pH 8.3 with NaOH], 0.025% bromophenol blue) for 10 min at 65°C. We quenched samples on ice for 10 min, then added equal volume 20× SSC (3 M NaCl, 0.3 M sodium citrate, pH 7.0). We placed a nylon membrane that had been equilibrated in 10× SSC for 10 min into a 96 well dot blot apparatus attached to a vacuum manifold. We collected samples, each containing 3 µg of RNA unless otherwise noted, in 100 µl volumes by aspiration. We aspirated 50 µl of 10× SSC through the membrane in duplicate immediately before and after samples were collected. Once all samples were aspirated, we air dried nylon membranes overnight and then baked them in an oven for 2 hrs. For prehybridization and hybridization, we sealed nylon membranes in a bag using a heated sealer. We prehybridized membranes for 4–6 hrs at 42°C and hybridized membranes overnight at 42°C. For DNA probes, we only used the coding region of each gene of interest (gliA, gliP, gliZ, gliT, gipA, and actin). After hybridization, we washed membranes as we would for Southern hybridization [65] and exposed them to a Typhoon 8600 PhosphorImager (GE Healthcare Life Sciences, Pittsburgh, PA) overnight. We quantified the intensity of hybridization using ImageQuant5.1 software (GE Healthcare Life Sciences, Pittsburgh, PA).
We extracted spent culture medium from RNA dot blot assays in 15 ml chloroform for 30 min at room temperature on an orbital shaker set to 250 rpm. We transferred the chloroform phase to a 50 ml conical tube (BD Biosciences, San Jose, California) and repeated the extraction twice for a total of 45 mls of chloroform. We dried open tubes under a hood until the chloroform was completely evaporated. We added 15 mls of chloroform to each tube and mixed, to concentrate the extracted material at the bottom. We dried open tubes under a hood until chloroform was completely evaporated. We added 1 ml of methanol to each tube and mixed to dissolve all methanol-soluble substances and transferred extracts to microcentrifuge tubes (Fisher Scientific, Pittsburgh, PA). We dried open tubes under a hood until all methanol was completely evaporated.
We dissolved extracts in 50 µl of dimethyl sulfoxide (DMSO), spun tubes to pellet insoluble debris, and transferred DMSO to fresh microcentrifuge tubes. We quantified gliotoxin levels by running samples through a reverse-phase high performance liquid chromatography (RP-HPLC) system with a Waters 996 photodiode array detector (Waters, Milford, MA). We ran samples through a Sonoma 2.1×250 mm C18 column (100 Å pore size) packed with 5 µM particles (VWR, Radnor, PA). The mobile phase consisted of H2O, 0.1% TFA (solution A) and 100% Acetonitrile, 0.1% TFA (solution B): 10% B up to 80% B over 30 min. The injection volume was 10 µl and flow rate was set at 0.4 ml/min. Gliotoxin eluted from the column at 14.7 min and absorbance was read at 268 nm wavelength. We determined gliotoxin concentrations by interpolation from a 9 point standard curve (39 ng to 10 µg) prepared using purified gliotoxin (Sigma-Aldrich Corp., St. Louis, MO).
The DNA amplicon microarray for Af293 was created previously [67]. We grew AMA.GL and AMA-gipA.GL in 25 mls CM (described above) at 37°C for 24 hrs in stationary cultures. The spore concentration was 5×106 spores/ml. We used two independent biological replicates for AMA.GL and AMA-gipA.GL growth assays. We prepared total RNA from freeze-dried mycelia using the TRIzol method [62]. We carried out RNA labeling reactions and hybridizations as described in the J. Craig Venter Institute Microarray Protocols (http://pfgrc.jcvi.org/index.php/microarray/protocols.html). We repeated all the hybridizations in dye-swap sets. We scanned and analyzed hybridized slides as described previously [67]. We averaged all replicates for the official data used in our analysis.
We injected the dorsal side of the thorax of CO2-anesthetized, adult, Toll-deficient D. melanogaster flies with a sterile 0.25 mm needle that had been dipped in a solution containing 107 spores/ml of A. fumigatus conidia. We infected 20–25 flies per strain for each virulence assay, which was repeated at least twice. Flies were kept in a 29°C incubator to maximize susceptibility to microbial challenge and monitored for 7 days. Flies that died within 3 hrs of the injection were not included in the survival graph, as these flies most likely died as a result of the puncture wound.
We amplified the 5′ flanking region (FR) and the 3′ FR from gipA using primers 01910 5′ F, 01910 5′ R, 01910 3′ F, and 01910 3′ R. We engineered a unique NotI site into 01910 5′ F to linearize the final deletion construct. We amplified the fragments from Af293 genomic DNA using e2TAK DNA polymerase (Takara Bio Inc., Otsu, Shiga, Japan) and following manufacturer recommendations. We cloned the gipA 5′ FR into pDONR P4-P1R and we cloned the gipA 3′ FR into pDONR P2R-P3, using BP recombination reactions (Invitrogen, Grand Island, NY). We transformed BP reaction mixes into TOP10 cells (Invitrogen, Grand Island, NY) by electroporation, as recommended by the manufacturer. We grew transformed cells on LB (described above) +50 µg/ml kanamycin at 37°C overnight. We picked colonies and transferred to 2 ml LB liquid +50 µg/ml kanamycin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany). We digested plasmid DNA with specific enzymes to verify the correct fragment orientation.
To create the deletion constructs, we combined pDONR P4-P1R-gipA 5′ FR, pDONR P2R-P3-gipA 3′ FR, and pDONR 221-AnpyrG [63] in an LR recombination reaction with pDEST R4-R3 as the destination vector (Invitrogen, Grand Island, NY) [68]. We transformed the LR reaction mix into TOP10 cells (Invitrogen, Grand Island, NY) by electroporation, as recommended by the manufacturer. We grew transformed cells on LB +100 µg/ml ampicillin at 37°C overnight. We picked colonies and transferred to 2 ml LB liquid +100 µg/ml ampicillin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany). We digested plasmid DNA with specific enzymes to verify the correct fragment orientation. We grew bacterial cultures containing the correct plasmid in 250 ml LB liquid +100 µg/ml ampicillin overnight in a 37°C shaking incubator. We purified plasmid DNA by banding on cesium chloride ethidium bromide gradients [65]. We designated the plasmid pDEST R4-R3-gipA 5′ FR-AnpyrG-gipA 3′ FR.
We grew A. fumigatus 1160 (obtained from FGSC) in MAG supplemented with uridine and uracil. We performed the transformation as previously described [64], linearizing the deletion construct with NotI to facilitate homologous recombination. We grew transformants on MMV supplemented with 0.2 M sucrose at 37°C for 3–5 days. We screened for mutants that were prototrophic for uridine and uracil. We prepared genomic DNA from transformant strains [66], and we identified deletion mutants by Southern blot analysis (Fig. S4) [65]. We made a DNA probe using the gipA 5′ and 3′ FRs to verify ΔgipA.
To create a vector for complementation, we PCR amplified the gipA coding region, flanked by a 3 kilobase promoter region and 500 base pair terminator region, using primers gipA 3 kb F and 6g01910 R. We amplified the fragment from Af293 genomic DNA using e2TAK DNA polymerase (Takara Bio Inc., Otsu, Shiga, Japan) and following manufacturer recommendations. We cloned the PCR product into pDONR HPH B [63], as described previously for pDHGL and designated this vector pDONR HPH-gipA. We grew ΔgipA in MAG medium and performed a transformation as previously described for pDHGL. For Southern hybridization, we made a DNA probe using the coding region of gipA (Fig. S5). We designated this strain as gipA(R).
We obtained controls (1160G and ΔgliZ) for growth assays. We created 1160G by transforming pDONR G [63] into Af1160 and we created ΔgliZ by transforming pDEST R4-R3-gliZ 5′ FR-AnpyrG-gliZ 3′ FR into Af1160, performed as described above for ΔgipA. We created the gliZ deletion construct as described previously for ΔgipA using primers gliZ 5′ F, gliZ 5′ R, gliZ 3′ F, and gliZ 3′ R. We used gliZ 5′ and 3′ flanking regions as a DNA probe for Southern hybridization (Fig. S4).
We grew ΔgipA on YAG supplemented with uridine and uracil +1 mg/ml 5-Fluororotic acid (5-FOA) at 37°C. Colonies that reverted to a pyrG- phenotype grew as outgrowths from the original streak. We prepared genomic DNA from these pyrG- outgrowths [66] and tested them for the presence of gipA by Southern hybridization (Fig. S6) [65]. We used the gipA coding region as a DNA probe. We designated this mutant as ΔgipA.0. We transformed pDEST R4-R3-gliZ 5′ FR-AnpyrG-gliZ 3′ FR into ΔgipA.0, as described above for ΔgipA. We used gliZ 5′ and 3′ flanking regions as a DNA probe for Southern hybridization (Fig. S7).
We amplified the C2H2 DNA binding region from gipA using primers gipA C2H2 F and gipA C2H2 R. We used Af293 as template and AccuPrime Pfx DNA polymerase (Life Technologies, Grand Island, NY) and per manufacturer's recommendations. We cloned the PCR product into pDONR 221 using a BP recombination reaction (Invitrogen, Grand Island, NY). We transformed the reaction mix into TOP10 cells (Invitrogen, Grand Island, NY) by electroporation, as recommended by the manufacturer. We grew the transformation mix on LB +50 µg/ml kanamycin at 37°C overnight. We picked colonies and transferred to 2 ml of LB liquid +50 µg/ml kanamycin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany). We digested plasmid DNA with specific enzymes to verify the correct insertion. We recombined the gipA C2H2 region into pDEST 15 using an LR recombination reaction (Invitrogen, Grand Island, NY). We transformed the reaction mix into TOP10 cells (Invitrogen, Grand Island, NY) by electroporation, as recommended by the manufacturer. We grew the transformation mix on LB +100 µg/ml ampicillin at 37°C overnight. We picked colonies and transferred to 2 ml of LB liquid +100 µg/ml ampicillin to grow overnight in a 37°C shaking incubator. We isolated plasmid DNA from each culture using a miniprep kit (Qiagen, Hilden, Germany). We digested plasmid DNA with specific enzymes to verify the correct insertion. This vector, pDEST 15-gipA C2H2, was used in a protein binding microarray analysis as previously described [69].
We created mutated gipA DNA binding sites in pDHGL using a QuikChange II XL Site-Directed Mutagenesis Kit (Agilent Technologies Inc., Santa Clara, CA), as recommended by the manufacturer. We utilized primers BSM1 F and BSM1 R, BSM2 F and BSM2 R to create vectors pDHBSM1 and pDHBSM2, respectively. We transformed pDHGL, pDHBSM1 and pDHBSM2 into Af293.1, as previously described for Af293.1-GL, except we plated transformants on YAG supplemented with uridine and uracil (described above), 0.2 M sucrose, and 300 µg/ml hygromycin. We verified correct transformants by Southern hybridization using the lacZ coding region as a probe (Fig. S8). We designated these strains as Af293.1-GL, Af293.1-BSM1 and Af293.1-BSM2, respectively. We obtained two independent isolates for each binding site mutant to verify that positional effects were not contributing to our results. To test the effect of the different gipA binding site mutants, we transformed pDONR AMA, pDONR AMA-gliZ, and pDONR AMA-gipA into two independent isolates of Af293.1-GL, Af293.1-BSM1, and Af293.1-BSM2, as described above for AMA-gipA.GL, except we grew transformants on YAG supplemented with 0.2 M sucrose. We designated these strains as AMA.GL, AMA-gliZ.GL, AMA-gipA.GL, AMA.BSM1, AMA-gliZ.BSM1, AMA-gipA.BSM1, AMA.BSM2, AMA-gliZ.BSM2, and AMA-gipA.BSM2.
We ground 50 µl lyophilized mycelia to a fine powder with acid-washed glass beads (400–650 µm) (Sigma-Aldrich Corp., St. Louis, MO). We suspended the ground powder in 200 µl protein extraction buffer (PEB) (60 mM Na2HPO4·7H2O, 40 mM NaH2PO4·H2O, 10 mM KCl, 1 mM MgSO4·7H2O, 1 mM EDTA, and 20 µM PMSF [added fresh], pH 7.0) by vortexing and incubated the samples on ice for 15 min, with additional vortexing every 5 min. We spun tubes for 15 min at 15,600× g at 4°C to pellet cellular debris and beads. We transferred supernatants, containing total protein, to fresh tubes on ice and measured protein concentration using a Bio-Rad protein assay kit (Bio-Rad, Hercules, CA). In a 96-well plate, we added 10 µl of protein in PEB and 90 µl of Z Buffer (60 mM Na2HPO4·7H2O, 40 mM NaH2PO4·H2O, 10 mM KCl, 1 mM MgSO4·7H2O, 50 mM β-mercaptoethanol [added fresh], pH to 7.0). For samples grown in repressing conditions (CM), the total protein added was 1 µg. For samples grown in non-repressing conditions (CD), the total protein added was 0.1 µg. To begin the β-galactosidase assay, we added 20 µl of 2-Nitrophenyl-β-D-galactopyranoside (ONPG) (Sigma-Aldrich Corp., St. Louis, MO), diluted to 4 mg/ml in Z Buffer, and placed the 96-well plate in a 37°C incubator. We timed reactions and stopped samples with 50 µl 1 M Na2CO3. We measured absorbance at OD420 and calculated Units of β-galactosidase activity/mg protein with the following equation: (OD420×TV)/(0.0045×T×V×C), where TV is total volume of the reaction in ml, T is time in min, V is volume of protein added in ml, and C is concentration of protein used in µg/µl.
We grew Af293.1 in MAG supplemented with uridine and uracil (described above). We transformed pDEST R4-R3-gliZ 5′ FR-AnpyrG-gliZ 3′ FR and pDEST-gipA 5′ FR-AnpyrG-gipA 3′ FR into Af293.1, as previously described for ΔgipA, except we plated transformants on YAG supplemented with 0.2 M sucrose. We used gliZ 5′ and 3′ FRs and gipA 5′ and 3′ FRs for Southern hybridization and designated these strains ΔgliZ.1 and ΔgipA.1 (Fig. S9). We chose one strain from the ΔgliZ transformation that did not show homologous recombination at the gliZ locus, but was prototrophic for uridine and uracil. We designated this strain pyrG+. We collected total RNA and performed dot blot analysis on pyrG+, ΔgliZ.1, and ΔgipA.1 to verify loss of gliZ and gipA, respectively (described in RNA dot blot Analysis) (Fig. S10). We cloned gliZ and gipA into pDONR AMA/HPH [63], as described above for pDONR AMA-gliZ and pDONR AMA-gipA, respectively, except we used primers gipA 3 kb F and 6g01910 R for the gipA cassette. We designated these plasmids pDONR AMA/HPH-gliZ and pDONR/HPH-gipA. We transformed these plasmids, along with pDONR AMA/HPH empty vector, into pyrG+, ΔgliZ.1, and ΔgipA.1, as described above for AMA-gipA.GL, except we grew transformants on YAG supplemented with uridine and uracil, 0.2 M sucrose, and 400 µg/ml hygromycin. We designated these strains as AMA.G, AMA-gliZ.G, AMA-gipA.G, AMA.Z, AMA-gliZ.Z, AMA-gipA.Z, AMA.A, AMA-gliZ.A, and AMA-gipA.A.
|
10.1371/journal.pgen.1000926 | Human Telomeres Are Hypersensitive to UV-Induced DNA Damage and Refractory to Repair | Telomeric repeats preserve genome integrity by stabilizing chromosomes, a function that appears to be important for both cancer and aging. In view of this critical role in genomic integrity, the telomere's own integrity should be of paramount importance to the cell. Ultraviolet light (UV), the preeminent risk factor in skin cancer development, induces mainly cyclobutane pyrimidine dimers (CPD) which are both mutagenic and lethal. The human telomeric repeat unit (5′TTAGGG/CCCTAA3′) is nearly optimal for acquiring UV-induced CPD, which form at dipyrimidine sites. We developed a ChIP–based technique, immunoprecipitation of DNA damage (IPoD), to simultaneously study DNA damage and repair in the telomere and in the coding regions of p53, 28S rDNA, and mitochondrial DNA. We find that human telomeres in vivo are 7-fold hypersensitive to UV-induced DNA damage. In double-stranded oligonucleotides, this hypersensitivity is a property of both telomeric and non-telomeric repeats; in a series of telomeric repeat oligonucleotides, a phase change conferring UV-sensitivity occurs above 4 repeats. Furthermore, CPD removal in the telomere is almost absent, matching the rate in mitochondria known to lack nucleotide excision repair. Cells containing persistent high levels of telomeric CPDs nevertheless proliferate, and chronic UV irradiation of cells does not accelerate telomere shortening. Telomeres are therefore unique in at least three respects: their biophysical UV sensitivity, their prevention of excision repair, and their tolerance of unrepaired lesions. Utilizing a lesion-tolerance strategy rather than repair would prevent double-strand breaks at closely-opposed excision repair sites on opposite strands of a damage-hypersensitive repeat.
| Telomeres consist of a repeated sequence located at each end of each chromosome. This repeated sequence is required for chromosomal stability and integrity, a function important for both cancer and aging. The DNA sequence of human telomeres is 5–10 kb of a repeated double-strand hexamer (5′TTAGGG/5′CCCTAA). In theory, this sequence is nearly optimal for acquiring UV-induced DNA damage. We developed a novel technique, the immunoprecipitation of DNA damage (IPoD), to study DNA damage induction and repair in the telomere and in coding regions (p53, 28S rDNA, and mitochondrial DNA). We find that human telomeres are hypersensitive to UV-induced DNA photoproducts and that the removal of those DNA photoproducts is almost absent. Cells containing persistent high levels of telomeric DNA damage nevertheless proliferate and chronic UV irradiation of cells does not accelerate telomere shortening. Telomeres are therefore unique in at least three respects: their biophysical UV sensitivity, their prevention of excision repair, and their tolerance of unrepaired lesions.
| Telomeric DNA consists, in all eukaryotes examined to date, of a tandemly repeated sequence located at each end of each chromosome. In humans, it is constituted of 5–10 kb of a repeated hexamer (5′TTAGGG/5′CCCTAA). Telomeres are required for chromosomal stability and integrity (reviewed in [1]).
Telomeres are hypersensitive to single-strand DNA damage induced by oxidative stress. This is thought to be due to the fact that sequences containing guanine triplets are highly sensitive to oxidation [2], [3]. When inserted in a plasmid, telomere sequence is 7-fold more sensitive to Fe2+/H2O2-induced strand breakage than bulk sequence [2]. Moreover, breaks induced in telomeres are repaired significantly more slowly than in other sequences, including interstitial guanine rich repetitive sequence tracts; repair is still incomplete after 19 days compared to complete repair at 1 day elsewhere [4]. In addition, the oxidation of telomeric DNA contributes to their premature shortening. The frequency of oxidative DNA damage at the telomere correlates with the amount of telomere lost during subsequent rounds of DNA replication [5]. It was proposed that the telomere enters DNA replication with greater oxidative DNA damage than the rest of the genome and this elevated damage contributes to telomere shortening [6]. Contrasting with this hypothesis, however, it has been shown that telomere shortening induced by oxidative DNA damage can be replication independent [3].
Ultraviolet light-induced DNA damage has been used for decades as a model to study DNA damage induction and repair. It is biologically relevant because UV is a complete carcinogen, requiring no additional treatments for tumor development, and is the preeminent risk factor in skin cancer development. The vast majority (>80%) of UV-induced damage in B-form DNA consists of cyclobutane pyrimidine dimers (CPD) [7], [8]. CPDs are intra-strand DNA lesions formed when two adjacent pyrimidines are joined across their 5–6 double bonds due to UV-excitation of one of them. The most frequent is the TT cyclobutane dimer [9]. These photoproducts are repaired by the nucleotide excision repair (NER) pathway, which nicks the DNA backbone and excises the damaged segment. Theoretically, the telomere sequence constitutes a perfect target for UV-induced DNA damage. First, the TT on the G-rich strand is repeated thousands of times in each chromosome. On the other strand, the 5′CCCTAA3′ would nominally generate low frequency CC and CT CPD, but two factors supervene: tracts of adjacent pyrimidines tend to generate multiple CPDs on the same molecule, due to cooperative denaturation of the helix by each successive CPD [10] and A:T tracts tend to transfer energy down the base stack until depositing it at a G:C pair [11], [12]. These potential CCCT dimer tracts are again repeated thousands of times in each chromosome. These considerations suggested that this sequence might constitute a hotspot for UV-induced damage. The presence of potential hotspots on both telomeric strands then raises the following spectre: if the cell attempts to simultaneously repair two nearby CPDs on opposite strands, the twin incision nicks would mimic a double-strand DNA break [13]–[15], triggering a DNA damage response and chromosome aberrations [16], [17].
Studying DNA damage induction and repair in the telomere is challenging. The vast majority of the techniques used to study DNA damage induction and repair in a specific part of the genome are PCR based [18]. Because telomeres are constituted of repeated sequences, there are no unique PCR-primer sites. Mismatch primers have been developed to analyze human telomere length by quantitative PCR [19]. However, since those primers can bind to any repeat element of the telomere sequence, they cannot be used in standard techniques to study DNA damage induction and repair, which rely on having one or two known DNA ends. An older study used a single-enzyme modification of the telomere restriction fragment technique (TRF) to study UV-induced CPD in telomeres [20]. However, it is now known that the TRF technique does not provide information on the true length of telomeres [21]: restriction enzymes used to cleave non-telomeric DNA (HinfI or RsaI) give TRF lengths that depend on the site of restriction in the pre-telomeric region. The situation is exacerbated by the fact that achieving complete digestion of genomic DNA using a single restriction enzyme is challenging. Thus studying the induction of DNA damage using the TRF technique does not provide information exclusively about telomeres but about a mixture of telomeric and pre-telomeric DNA. Pre-telomeric DNA is now known to be one of the most rapidly-repaired regions of the genome [22], skewing lesion measurements if this region is included.
We developed a novel method, based on the chromatin immunoprecipitation technique (ChIP), to study single-strand DNA damage. This technique, “immunoprecipitation of DNA damage” (IPoD), allows the separation of damaged DNA from undamaged. The result is two fractions that can each be quantitated by PCR using primers specific for the gene under study. Previously developed primers specific for the human telomeric sequence [19] can be used in this technique, allowing the study of single-strand DNA damage induction and repair in this region. Using the IPoD technique, we have studied UV-induced CPD induction and repair in the telomere as well as in the p53 tumor suppressor gene, in 28S ribosomal DNA, and in a portion of mitochondrial DNA. We find that the telomere sequence is highly sensitive to the induction of CPD by UV light. Moreover, we show that the repair of those UV-induced CPD in telomeres is nearly absent.
IPoD is based on the ChIP technique [23]. Instead of immunoprecipitating a protein covalently cross-linked to DNA, IPoD directly immunoprecipitates DNA fragments containing a DNA structural alteration. Here we use the IPoD technique to study the CPD damage induced on a DNA strand by UV radiation [9]. The technique is schematized in Figure 1A. As the level of DNA damage in a specific region of the genome increases, the number of immunoprecipitated fragments from this region will increase. UV-irradiated DNA, but not unirradiated DNA, yielded an IP fraction using antibody against CPD but not with antibody to Bcl-xL protein or with antibody omitted (Figure S1). UVC has been used in this study to minimize the introduction of photosensitized oxidative DNA damage that accompanies UVB.
The quantity of specific genomic DNA fragments present in the IP fraction was measured, after removing CPDs using photolyase, by PCR amplification using primers specific for the p53 tumor suppressor gene, the 28S ribosomal RNA repeat region, the CYTB gene of mitochondrial DNA, and telomeric DNA. The telomere sequence is composed of a 6-mer concatenated to greater than 5 kb, complicating the design of PCR amplification primers. A 5′ 21-mer primer composed of telomeric repeats is certain to have a complement on the 3′ primer, so primers will anneal together instead of annealing to the telomeric DNA target. Cawthon [19] describes telomeric primers containing mismatches that prevent primers from annealing to each other, thus achieving preferential annealing to telomeric DNA. Because the particular site at which any primer binds on the telomere sequence is random, the resulting PCR product is not a sharp band but a smear.
For the exponential PCR process to be used quantitatively, it must contain an internal control, as in real-time PCR, or be carried out so that all samples have been amplified by the same factor of 2n, that is, with all samples lying on the log-linear part of the amplification curve so that they can be compared to a calibration curve. No internal control is possible with IP, so we adjusted the amount of starting DNA material and the number of PCR cycles to achieve log-linearity for each primer. Figure 1B (upper two panels) shows that the signal from the PCR amplified IPoD-immunoprecipitated DNA is proportional to the UVC dose for 3 different genomic regions. Each genomic region's signal is normalized to that region's signal at 20 J/m2. The signal was linear up to 30 J/m2 UVC (Figure S2). Above this dose range, the slope decreased. Doses above 20 J/m2 UVC are lethal so the present experiments did not enter that range. The high-dose slope reduction could be due to sustaining more than 1 CPD per DNA fragment, saturating the anti-CPD antibody with CPDs, or depleting PCR reagents. Linearity at doses below 30 J/m2 UVC indicates that: a) CPDs are not missed because they occur in DNA segments that will already be IPd due to another CPD; b) the many telomere copies do not saturate the PCR reaction; and c) CPDs or (6-4) photoproducts remaining in the fragment during PCR do not cause a dose-dependent dropout of sample.
To confirm the last point, we also amplified the IP fraction without first reversing remaining CPD with photolyase (Figure 1B, lower panel). When normalized to the signal at 20 J/m2, the shape and slope of the dose-response curve were unchanged for both single-copy and repeat genes. Because PCR-blockage is sometimes used as a relatively insensitive lesion assay, this might seem paradoxical. But the goal of the PCR blockage assay is to determine whether the extent of amplification is reduced compared to undamaged DNA, by measuring the percentage of fragments that have no lesions between the PCR primers. In contrast, IPoD has already identified the CPD-containing fragments via the IP step, so the CPDs can lie outside the PCRd region. The IPoD amplification serves only to make visible a particular set of CPD-containing fragments present in the IP sample. Even when some photoproducts are present, as in the absence of photolyase, the signal is nearly normal: a) 60% of the ∼750 bp sheared fragment lies outside the ∼300 bp PCR fragment; thus, even if photoproducts are a complete block to PCR, the PCR primers are assaying a CPD-target region external to the PCR primers rather than internal plus external. b) Diminution of a gene's PCR signal due to a photoproduct internal to the primers is equal between genes, on average, because every IPd molecule has by definition at least one cyclobutane dimer and, at the UV doses used, typically no more than one dimer per molecule. c) PCR inhibition is only partial because i) Taq polymerase can slowly bypass CPDs [24] and ii) partially-extended fragments will, in the next PCR cycle, anneal to a different partner and extend further; thus the internal region is sampled as well.
To compare the level of UV-induced CPDs in telomeres with the level in other genomic regions, we calculated for each region the absolute percentage of the input that was IPd (IP/Input). This absolute number circumvents differences in PCR efficiency or copy number. The IP fraction was amplified using primers for p53, 28S rDNA, and the telomere after removing CPD with photolyase. For the corresponding Input DNA, various dilutions were amplified and a calibration curve of PCR signal vs dilution was constructed. The PCR signal from the IP was compared with the curve to determine the dilution factor matching the IP signal, and thus the ratio IP/Input.
At 20 J/m2, 14% of the telomeric DNA fragments were damaged (Figure 2), whereas approximately 2% of fragments from the p53 or 28S rDNA genes were damaged at the same dose. The same ratios were obtained whether or not remaining CPD were reversed with photolyase prior to PCR amplification (not shown). Therefore, the telomeric region is 7 times more sensitive than two other regions of the genome. To determine whether one of the telomeric DNA strands was responsible for this sensitivity, we examined the strands separately. Because each strand of the telomere contains only 3 of the 4 possible nucleotides (only GAT for the 5′TTAGGG strand and only ATC for the 5′CCCTAA strand), we performed a strand-specific amplification of the telomere by omitting one nucleotide from the reaction. In addition, an initial linear amplification using only one of the 2 primers and 3 of the 4 nucleotides was performed for 30 cycles. Linear amplification was followed by a standard PCR amplification of the linear-amplified DNA (see Materials and Methods). Each strand was more sensitive than p53 or rDNA (Figure 2), with 16% of the 5′CCCTAA strand fragments being damaged at 20 J/m2 UVC and 6% of the 5′TTAGGG strand.
The telomeric sensitivity was not due to a difference in the frequency of dipyrimidine sites (the site of formation of cyclobutane dimers). This frequency was 29.5 dipyrimidine sites per 100 nucleotide in the p53 fragment, 28.9 in the rDNA fragment, and 33.3 in the telomere. We also examined possible artifactual explanations for the telomeric sensitivity. First, repeated DNA at the ends of chromosomes might sonicate differently, producing more-readily IPd fragments. A Southern blot showed that the sizes of sheared telomere and p53 DNA are the same (Figure S3A). Second, telomeric DNA might have a conformation more accessible to antibody or enzymes. A similar Southern blot experiment revealed that photolyase could completely reverse cyclobutane dimers in both telomeres and p53, suggesting that, at least in naked DNA, accessibility differences do not play a role (Figure S3B). Thirdly, the large number of telomeric repeats might create shorter PCR fragments, which would PCR more efficiently. But the number-average molecular weight of the telomere smear is 250–500 bp, the same range as the ∼300 bp p53 and 28S bands. Finally, we considered that more ‘copies’ of the (diluted) telomeric repeat are present in the PCR reaction than are p53 copies, but this is also true for its pre-IP control.
To confirm the UV hypersensitivity of telomeres independently of IPoD, and to test whether the telomere's hypersensitivity was due to its DNA sequence independent of telomere-bound proteins such as shelterins or chromatin-induced DNA conformation, we examined synthetic oligonucleotides. Four different double-stranded 102-mer oligonucleotides were constructed in which the central 60 bp were varied to include either: 10 repeats of the telomere sequence (5′TTAGGG/CCCTAA) (“Telomere”), 10 repeats of a modified 6-mer (5′TTCAGG/CCTGAA) having the same number of potential UV photoproduct sites (dipyrimidine sites) (“Repeats”), or a single random sequence containing the same number of potential UV photoproduct sites (two examples, “Equi-diPyr #1” and “Equi-diPyr #2”). Each 102-mer was irradiated with either 100 or 500 J/m2 UVC (0.1 – 0.5 CPD per molecule). The irradiated double strand oligonucleotides were directly applied onto a nylon membrane (without PCR amplification) using a dot-blot apparatus and CPD-containing DNA was detected using a CPD-specific antibody (Figure 3). The quantification shows that the telomere repeat was 5 times more sensitive to UVC-induction of CPD than either of the non-repeated sequences. Surprisingly, the non-telomeric 6-mer repeat (5′TTCAGG) was 3 times more sensitive than the random (non-repeated) sequences. This result suggests that repeatedness per se renders dipyrimidine-containing oligonucleotides more sensitive to UV, with telomeric sequences being particularly sensitive.
To determine the number of repeats needed to confer sensitivity to CPD formation, we designed 102-mer double-strand oligonucleotides having increasing numbers of telomeric repeats (“Telo” series). As control, oligonucleotides were designed to have a dipyrimidine-containing region of the same length as the corresponding telomeric repeats but not arranged as repeats (“Equi” series). Outside the repeated region or the corresponding dipyrimidine-containing region, the oligonucleotide lacks dipyrimidine sites. For the “Equi” series, increasing the length of the dipyrimidine-containing region linearly increased the number of CPDs induced, as expected (Figure 4). The Telo series behaved similarly to the Equi oligonucleotides up to 4 repeats. Strikingly, a positive effect of repeats on UV induction of CPD appeared around 5 repeats, as if the DNA had undergone a phase transition. The oligonucleotide containing 5 telomere repeats was 3 times more sensitive than the non-repeated oligo. At 7 repeats, a plateau was reached at which sensitivity to CPD formation was 4–5 times greater in the oligonucleotide containing repeats than in the non-repeated oligo. Limitations on synthesizing longer telomeric oligonucleotides prevented us from determining whether the UV-susceptibility of repeats continues to increase with repeat number – with the plateau merely reflecting the fact that double-strandedness is partially lost at DNA ends – or truly plateaus due to complete acquisition of an altered conformation.
To examine photoproduct repair, sub-confluent human diploid lung fibroblasts (WI38) were irradiated with a minimally lethal dose of UVC (10 J/m2) and harvested at different time points 0–48 hours post-irradiation. Photoproduct-containing DNA was then isolated using IPoD and, after photoreversing CPD, amplified using primers specific for the telomere region (“Telomere”), mitochondrial DNA (“mtDNA”), the gene for the RNA component of ribosomal subunit 28S (“28S”), and tumor suppressor gene p53 (“p53”). p53 was used as a positive control for fast repair by the transcription-coupled NER system (TCNER) [25] because p53 is actively transcribed in human cells, especially after a genotoxic stress such as UV irradiation. The repair rate observed here will reflect both DNA strands and thus will be an average of TCNER on the transcribed strand and slower global genomic NER (GGNER) on the non-transcribed strand. CPDs in the 28S gene of mammalian cells are known to be repaired only by GGNER and not by TCNER [26]–[28], so it serves as a positive control for normal GGNER. In contrast, NER proteins are not present in mitochondria and CPD are not repaired in mtDNA [29]–[31]; thus mtDNA serves a negative control for repair and would indicate any apparent photoproduct loss due to cell dilution during replication.
We found that, 48 hours post-UVC, approximately 70%, 40% and 10% of CPD were removed from p53, 28S and mtDNA DNA regions, respectively (Figure 5A). Repair of CPD in the telomere region was comparable to or less than that seen in the mtDNA negative control, less than 10% after 48 hr, indicating that the NER system is ineffective in telomeres. To ensure that the lack of repair in the telomere region was not specific to the cell line used, the growth stage, or the UV dose, the experiment was repeated in confluent (quiescent) skin fibroblasts at 20 J/m2 UVC, with the same result (Figure 5B).
Cyclobutane pyrimidine dimers are profound blocks to DNA replication forks in mammalian cells, triggering cell cycle arrest and DNA damage responses through the ATR pathway [32], [33]. Oxidative damage at telomeres interferes with maintenance of the D loop and induces telomere shortening [3], [6], [34]. To determine how the elevated and persistent levels of CPD affect telomere maintenance, we investigated UV-induced telomere shortening.
Human diploid fibroblasts were chronically irradiated with minimally-lethal doses of UVB, receiving 0 to 200 J/m2 UVB 1 day after each passage (approximately every 5 days). After 16 passages, DNA was isolated and approximate telomere length was measured using the telomere restriction fragment (TRF) technique [35]. At passage 12 (“X12”), the mean telomere length of un-irradiated cells was approximately 12 kb. At passage 28 (“X28”), the telomere length was approximately 8 kb (Figure 6), corresponding to the expected telomere shortening with increasing passage level. Irradiating cells with 10 to 200 J/m2 of UVB 16 times did not increase the rate of telomere shortening. Therefore, a) normal telomere shortening is not accelerated by unrepaired CPD and b) unrepaired CPDs are not removed by telomere shortening. Evidently, the telomere possesses an efficient tolerance mechanism for cyclobutane pyrimidine dimers.
Every eukaryote has telomeric DNA consisting of short sequences repeated thousands of times at the end of each chromosome. To these repeats has been attributed the role of preserving genome integrity via the stabilization of chromosomes. By “capping” chromosomes ends, telomeres protect them from recombination. A role of “longevity clock” has also been attributed to telomeres. Eukaryotes begin life with full-length telomeres and, at each cell division, telomeres shorten to finally reach a point where the cell enters into a senescence state. In view of these critical roles in genomic integrity, the telomere's own integrity should be of paramount importance to the cell. The present results show that telomeres are unique in at least three unexpected respects: their biophysical sensitivity, their prevention of repair, and their tolerance of unrepaired lesions.
Telomeres were found to be 7 times more sensitive to UV-induced CPD than other DNA regions (Figure 2). This observation was made in a cellular context, so the proteins and secondary structure of the chromatin might be involved in this hypersensitivity. To distinguish these possibilities, we tested the sensitivity of telomere sequence inserted in a 102-mer oligonucleotide. Because this oligonucleotide was irradiated in vitro, it was free of any cellular context. This oligonucleotide showed hypersensitivity comparable to the telomere DNA sequence in vivo (Figure 3). Thus the cellular context is not the major contributor to the UV hypersensitivity of the telomere. What, then, can explain it? We tested different oligonucleotides for UV sensitivity and found that short repeats, like the telomeric sequence, render those oligonucleotides more sensitive. An oligonucleotide containing 10×6-mer repeats was ∼5-times more sensitive than an oligonucleotide containing the same frequency of dipyrimidine sites but randomly distributed (not in repeats). Surprisingly, the sensitivity of the telomeric repeat underwent a sudden transition at 5 repeats, suggestive of a structural phase change (Figure 4). This result means that the expected sensitivity based on DNA sequence is not the entire source of UV sensitivity. The biophysical nature of this transition, and its effect on the distribution of DNA photoproducts, will require biophysical investigation. G-rich single strands undergo a variety of interactions such as Hoogsteen base pairing and G-G stacking. These can create G quadruplexes, parallel-stranded helixes, A- and Z-form DNA, hairpins, and local melting. In telomeric and trinucleotide repeats, the stability of the various structures depends on the number of repeats [36], [37]. The behavior in double-stranded DNA is less studied.
A region of the genome so critical to cell survival and genomic integrity would be expected to preserves its own integrity after a genotoxic stress. Yet little is known about how telomeric DNA does this. The finding that telomeres are hypersensitive to UV-induced DNA damage prompted the expectation that repair of this DNA damage would be rapid, to prevent DNA damage accumulation in this region. What we found was the contrary. Repair was almost absent in telomere regions, proceeding as slowly as in mitochondrial DNA where NER proteins are absent (Figure 5). Two days after UV irradiation, CPD were still present in telomeres but had been half removed from coding regions (p53 or 28S genes) and probably entirely removed from the transcribed p53 strand. The repair defect could be active or passive. In the passive category, compaction of telomeric heterochromatin may prevent access of repair proteins [38]. Also, telomeric DNA has been reported to have partial A-DNA character [39], which predisposes to trans- rather than cis-isomers of CPD [8]. Little is known about the repair of trans-isomers of CPD and they may be more difficult for the NER system to recognize or remove. In the active category, some of the many protein factors bound to telomeres (reviewed in [40]) may inhibit the repair system in this region.
There are two reasons suppression of excision repair can be desirable. The high frequency of CPDs in the telomere, together with the telomere's repeat nature, may generate multiply damaged sites (MDS). MDS are sites where DNA lesions are closer than ∼20 bp on opposite strands [41]. After the incision nicking that is the first step in excision repair, multiply damaged sites result in double-strand DNA breaks. This has been observed for UV-irradiated DNA containing halogenated nucleotide analogs in close proximity [41]. Double-strand breaks, in turn, are clastogenic and lethal events. At an MDS, displacement of the lesion-containing oligonucleotides during the second step of excision repair will also create overlapping daughter strand gaps [13]–[15]. This event increases the permissible distance between CPDs. In unique-sequence DNA, such MDSs would be rare, but in repeats they could be the rule when photoproduct frequency is high. The absence of telomere shortening after chronic UV irradiation (Figure 6) indicates that, in fact, such double-strand breaks have been avoided.
The fact that cell proliferation was unhindered by chronic UV irradiation, despite the presence of CPD in their telomeres, raises a new question: how can a cell tolerate DNA damage in its telomeres? During mammalian DNA replication, a bulky lesion such as a CPD typically blocks replication fork progression [42], [43]. This blockage leads to single-stranded DNA that activates ATR-dependent stress responses such as G2/M arrest and apoptosis [44]. To avoid these events at unrepaired CPDs, the replication mechanism uses DNA polymerases capable of bypassing CPD. In E. coli, the SOS response activates polV to a translesion synthesis polymerase by transferring RecA-ATP to it from a RecA filament [45]. In human cells, the XPV gene (defective in the xeroderma pigmentosum variant complementation group) codes for pol eta, a polymerase able to bypass CPD by incorporating A opposite a T or C in a CPD (reviewed in [46]). Correspondingly, cells from a squamous cell carcinoma from an XPV patient were found to generate recurrent chromosome abnormalities as they were passaged in vitro. These were dicentric chromosomes, particularly telomere–telomere bridges, indicative of telomeric damage [16]. It seems likely, then, that CPD accumulating in the telomere are especially reliant on bypass to avoid replication gaps. In the absence of bypass, these replication gaps would be frequent enough to trigger telomeric double-strand breaks and telomere–telomere bridges, the same kinds of genetic catastrophes that repair suppression aims to avoid.
Each experiment was performed with two different primary human fibroblast cell strains. The first strain was derived from breast reduction tissue from a healthy 25-year old female [47]. The other strain was the commercially available WI38, derived from lung tissue of a male foetus (ATCC, Manassas, VA). Cells were grown in high-glucose DMEM (Gibco Invitrogen) supplemented with 10% FBS and 1% penicillin/streptomycin. Cells were UV-irradiated at room temperature after replacing the medium with cold sterile phosphate buffered saline (PBS). The two cell strains have different UVC sensitivities (Figure S4). The UVC source was a germicidal lamp emitting at 254 nm. Using UVC rather than UVB avoids potential complications from photosensitized oxygen radical formation. For the telomere shortening experiment, UVB was used to maximize the likehood of telomere shortening; the source consisted of three fluorescent tubes (FS20T12/UVB/BP, Philips) filtered through a sheet of cellulose acetate to eliminate wavelengths below 290 nm (Kodacel TA-407 clear, 0.015 inch thickness; Eastman-Kodak Co.). Dose rate was measured prior to each experiment using a UVX UV-meter (UV Products, Upland, CA).
Purification of the DNA was performed using DNeasy Tissue Kit (Qiagen, Valencia, CA), according to the manufacturer's protocol. Purified DNA was sonicated to 500–1000 bp fragments (Branson sonifier 250, microtip, at 30% power, 3×15 sec on ice), precipitated with NaCl/ethanol, and resuspended in resuspension buffer (0.01% SDS, 1.1% Triton X 100, 1.2 mM EDTA, 16.7 mM Tris-Cl pH 8.1, 167 mM NaCl). DNA was denatured by boiling 10 min, incubated with the CPD-specific antibody (D194-1, MBL, Woburn, MA) [48] overnight at 4°C and then with a rabbit anti-mouse secondary antibody for 1 hour. The anti-CPD antibody was used in molar excess to CPD to ensure that each damaged dipyrimidine was pulled down regardless of its local sequence or slight variations in the binding affinity of the antibody to each dipyrimidine type. Molar excess is indicated by the linearity of the dose-response with respect to substrate (Figure 1B). Antibody-bound DNA was pulled down using Staph A beads (Calbiochem). The bead/DNA complexes were washed 2 times with wash buffer 1 (2 mM EDTA, 50 mM Tris-Cl pH 8.0) and 4 times with wash buffer 2 (100 mM Tris-Cl pH 8.0, 500 mM LiCl, 1% NP40, 1% deoxycholic acid). DNA was eluted from the staph A beads with elution buffer (50 mM NaHCO3, 1% SDS) and the eluted DNA was cleaned using a Qiagen PCR purification kit to remove salts and SDS prior to PCR.
In the indicated experiments, CPD were removed before the PCR reaction (but after the IP step) using cloned E. coli CPD photolyase (kindly provided by Drs. C. Selby and A. Sancar). The CPD photoreactivation mix (10 mM Tris-HCl pH 7.6, 10 mM NaCl, 2 mM EDTA, 20 mM DTT, 0.2 mg/mL BSA, 0.1 µL CPD photolyase) was added to the DNA and exposed for 1 h to UVA light from eight F20T12BL lamps (Spectra Mini, Daavlin Co., Bryan, OH) passed through filters to remove UVB and UVC. The DNA was then cleaned using a PCR purification kit (Qiagen).
For PCR reactions, 20 cycles of amplification were performed on a Biometra TGradient thermal cycler with Taq polymerase in 10 mM Tris/HCl, 1.5 mM MgCl2, 50 mM KCl, pH 8.3 and 200 µM each dNTP (Roche Molecular Biochemicals, Indianapolis, IN). A test run of PCR using different amounts of starting material was done on each sample and on each primer set to ensure the amplification lay in the exponential portion of the amplification reaction. The following primers were used: For the telomere sequence: 5′GGTTTTTGAGGGTGAGGGTGAGGGTGAGGGTGAGGGT and 5′TCCCGACTATCCCTATCCCTATCCCTATCCCTATCCCTA [19]. The underlined bases are mismatched with respect to the telomere sequence. For the p53 gene: 5′CTGCCTCTTGCTTCTCTTTTCC and 5′GGTTTCTTCTTTGGCTGGG, giving a PCR product of 309 bp. For 28S ribosomal DNA: 5′GTAGAATAAGTGGGAGGCCCCCGG and 5′AGGCCCCGCTTTCACGGTCTGTATTCG, giving a PCR product of 368 bp. For the CYTB gene in mitochondrial DNA: 5′CCCTAGCCAACCCCTTAAAC and 5′TTGGCTTAGTGGGCGAAATA, giving a PCR product of 297 bp. The agarose gel was scanned and quantification was done using ImageQuant 5.0 software (Molecular Dynamics). For p53, 28S and mtDNA, the band was simply quantified and the background was subtracted from the signal. For the telomere sequence, the PCR primers can anneal varying distances apart on the telomeric repeat, so the PCR product is not a single-size product but rather an assortment of DNA fragments over a size range. We therefore ran telomeric samples on the agarose gel for a few minutes (to let the DNA enter the gel and to separate the PCR product from the primers), making the smear band-like. The entire smear was quantified using the same technique as for coding regions.
Oligonucleotides used in dot-blot experiments are depicted in Table 1. 400 ng of each double-strand oligo was irradiated with the indicated UVC doses using a 254 nm source. The irradiated DNA was denatured and applied onto a nitrocellulose membrane using a dot-blot apparatus. CPD-containing DNA on the membrane was visualized using a CPD-specific antibody (D194-1, MBL, Woburn, MA) [48] followed by a secondary anti-mouse-HRP antibody (Santa Cruz Biotechnology, Santa Cruz CA) and revealed by chemiluminescence (Denville, Metuchen, NJ). Different film exposures were scanned and quantification was done using ImageQuant 5.0 software (Molecular Dynamics).
Cells were irradiated with different UVB doses (0, 10, 50, 100 and 200 J/m2). After the irradiation, cells from each condition were allow to grow until they reached full confluency. When cells from every exposure condition reached 100% confluency, they were all passaged 1∶4. This precaution was taken to assure that UV-irradiated cells did not undergo fewer population doublings than unirradiated ones at the same passage number. (A disadvantage of this design is that mortality at the higher UV doses would cause more divisions of the remaining living cells to compensate, possibly leading to faster telomere shortening at these doses. However, because telomere shortening was not seen, this absence is conclusive.) UV-induced telomere shortening would be obscured if UV also reduced the number of cell doublings by decreasing the cell density at confluence. This effect would reduce the extent of normal, replication-related, telomere shortening. The cell density reduction apparently did not occur here. Because each cell lineage was split at the same ratio, a 25% reduction in cell density of treated cells compared to untreated would result in a (0.75)16 = 100-fold difference in cell number after 16 passages (from X12 to X28). But no difference in the final amount of DNA harvested was observed between any of the UV doses.
Terminal restriction fragment length measurements were obtained using the Telo TTAGGG telomere length assay kit (Roche Molecular Biochemicals, Indianapolis, IN) as done previously [47]. Briefly, 2 mg of HinfI/RsaI-digested genomic DNA were separated on 0.8% agarose gels and Southern blotted onto a Hybond-N+ nylon membrane (Amersham Biosciences, Piscataway, NJ). After UV-fixation of DNA fragments onto the membrane, membranes were hybridized with digoxigenin-labeled telomere-specific probe (TTAGGG)4. After washing out non-bound probe, membranes were incubated with a digoxigenin-specific antibody covalently coupled to alkaline phosphatase. Finally, the telomere fragments were visualized by a chemiluminescent substrate (CDP-star, Roche Molecular Biochemicals, Indianapolis, IN). TRF lengths were determined by comparing the signals relative to a standard molecular weight using ImageQuant 5.0 software (Molecular Dynamics). All lanes were divided into 75 intervals, and the mean TRF length was defined as S(ODi)/S(ODi/Li), in which ODi is the chemiluminescent signal and Li is the length of the TRF fragment at position I [49]. Although TRF fragments have one terminus in the pre-telomeric region, changes in TRF length reflect changes in telomere length.
|
10.1371/journal.pcbi.1004256 | Dissecting the Calcium-Induced Differentiation of Human Primary Keratinocytes Stem Cells by Integrative and Structural Network Analyses | The molecular details underlying the time-dependent assembly of protein complexes in cellular networks, such as those that occur during differentiation, are largely unexplored. Focusing on the calcium-induced differentiation of primary human keratinocytes as a model system for a major cellular reorganization process, we look at the expression of genes whose products are involved in manually-annotated protein complexes. Clustering analyses revealed only moderate co-expression of functionally related proteins during differentiation. However, when we looked at protein complexes, we found that the majority (55%) are composed of non-dynamic and dynamic gene products (‘di-chromatic’), 19% are non-dynamic, and 26% only dynamic. Considering three-dimensional protein structures to predict steric interactions, we found that proteins encoded by dynamic genes frequently interact with a common non-dynamic protein in a mutually exclusive fashion. This suggests that during differentiation, complex assemblies may also change through variation in the abundance of proteins that compete for binding to common proteins as found in some cases for paralogous proteins. Considering the example of the TNF-α/NFκB signaling complex, we suggest that the same core complex can guide signals into diverse context-specific outputs by addition of time specific expressed subunits, while keeping other cellular functions constant. Thus, our analysis provides evidence that complex assembly with stable core components and competition could contribute to cell differentiation.
| A key challenge in cellular network biology is to understand how protein complexes are cell-type or condition-specific assembled and disassembled. Cell differentiation is a major cellular reorganization bringing about fundamental changes in the new differentiated cell type. As many genes are expressed throughout all stages and only their expression levels differ, the question arises of how specific functions can be mediated. Here, focusing on the calcium-induced differentiation of primary human keratinocytes, we describe motifs of protein complex assemblies. We found that a large proportion of complexes contain both proteins expressed at similar levels in all stages of differentiation (non-dynamically expressed) and proteins with variable expression between (dynamically expressed). Using structural information we found that subunits tend to be replaced at structural overlapping surfaces of proteins. When applying our concepts to a manually annotated large TNF/NFkB signaling complex, we find a stable core associated with both a dynamically changing module and several stable modules. We propose this as a ‘constant signalosome ready to work,’ where a stable core is associated with a dynamic periphery. Altogether, our analysis highlights the importance of understanding the dynamic assembly and disassembly of complexes, taking 3D structural information into consideration, rather than only considering networks of individual proteins.
| A key question in cellular network biology is how protein complexes assemble and disassemble in a time-dependent manner. Coordinated changes in the transcriptome and proteome occur during cellular differentiation [1–3], during cell reprogramming [4], or after growth factor stimulation [5] to name a few examples. Recent work in yeast has predicted that complexes change in composition during the cell cycle, and that complexes consist of both constitutive but non-dynamically expressed, and dynamically expressed subunits, leading to the proposal of ‘just-in-time assembly’ of complexes [6]. Consistent with this concept, relating expression data in different human cell types and tissues to protein complexes showed that non-dynamically expressed proteins extensively interact with tissue-specific expressed proteins, suggesting a tight interplay between core and tissue-specific proteins [7]. However, the molecular details underlying the assembly of complexes (‘complex assembly motifs’) are largely unexplored. This includes the definition of complexes themselves, e.g. as molecular machines (stably associated complexes) or pleiomorphic ensembles (complexes that assemble on demand) [8]. For example, what is the proportion of complexes that are permanently assembled, changed during different cellular conditions, or contain both non-dynamic and dynamic subunits (‘di-chromatic’ complexes)? Are subunits replaced at structurally overlapping or compatible surfaces of proteins? What is the role of evolutionarily-related paralogs?
Complementing protein interaction networks with three-dimensional structural information for binding interfaces has provided an improved functional understanding of cellular protein networks [9–15]. A recent study combining structural modeling with network analyses has revealed two types of interactions with a common hub protein: mutually exclusive interactions through a single interface (also called ‘XOR’) and compatible interactions through multiple interfaces (also called ‘AND’) [10]. This study has also shown that hub proteins characterized by single interfaces evolve faster and are enriched in paralogs [10]. In another study, modeling ErbB signaling through combining network and structural analyses, has suggested that competing protein interactions at single interface hubs produce variations in signaling responses [14].
Based on the above studies, we reasoned that the assembly of complexes where proteins compete for a common stable core could play a role in cell differentiation. To test this hypothesis and to define complex assembly motifs, we focused on the calcium-induced differentiation of primary human keratinocytes (PHK) as a model system for a large cellular reorganization process [16,17]. The epidermis of mammalian skin develops from a single layer of keratinocytes (interfollicular basal stem cells) into a multi-layered stratified epithelium. Keratinocyte differentiation is a well-suited model system, as differentiation of primary keratinocyte cells can be induced in vitro in cell culture by the addition of calcium [18]. In our previous work we quantified the transcriptome during differentiation, and we identified functionally important circadian oscillations [3]. Here, we performed a detailed analysis of all gene expression changes associated with keratinocyte differentiation followed during 45 hours and integrated this information with protein complexes to analyze their reorganization. We inferred that half of human protein complexes present during differentiation contain both non-dynamically and dynamically expressed subunits. Some di-chromatic complexes contain a stable core that associates with dynamic genes belonging both to similar clusters (concerted gene expression changes) and different clusters (opposing gene expression changes). In many cases, di-chromatic complexes with genes exhibiting opposing expression changes belong to complexes known to be involved in general or keratinocyte-specific differentiation processes and pathways, such as EGF/TGF-α signaling, TNF-α/NFκB signaling, Notch/γ-secretase, ubiquitination, cell cycle arrest, and chromatin remodeling complexes. Using three-dimensional structural modeling, we predicted physical interfaces and distinguished between mutually exclusive (XOR) and compatible (AND) interactions [13,19]. We found that dynamic proteins binding to a common non-dynamic protein are enriched for mutually exclusive interactions, suggesting that changes in complex assemblies can occur through variation in the abundance of proteins that compete for binding to XOR nodes. In addition, compensatory expression changes of paralogs suggest that these proteins—while keeping a constant essential function for cell viability—have differential functionalities, which serve a specific role for cell type-specific functions. Altogether our analysis highlights the importance of understanding the assembly of complexes and taking 3D structural information into consideration, rather than elucidating networks of individual proteins.
Differentiation was initiated in human primary keratinocyte stem cells by the addition of CaCl2 [3, 18]. Cells were harvested over 45 hours at 5-hour intervals with three biological replicates. At each time point cells were lysed, mRNA was isolated, and expression profiles were measured using Agilent microarrays [3] (Fig 1A and S1 Table). 21,113 probes mapping to 16,720 genes were detected as expressed. We defined genes that are changing (‘dynamic’) and those that are constantly expressed (‘non-dynamic’) using a 2-fold expression change cut-off and a Chi-squared test on time points of biological replicates (see methods). As a more stringent criterion for classifying dynamic genes, we used a threshold of 4-fold, defining these genes as ‘super-dynamic’ (S1 Table). In summary, out of the 16,720 expressed genes, 6,137 are ‘non-dynamic’ (P> = 0.01, Chi-squared test, and fold change <2), 6,096 are ‘dynamic’ (P<0.01 and fold change > = 2), 1,317 of the dynamic genes are also ‘super-dynamic’, (P<0.01 and fold change > = 4), and 4,487 genes do not fall into one of the above categories (classified as ‘unresolved’). In general, we used the super-dynamic genes for our analyses, unless otherwise stated.
Classification of proteins encoded by the super-dynamic genes based on DAVID [20,21], UniProt [22], and manual literature searches (e.g. [23]) (S1 Fig) uncovered a large fraction of proteins involved in: i) signaling (transcription factors, and the adhesion, chemokine, calcium, immune/Toll, ubiquitin-like, apoptosis, Wnt, TGFβ, interferon, and Notch signaling pathways (S2 Fig) (32%); ii) housekeeping functions (cytoskeleton, cell cycle, solute transport carriers, histone, and chaperone proteins, or formation of tight junctions (S3 Fig) (22%); iii) metabolic enzymes for lipid, amino acid, steroid, purine, and flavine interconversions and lipid binding (S4 Fig) (13%), iv) proteins needed for the progressive steps towards the formation of the cornified envelope (metallo and serine proteases), crosslinking enzymes (e.g. transglutaminases) and substrates (e.g. loricrin, involucrin, and small proline rich proteins) that provide structural stability and elasticity, keratins (mechanical resistance), and lipid modifying enzymes (water repellence) (S5 Fig) (8%) [23]. For 8% of the super-dynamic genes only one Pfam domain prediction can be assigned, while 17% of them have no known function or Pfam domain annotations. Thus, the process of keratinocyte differentiation is accompanied by concerted changes in metabolic, signaling, and housekeeping pathways. We also confirmed the induction of known markers for keratinocyte differentiation (involucrin [IVL], filagrin [FLG], cystatin [CSTA, CSTB]) (S1 Table) [23]. Involucrin expression was additionally confirmed on the protein level by immunostaining (S6 Fig).
We used K-means clustering to classify the temporal profiles of the 1,317 super-dynamic genes and identified eight optimized clusters (Fig 1B). The clusters are arranged with opposing temporal profiles on the top and on the bottom: clusters 1 and 5 contain early highly transient expressed/repressed genes, clusters 2 and 6 contain early transient expressed/repressed genes, clusters 3 and 7 contain early highly sustained expressed/repressed genes, and clusters 4 and 8 contain early sustained expressed/repressed genes. Consistent with the function of the Ets transcription factor ELF3 in promoting differentiation of keratinocytes [24,25], we see strong immediate and sustained expression of ELF3 (cluster 3), followed by delayed expression of genes regulated by ELF3 like KRT4 and TGFβ (cluster 8), and the envelope protein SPRR2A (cluster 3) [26,27].
The clusters reveal a moderate stage-specific expression of functionally related protein classes and biological processes. For instance, proteins related to the formation of the cornified envelope are strongly induced at later time points (clusters 3, 4, 5, and 8). Cluster 7 (repressed genes) is enriched for cell cycle-related proteins in the housekeeping category, as expected during the onset of differentiation [28]. This cluster has an antagonistic behavior when compared to the cell-cycle enriched cluster published in a cellular reprogramming study [4]. Signaling-related pathways in clusters 1 and 2 involve transiently expressed transcription factors, cytokines, and proteins involved in Wnt, adhesion, TGFβ, and TNF signaling (S1 Table). Likewise, enzymes involved in lipid and amino acid metabolism are present in most clusters. The absence of a strong association between functional categories and clusters suggests a functional replacement within sub-categories, possibility brought about in some cases by dynamic rearrangements within protein complexes.
During keratinocyte differentiation, a switch in the expression of paralogous gap junction subunits is needed for the changes in gap junction permeability required for epidermis formation [29], which we see in our study (Fig 2A). To see if this is a general case for all paralogous pairs, we annotated expressed genes in keratinocytes with homology information from EnsemblCompara [30] limiting our analysis to paralogous pairs in which both genes are dynamically expressed, thereby reducing our set from 11,582 to 2,260 paralogous pairs. Computing Pearson correlation metric for all pairs, we found 950 to have highly correlated (r> = 0.6, P<0.07) dynamic expression profiles during skin differentiation with a subset of 235 qualifying as super-dynamic, and 281 pairs to have highly anti-correlated dynamic expression (r< = -0.6, P<0.07) with 19 of these pairs being super-dynamic (Fig 2B and S2 Table). When compared to super-dynamic and dynamic random pairs, we found that paralagous pairs are more likely to have correlated expression (Wilcoxon rank sum test; P = 4.5e-37, for super-dynamic S7A Fig; and P = 5.4e-61 for dynamic, S7B Fig). An example where gene duplicates can act together to bring about a functional change is the formation of cornified envelope (e.g. the transglutaminase substrates, small proline-rich proteins) (Fig 2C and S8 Fig). This could suggest a requirement for increased gene dosage in particular specific stages of cell differentiation as the likely reason for the duplication of these genes [31–33].
We also found nineteen anti-correlated pairs where both proteins are super-dynamic (Figs 2D–2F and S9 Fig). Examples include PLEKHA 6/7/4 (Fig 2D; for PLEKHA-7 a functional role has been demonstrated in recruiting paracingulin to tight junctions [34]), WNT7A/7B/5A (Fig 2E), NDRG1/2/4 (Fig 2F; NDRG2 is expressed in response to TGFβ and inhibits proliferation [35]) and CCNA/B proteins (S9 Fig). One explanation for this observation might be that there is temporal sub-functionalization, with one paralog being expressed early, the other late during the course of differentiation as the paralogs have taken on different functions. Therefore, the expectation would be that anti-correlated paralogous pairs are more divergent in sequence level than correlated ones. Using Pfam domain annotations we compared correlated to anti-correlated dynamic paralogous pairs (here we used dynamic pairs to increase the numbers for statistical analysis since not every protein has Pfam annotations) and found that correlated paralogous pairs are more similar in their Pfam domain composition (different domain, or domain missing) than anti-correlated pairs (1.3 versus 1.7; Wilcoxon rank sum test; P = 0.009) (Fig 2G and S10 Fig). Likewise, anti-correlated dynamic paralogous pairs have greater sequence divergence (S11A and S11B Fig) and a greater number of differences in amino acid sequence length (S11C and S11D Fig). Furthermore, anti-correlated genes have a higher fraction of evolutionary old duplicated genes compared to correlated genes (S12 Fig).
In summary, dynamic paralogs are enriched in gene pairs that are correlated in expression. This suggests that a requirement for increased gene dosage is the likely reason for the duplication of these genes. Paralogous genes with anti-correlated expression changes have greater amino acid sequence and Pfam domain divergence. This implies that anti-correlated paralogs have taken on specialized roles during differentiation.
The finding that divergent paralogous pairs tend to have anti-correlated expression patterns could suggest replacement of subunits in protein complexes to expand functional complexity. To see if this is a general feature of protein complexes during differentiation, we obtained a list of protein complexes from the CORUM database (a resource of manually annotated protein complexes from mammalian organisms [36]), and mapped it onto the non-dynamically expressed, unresolved, or dynamically changing proteins (Fig 3A and S13 Fig and S3 Table). After removing complexes containing unresolved genes, 19% of the complexes did not have significantly changing gene expression for any of their subunits during the keratinocyte differentiation process (all non-dynamic), 26% of the complexes had dynamically changing expression for all subunits, and for half of the complexes (55%) we find a mix of behaviors (here called ‘di-chromatic’) (Fig 3B). We represented the previously clustered 1,317 super-dynamic genes in the context of complexes to which they belong and also assigned dynamic genes to each one of the eight super-dynamic clusters through comparative correlation analysis with average expression profile of the aforesaid clusters (Fig 4; S1 Network). We found that super-dynamic genes in the same complex or closely connected in the network often belong to different clusters (72%). However, this ratio decreases to 43% if clusters with similar behaviors (1 and 2, 3 and 4, 5 and 8, 6 and 7) are combined.
To further characterize the expression profiles of di-chromatic complexes, we classified the CORUM complexes into functional groups (S3 Table). Then we represented the expression changes for functionally related complexes or complexes sharing components (Fig 5 and S14–S16 Figs). Among many cellular processes known to be important for keratinocyte differentiation or general cell differentiation (Fig 5A), we find examples of complexes with a stable component or core and a dynamic periphery. Examples in cell signaling include EGFR and TNF-α/NFκB pathways. EGF signaling is important for keratinocyte differentiation [37] and we find two complexes involved in its recycling and degradation. In both complexes, one component (STAM2 and CBLB) is constant and the other is super-dynamic (RIN1 and SH3BKP1, respectively), changing in opposite directions in the two complexes. RIN1 regulates EGFR degradation in cooperation with STAM [38], and SH3BKP1 prevents epidermal growth factor receptor degradation by the interruption of c-Cbl-CIN85 complex [39]. Thus, the opposite behavior of these complexes will favor EGFR stabilization and consequently, EGF signaling and keratinocyte differentiation. In this respect we also find anti-correlated changes of the paralogous receptors EGFR and ErbB2, which slow down EGFR recycling through heterodimer formation [40].
The TNF-α/NF-κB signaling pathway is required for normal epidermal development and homeostasis [41–43]. As the CORUM database is not complete [44], we combined the CORUM complex information with a detailed structural and pathway analysis based on the literature [45, 46] (S15 Fig). We identified a core TNF receptors/scaffold complex (TNFRSF1, TRADD, RIPK1, TRAF2, IKBG1, CHUK, CDC37 and HSP90AA1) associated with a MAPK pathway (i.e. MAP2K7/ and MAPK8/9) which is composed of non-dynamic and unresolved genes, while the rest of the pathway in general is composed of dynamic components. This is an example of a stable core module associated with a dynamic peripheral module. Most of the super-dynamic genes are found mainly in the transiently expressed clusters 1 and 2, suggesting relevant concerted changes during keratinocyte differentiation (Fig 5B). MAP3K8 (Cot) is in cluster 4 (early activated and sustained expression) and can form larger complexes containing NFKB1, thereby promoting signaling. Likewise, CFLAR is in the same cluster 4 and is an inhibitor of FADD thereby preventing apoptosis. NF-κB/RelA has been shown to control cell-cycle exit in keratinocytes [47].
NF-κB-independent signaling commencing at the level of the non-dynamic core connects to MAPKs (such as NIK, NAK, TAK1, and Cot) and PKC isoforms [48]. Interestingly we see that TRAF1 expression is in the same cluster as the NF-κB components. It has been described that TRAF1 promoter has NF-κB binding sites and is strongly activated by TNF-α [49]. TRAF1 cannot bind to the TNF receptor directly, but is recruited though binding to TRAF2 [50] and overexpressing TRAF1 does not affect the interaction between TRAF2 and FADD [51], suggesting a compatible complex formation. TRAF1 binding to TRAF2 results in blocking apoptosis [52] and analysis of TRAF1-/- mice suggests that TRAF1 inhibits TNF-α signaling [49]. Thus, TNF-α/NF-κB induced dynamic expression of TRAF1 creates a negative feedback loop that could mediate anti-apoptotic functions [52] and decrease NF-κB activation without affecting other possible constitutive function of TNF receptor. We propose that the stable core module connects to both a dynamic peripheral module important for cell cycle arrest during differentiation (via NF-κB heterodimers and prevents apoptosis (via TRAF1), and several stable modules that function independent of NF-κB by binding to TRAF or the IKK complex and which should play a housekeeping function [48] (see S15 Fig for a summary of different signaling functions).
In the Notch/γ-secretase pathway [53], the γ-secretase (APH1B, promoting differentiation) is up-regulated and delta like 1 (DLL1, blocking differentiation) is down-regulated (Fig 5B). Both proteins associate to the same stable core (PSENEN and NCSTN), reinforcing a biological role for di-chromatic complexes. In the ubiquitination/degradation pathway, the CAND1 assembly factor of the E3 ubiquitin ligase complex [54] is up-regulated transiently and the S-phase kinase-associated protein 2 (SKP2, promoting cell cycle) is down-regulated as the target of the E3 ubiquitin ligase complex (S14A Fig). Skp2–Skp1 abrogates the inhibitory influence of CAND1 on the neddylation of Cul1 by promoting the dissociation of the cullin–CAND1 complex, whereas substrate, together with substrate-presenting components, prevents the action of CSN to deneddylate cullin [55]. It has been described that high levels of Skp2 are needed for proliferation in stratified epithelia [56]. In this respect it is noteworthy that CAND1 also causes elevation of p27, which has been demonstrated to be important during pre-adipocyte differentiation [57].
Finally, two complexes with interesting di-chromatic behavior are related to chromatin remodeling and known to be important for regulating gene expression changes during keratinocyte differentiation [58,59]. Emerin is a nuclear membrane protein, which is involved in tissue-specific gene regulation [60] and expressed constantly during the keratinocyte differentiation (S14C Fig). The emerin binding protein LMO7 is a cell type-specific transcription factor that is strongly up-regulated during differentiation. It acts by escaping actin (ACTB)-mediated inhibition [61,62] and therefore its levels need to increase strongly. At the same time, Laminin B1 (LMNB1; important for DNA replication) is transiently down-regulated. Finally, the BAF (SWI/SNF) complexes are known for their combinatorial assembly providing functional specificities [63,64] (S14D Fig). BRCA1 can directly interact with the BRG1 subunit of the SWI/SNF complex and is down-regulated during differentiation. This may liberate the SWI/SNF complexes to take part in chromatin remodeling, which are either constantly expressed or moderately down-regulated.
Generalizing, there are three different types of assembly motifs: In some complexes, dynamic genes are added or removed from the complexes and these are predicted to be compatible (AND; [14]) interactions (e.g. TNF-α/NFκB signaling and Notch complexes). For other complexes we observed opposing expression changes of subunits, which are potentially competing for the same binding interface (XOR interactions; i.e. CULIN SKP2 CAND1 [14]). Finally, there are large assemblies where we observed a mixed behavior (i.e. EGFR/TGF-β; S16 Fig). In summary, we find a di-chromatic behavior with a stable core for many complexes involved in cell differentiation. We suggest that the different assembly motifs with respect to compatible (AND) and mutually exclusive (XOR) surface interactions should be classified using 3D structural information to analyze which type of assembly motifs dominate during keratinocyte differentiation.
Proteins that bind mutually exclusively to the same domain on a shared binding partner protein prevent each other’s binding through steric hindrance depending on concentration and localization [10, 13, 14]. Steric hindrance and competition could be a mechanism to achieve cell type-specific functions if a competing protein is expressed at a higher level in a specific cell type or tissue [14]. To determine if the replacement of subunits of complexes during the differentiation process happens at mutually exclusive surface interactions, we structurally analyzed all CORUM complexes with less than 20 members using the SAPIN software framework (S17 Fig) [19]. SAPIN identifies the protein regions that could be involved in an interaction, provides template structures, and then performs structural superimpositions to identify compatible and mutually exclusive interactions. If a protein has at least two interacting partners, the domains mediating the interaction are superimposed on the reference domain, and the interacting domains are analyzed for compatibility (AND) or mutual exclusiveness (XOR) (S4 Table). Next, we combined expression classification during keratinocyte differentiation (non-dynamic vs. dynamic) with compatible and mutually exclusive interaction types (S18 Fig). Out of six possible cases (obviating the unresolved group), three cases were selected that we could interpret in terms of competition (Fig 6A). In case 1, the hub protein (common protein with at least two different binding partners) is non-dynamically expressed while the attachment proteins are dynamic. In case 2, all three interacting proteins are dynamic. In case 3 all three interacting proteins are non-dynamic. Interestingly, case 1 is significantly enriched for gene products with mutually exclusive surface interactions (XOR) (61% compared to 39%, Fisher’s exact test; P = 1.9e-6), reinforcing our hypothesis that dynamic genes tend to be involved in competing interactions for a common (constitutively non-dynamically expressed) binding partner (Fig 6 and S19–S21 Figs). Case 3 is significantly enriched for AND (61% compared to 39%, Fisher’s exact test; P = 2.03e-12), representing stable complexes that do not change their composition during the differentiation process. Case 2 (all three proteins dynamically changing) is also enriched for XOR interactions compared to AND (65% compared to 35%, Fisher’s exact test; P = 8.66e-04). When focusing on the 32 XOR nodes with at least two super-dynamic genes, 26 XOR nodes contain genes from at least 2 different clusters suggesting that opposing or at least different expression profiles may impose different functional outputs possibly to compete at single interface hubs. Thus, variation in the concentration of proteins (belonging to different clusters) that bind to XOR nodes may cause complex re-assembly through competition and thus achieve a different functional output in the differentiated keratinocytes or during the process of differentiation.
Using calcium-induced differentiation of primary keratinocytes as a model system for a substantial cellular re-organization, we analyzed transcriptome changes during 45 hours. We discovered a large proportion of genes change their expression during the differentiation process (36% dynamic and 7.9% super-dynamic genes), which is comparable with other cell differentiation or reprogramming studies [2,4, 65]. Our set of dynamically up-regulated genes contains 39 out of 53 previously described keratinocyte differentiation markers, 25 of which are super-dynamic [66]. The super-dynamically changing genes, aside from those with known keratinocyte function, span metabolism [67], signaling, and housekeeping cellular functions. Interestingly a quarter of super-dynamic genes are of unknown functions based on UniProt annotations and manual literature searches. The super-dynamic genes were partitioned into eight clusters, with only those genes needed for establishing the final physiological functions of the cornified envelope (e.g. water repellence, structural stability, and mechanical resistance) exhibiting a clear GO-enrichment. Many metabolic, signaling and housekeeping genes were found in all clusters. Interestingly, the shapes of the clusters are similar to those observed in other cell differentiation or reprogramming studies [4,65]. Yet, due to an absence of strong enrichment in GO terms in this study and others, direct comparison between the functions of the clusters could not be conducted. Both correlated and anti-correlated paralogous pairs have been discovered before, during reprogramming of somatic cells to pluripotency [4]. Similarly, we found that dynamic paralogs are enriched in correlated expression changes, which may represent examples of gene duplicates being maintained to increase gene dosage/expression levels [31–33]. In addition, gene duplication has also been shown to contribute to the robustness of complex formation [68]. We also found that dynamically changing paralogs can be anti-correlated. Here we showed that anti-correlated paralogous pairs have a greater amino acid sequence and Pfam domain divergence and are evolutionary older genes, suggesting that paralogs have attained specific roles during differentiation.
To analyze how dynamic and non-dynamic genes are integrated into complexes, we used the CORUM database to map the changes in gene expression onto protein complexes. Earlier work analyzing complex assemblies during the yeast cell cycle revealed that complexes consist of both constitutively non-dynamic and dynamically expressed subunits (‘just-in-time assembly’) [6]. Our study of a mammalian differentiation process augments these efforts as we find a larger fraction of di-chromatic complexes, containing both dynamic and non-dynamic genes. When we applied these concepts to a manually annotated large signaling complex to include directionality, the TNF-α/NF-κB signaling, we found a stable core associated to both a dynamically changing module and several stable modules. We propose this as a ‘constant signalosome ready to work’, where a stable core—involving TNF receptor/TRAF2—associates with a dynamic periphery with different functions: i.e. NF-κB signaling, TRAF1 resulting in anti-apoptotic effects without affecting TRAF2/TRAD. On the level of EGFR signaling, a picture emerges whereby several sub-complexes consisting of adaptors, Raf kinases, ERK kinases, etc. show a mix of both, dynamic and non-dynamic subunits. In contrast to what is proposed in the hour-glass model of signaling [69,70], our analysis suggests that several sub-complexes act in parallel, containing both common non-dynamic and different dynamic subunits.
An alternative view of complexes (‘machines vs. ensembles’) has been proposed by Loew and colleagues, which holds that in addition to stable signaling complexes and molecular machines, such as the ribosome or the proteasome, a considerable combinatorial complexity arises from different compositions of complexes [8,71,72]. We propose that these two concepts do not exclude each other and are probably both used in signal transduction. However, it is currently unclear how ‘ensemble-like’ signaling complexes actually are. Both, computational and experimental work is critical to answer this question and in particular structural analyses help to identify the two cases by distinguishing compatible (more machine-like) from mutually exclusive (more ensemble-like) complexes [10,71].
Structural information about protein domains was used to distinguish compatible and mutually exclusive protein-protein interactions among hub proteins and their binding partners [10,19]. Here, we used the SAPIN webserver to distinguish compatible from mutually exclusive binding surfaces [19]. Other successful methods for structural characterization are based on evolutionary conservation of the interacting residues, e.g. PrePPI [73], Inferred Biomolecular Interaction Server—IBIS webserver [74], and Interactome3D [15]. We identified that dynamic genes binding to a common non-dynamic hub are enriched in mutually exclusive interactions (XOR), suggesting that changes in complex assemblies during differentiation could also be caused through variation in the abundance of proteins that compete for binding to XOR nodes [14]. In fact we show here that XOR nodes involving super-dynamic genes are often from different clusters, supporting this hypothesis. Thus we go one step further and propose a model where at least some protein complexes exist assembled stably at the core, and their behavior is modified by the attachment and detachment of accessory proteins at the periphery of the signaling cascade in response to various conditions. This is in agreement with the concept of achieving cell type-specific complexes through the interaction of core (cell/tissue general) and peripheral (cell type/tissue-specific) proteins [7,75]. However, more detailed computational modeling, including information on XOR and AND nodes are needed to investigate the global impact on protein competition, e.g. how changes in protein abundance propagate through a larger PPI network [71].
We have found examples where anti-correlated changes in gene expression of subunits in a complex (Ubiqutin/degradation), or in two complexes competing for the same component (EGFR recycling complexes), which reflect very nicely what is known about factors important for keratinocyte differentiation. This supports the idea of protein competition as a driver of biological processes. In fact, computational modeling of the yeast PPI network has suggested that changes in concentration spread locally and decrease exponentially within the network, as a function of the distance from the perturbed node [76]. Previous experimental and computational work investigating competition at the RAS node has confirmed these rather local effects of competition [14]. Furthermore, it is known that competing protein interactions can induce switch-like cellular behaviors, such as apoptosis versus differentiation [77], or self-renewal versus differentiation [78]. However, cellular fates also crucially depend on the crosstalk between signaling pathways initiated at different phosphorylation sites providing spatiotemporal separation and acting as molecular switches [79]. As our gene expression changes did not allow us to analyze phosphorylation events, it is difficult to speculate to what extend phosphorylation levels contribute to network and complex reorganization processes during keratinocyte differentiation. Likewise, the role of homooligomeric complexes is neglected despite being the dominating type of interaction. In the future, it will be important to integrate protein interaction networks resolved at the level of domains as well as phosphorylation events and homooligomeric complexes, in order to provide the complete picture.
Altogether our analysis highlights the importance of understanding the dynamic assembly and disassembly of complexes taking 3D structural information into consideration, rather than unraveling networks of individual proteins.
The time-course microarray data analyzed in this work was generated in our previous study [3]. However, in this section of the methods, we re-outline the details of this experiment. Total RNA in the amount of 500 ng from calcium treated primary human keratinocytes collected every 5 hours in triplicates for a total of 45 hours (i.e. 10 samples) and labelled using Agilent LowInputQuick Amp Labelling kit following manufacturer instructions. mRNA was reverse transcribed in the presence of T7-oligo-dT primer to produce cDNA, The cDNA was then in vitro transcribed with T7 RNA polymerase in the presence of Cy3-CTP to produce labeled cRNA and this labeled cRNA was in turn hybridized to Agilent Human Gene Expression 4x44K v2 microarrays (ID026652). Signals from probes were obtained by the Agilent’s Feature Extraction custom software, were corrected for background noise using the normexp method [80] available in the R package limma from Bioconductor, and then normalized between arrays to assure comparability across samples using quantile normalization [81]. Lastly, the dataset was log2-transformed. The dataset was filtered to remove and collapse all Agilent control spots and to only include probes that were present above stochastic background expression. To do this we filtered any probes with expression profiles completely below 7 which is a figure obtained by taking the median of the entire dataset and rounding it down to a single digit. This procedure yielded 21,113 probes, which were then mapped to 16,720 genes where the probe with the highest mean expression represents a gene. For some analyses such as K-means clustering the dataset was mean-centered and scaled such that the mean of each gene is zero.
To distinguish genes that are changing (i.e. transiently or continuously up- or down-regulated) from those that are constant (i.e. stably expressed) across our differentiation time-course, we used Chi-squared test where for each gene, the expected value at each time point is equivalent to the gene’s average expression across all 10. The equation is as follow:
X2=∑iN(X¯i−X¯)2N−1
where N is the number of samples or time points (i.e. 10), X¯i is the mean expression of gene x across three experimental repeats at time point i, and X¯ is the overall mean expression of gene x across all 10 time points and all repeats. Note that in this formula there is no normalization of (X¯i−X¯)2 by the experimental error computed as SE = SD/N. Instead we computed a universal error and scaled the X2 distributions of our data to the standard X2 distribution for 9 degrees of freedom.
In addition to Chi-squared test we employed an empirically derived fold change threshold of 2 and a strict threshold of 4. Each gene’s expression peak was compared to its trough to ensure the difference satisfies the appropriate threshold and the genes were classified as follow:
non-dynamic genes have to satisfy P(X2)> = 0.01 and fold change < 2
unresolved genes either have P(X2)<0.01 and fold change < = 2 or P(X2) > = 0.01 and fold change > = 2
dynamic genes P(X2)<0.01 and fold change> = 2.
super-dynamic genes are a subset of dynamic genes which have fold change> = 4
We also conducted corrections for multiple testing using Benjamini’s method. While the above analysis had classified 6,096 genes as dynamic (1,317 as super-dynamic), after Benjamini correction for multiple testing and applying a p-value cutoff of 0.01 (1%), these two numbers remained the same while the number of non-dynamic genes increased from 6,137 to 6,413 and the number of unresolved genes decreased from 4,487 to 4,211. Thus, a subset of 200 genes got shifted from the unresolved category to the non-dynamic category after Benjamini correction. Since the majority of our analyses were based on dynamic and super-dynamic genes such as clustering and mapping to complexes and these classes did not change after multiple testing correction, we kept the gene classes as originally defined without Benjamini correction.
We used R version 2.14 for the majority of the statistical analyses along with Perl and AWK for text processing. K-means clustering was performed using the R kmeans function. To optimize and select the K value or the number of cluster centers, we exhaustively performed K-means clustering with K value ranging from 2 to 500 and then for each set of clusters, computed an F-score as well as an average silhouette score available from the standard statistics and cluster packages, respectively. The K value, which resulted in the highest average silhouette score and F-score, was selected as the optimal number of cluster centers.
A reference set of complexes was obtained from the CORUM database [36] of curated mammalian protein complexes. Out of a total of 1,331 human complexes available in the latest release (February 17th, 2012), 44 homo-oligomer complexes consisting of only one type of protein were filtered out. 84 complexes were removed, which were fully non-expressed across our skin differentiation expression time course. 154 (<13%) complexes contain subunits that are partially non-expressed in keratinocytes while 1,049 complexes contain subunits, which are all expressed during skin differentiation. The resulting set of 1,203 complexes consists of 2 or more distinct proteins per complex.
Paralog annotations were obtained from EnsemblCompara through the BioMart portal (database version EnsemblGenes71). The method identifies true paralogs by computing a phylogenetic tree across the whole set of protein-coding genes with one pipeline which includes TreeBeST [33].
Analysis of mutually exclusive (XOR) and compatible (AND) binding was done using the SAPIN web framework [19]. SAPIN identifies the protein parts that could be involved in the interaction and provides template structures and then performs structural superimpositions to identify compatible and mutually exclusive interactions. We analyzed all complexes in the CORUM database with a complex size of less than 20 members. Each CORUM complex was broken down into combinations of three different proteins, and the respective protein sequences were used as the input for the SAPIN webserver. The workflow in SAPIN is:
Assigning all domains using Pfam (pfam_scan.pl with default parameter) based on the protein sequences.
Searching the 3DID [82] database to find potential domain-domain interaction hits of binary interaction partners provided.
Finding the best template of PPI structures (containing interacting domain A and domain B) by using the InterPreTS scoring function [83]. The interaction structures are evaluated by InterPreTS based on interface sequences aligned with 3Dstructures (we used MUSCLE [84] with default parameter). SAPIN selects one template that has more than 2.33 Z-score or the best score one among all candidates.
If two PPI structures share a same domain, structural superimpositions are performed based on the domain structure using Combinatorial Extension (CE [85]).
Analyzing for backbone clashes using FoldX [86] with superposed structures, and based on a backbone clashes threshold (15% of interface residues) the interactions are assigned compatible (AND) or mutually exclusive (XOR).
The reliability of the predictions with respect to sequence similarity to the template complex, we measured the sequence identity between the reference and homolog domains from the alignment based on the Needleman-Wunsch Algorithm [87]. Based on this, we determined a Z-score for the percentage of van der Waals backbone clashing among interface residues calculated for either AND or XOR, dependent on sequence similarity [19].
If a protein has at least two interacting partners, the domains mediating the interaction are superimposed on the reference domain, and the interacting domains are analyzed for compatibility (AND) or mutual exclusiveness (XOR). Note, that even two protein having similar domains do not necessarily bind at the same site/interface, but there are frequent cases of similar domains binding in a different way. As the best template suggestion based on sequence homology using INTERPRETS [9]. As the sequence is scored in INTERPRETS against a set of 1000 random sequences, the selection of the best template is not deterministic and can result in different results in different runs. We provide the average of the XOR/AND likelihood of independent runs in S4 Table.
For calcium-induced differentiation experiments, keratinocytes were seeded into 35 mm plates and grown in Keratinocyte Serum-Free Medium with supplements (KSFM; GIBCO) [3]. After reaching 70% confluence Keratinocyte Serum-Free Medium was exchanged for EMEM (Lonza) supplemented with 8% chelated FBS, EGF (10 ng/ml), 1% penicillin/streptomycin and 0.05 mM CaCl2. After 12 hour time point 0 h was collected, and the residual keratinocytes were synchronized for 2 h with EMEM containing 20% chelated FBS, EGF (10 ng/ml), 1% penicillin/streptomycin, and 0.05 mM CaCl2. After synchronization cells were washed once with PBS and cultured in EMEM, supplemented with 8% chelated FBS, EGF (10 ng/ml), 1% penicillin/streptomycin, and either 0.05 mM or 1.2 mM CaCl2, corresponding to non-calcium and calcium treatment, respectively. Cells were collected every 5 hours during a period of 45 hr.
Keratinocytes were seeded onto 16-chambered LabTek slides (Nuncbrand). Kerationocyte differentiation was induced through addition of Calcium (as before), and sample were taken at the respective time points. Cells were fixed in 4% paraformaldehyde for 20 min and blocked with 4% BSA in PBS 1× (blocking solution). Primary staining was done using an antibody against Involucrin (Abcam, ab28057; dilution 1:1000 in blocking solution) followed by an Alexa Fluor 568 secondary antibody (Invitrogen Probes; dilution 1:200 in blocking solution). Cells were counterstained with DAPI (Sigma-Aldrich; concentration 1 μg/mL) and mounted using Mowiol solution. Staining was visualized on a Leica TCS SP5 confocal microscope with a 40× 1.25 NA objective at a zoom factor of 3 (1024 × 1024 pixels; 0.126 μm/pixel).
|
10.1371/journal.pcbi.1000212 | A Model of Brain Circulation and Metabolism: NIRS Signal Changes
during Physiological Challenges | We construct a model of brain circulation and energy metabolism. The model is
designed to explain experimental data and predict the response of the
circulation and metabolism to a variety of stimuli, in particular, changes in
arterial blood pressure, CO2 levels, O2 levels, and
functional activation. Significant model outputs are predictions about blood
flow, metabolic rate, and quantities measurable noninvasively using
near-infrared spectroscopy (NIRS), including cerebral blood volume and
oxygenation and the redox state of the CuA centre in cytochrome
c oxidase. These quantities are now frequently measured in
clinical settings; however the relationship between the measurements and the
underlying physiological events is in general complex. We anticipate that the
model will play an important role in helping to understand the NIRS signals, in
particular, the cytochrome signal, which has been hard to interpret. A range of
model simulations are presented, and model outputs are compared to published
data obtained from both in vivo and in vitro
settings. The comparisons are encouraging, showing that the model is able to
reproduce observed behaviour in response to various stimuli.
| Monitoring the brain noninvasively is key to solving various biological and
clinical problems. Near-infrared spectroscopy (NIRS) is a technique that can
measure changes in the colour of the brain. The brain has an absolute
requirement for oxygen; the spectroscopically observed colour changes are due to
the proteins that deliver (haemoglobin) and consume (mitochondrial cytochrome
c oxidase) oxygen. Haemoglobin changes colour when it binds
oxygen. The changes in cytochrome c oxidase are due to the
electron occupancy (reduction) of a particular copper metal centre in the
enzyme. The way that the state of this enzyme changes in various situations is
poorly understood. Currently there is no theoretical model that can be used to
decode simultaneously all of the spectroscopic changes in these proteins, and
thus limited information about the underlying biochemistry and physiology can be
extracted from the NIRS signals. We therefore constructed such a model, ensuring
that it is consistent with the scientific literature, in vivo
data, and the underlying thermodynamic principles. The model was able to predict
the physiological and spectroscopic responses to a wide range of stimuli,
including changes in brain activity and oxygen delivery. It is likely to be of
significant value to a wide range of clinical and life science users.
| In recent years there has been widespread use of near infrared spectroscopy (NIRS) to
monitor brain oxygenation, haemodynamics and metabolism [1],[2]. Initially the primary
chromophores of interest were oxy- and deoxy-haemoglobin, with changes (termed
ΔHbO2 and ΔHHb, respectively) being measured using differential
spectroscopy systems [3]–[5]. Technical
developments made possible the measurement of absolute tissue oxygen saturation
(TOS). This quantity has been variously labelled rSO2 (regional saturation of
oxygen, Somanetics INVOS systems), TOI (tissue oxygenation index, Hamamatsu NIRO
systems) and StO2 (tissue oxygen saturation, Hutchinson InSpectra
systems). TOS provides a percentage measure of mean oxygen saturation across all
vascular compartments in the tissue of interest. TOS has been used extensively as a
marker of tissue oxygenation in a range of applications [6]–[9] but its
relationship to underlying physiology is still under investigation [10],[11].
In addition to the haemoglobin chromophores, the CuA centre in cytochrome
c oxidase (CCO) is a significant NIR absorber. Measurement of
the changes in oxidation level of this centre give rise to a signal, here referred
to as the ΔoxCCO signal, which has been extensively investigated as a marker
of cellular oxygen metabolism [12]–[15]. A number of clinical
studies have been performed to elucidate its role as a measure of cerebral well
being [16]–[18].
Although in the case of TOS and ΔoxCCO there are no obvious “gold
standard” measurements against which a direct experimental validation can
be performed, these NIRS signals undoubtedly encode information of biological and,
potentially, clinical importance on tissue oxygen levels, blood flow, metabolic rate
(CMRO2), and other underlying state variables in the brain. However
the mapping between NIRS signals and the underlying variables is not
straightforward, as a number of different causes may give rise to the same signal
changes. The data on CCO redox state is particularly difficult to interpret because
of the potential complexity of the correlations between physiological changes and
mitochondrial redox states [12],[19].
Thus in order to correctly interpret and maximise the clinical usefulness of the
information that can be extracted from NIRS data, a model of the underlying
physiology is required. This is our aim in this paper. The model we construct is
based on thermodynamic principles, and is to date the only model which attempts to
predict the state of the CuA centre in cytochrome c
oxidase in an in vivo setting. It is designed to be able to
simulate responses to physiologically and clinically important stimuli (listed
below), and is able to reproduce several experimental data sets including both
in vivo data, for example on NIRS signal changes during
functional activation [5], and in vitro data on
mitochondrial flux and redox state during hypoxia and uncoupling [20].
Moreover our simulations suggest important practical conclusions: For example, that
the ΔoxCCO signal contains information independent of that contained in the
other NIRS signals, and that physiological variability between individuals has the
potential to affect its relationship with the haemoglobin signals.
The model is designed to respond to four input stimuli, which have been chosen both
because they are physiologically important, and because there is considerable data
on the response of NIRS signals to these inputs. The stimuli can be expected to
cause changes in the different NIRS signals via a variety of different physiological
pathways. They are
One key consideration has been to keep the model small enough to allow eventual
optimising of key parameters to an individual's data. This would be
required if the signals were to be used to interpret physiological changes in an
individual, for example in the clinical setting. For this reason rather than
attempting to append a model of mitochondrial metabolism to the large and complex
BRAINCIRC model [25], we have used this model as the basis for a much
simpler model.
In order to increase readability, model differential equations, and tables of model
variables and parameters are presented in Text S1. The model was written and simulated in
the open source BRAINCIRC interface [26] and is available for download [27],
complete with instructions on how to reproduce each of the simulation plots
presented in this paper.
The model consists essentially of two components. The first is a submodel of the
cerebral circulation, which is known to respond in complicated ways to a variety
of stimuli – physical, chemical and neuronal [28]. Though much of
the physiology is still under investigation, there are a variety of more or less
simplified models which attempt to capture some features of this control. Among
these are the models of Ursino and co-workers ([29],[30] for
example), the model of Aubert and Costalat [31], and the BRAINCIRC
model [25] described in [32] and still under
active development. All of these models have contributed to the construction of
the model described in this paper.
The second component of the model presented here is a submodel of mitochondrial
metabolism. Several such models exist, notably the models of Korzeniewski and
co-workers (e.g., [33]) and Beard and coworkers [34],[35]. These models have also played a large part
in the construction of our model, and processes here are often either
caricatures or refinements of processes in these models. The two components of
the model are linked via oxygen transport and consumption.
The basic structure of the model is illustrated in Figure 1. In order to aid model validation, a
smaller mitochondrial model appropriate to in vitro situations
will also be introduced later. In particular this model omits all processes
relating to blood flow, with oxygen being supplied directly to the mitochondria.
Following the normal simplification in most chemical models, all chemical
reactions are assumed to take place in solution in compartments. A reference
brain volume is assumed (although never needed explicitly) and other volumes are
calculated as fractions of this reference volume. Thus “blood
volume” and “mitochondrial volume” will refer to
blood/mitochondrial volume per unit brain volume. Processes
take place at two sites: in a blood compartment, divided into an arterial
compartment with variable volume, a capillary compartment with negligible
volume, and a venous compartment with fixed volume; and a mitochondrial
compartment with fractional volume Volmit which can be interpreted as
ml mitochondrial volume per ml tissue. The arterial volume Volart and
venous volume Volven are expressed as fractions of
normal total blood volume, so that in normal conditions,
Volart+Volven = 1.
In other words they measure ml arterial/venous blood per ml normal blood
volume.
Following [33], the presence of buffers in the mitochondria
serves effectively to enlarge mitochondrial volume as seen by protons. We define
an effective mitochondrial volume for protons
VolHi = RHiVolmit whereCbuffi and dpH are constants.
As discussed above, all volumes are taken as fractions of a reference volume and
are thus, strictly speaking, dimensionless. When the reference volume is not
clear the complete units will be presented. In general, chemical concentrations
are millimolar (mM), with the reference volume being implicit (so for example
concentration of a substance Y in mitochondria has units millimoles Y per litre
of mitochondrial internal volume). The exceptions are when a unit conversion is
carried out to follow convention or to facilitate comparison with data, as in
the case of NIRS quantities which are generally in μM
and where the reference volume is brain volume even when the quantity is
confined to some specific compartment. All blood pressures and partial pressures
of gases are in mmHg. For readability, units will be generally omitted from the
text but are presented in the Sections B and C of Text
S1.
The first component of the model is a basic representation of the mechanics of
cerebral blood flow. This part of the model is a simplification of the detailed
biophysics in [29], where regulation occurs at two
sites—a proximal and a distal arterial compartment, each responding to
stimuli differently. We constructed a version of this model with a single
compartment which was able to reproduce steady state responses to stimuli
adequately, and so in our model here, a single compartment is used. Certain
processes are omitted, including the viscous response of blood vessels, and the
complexities of the venous circulation. The conductance of the circulation,
G, determines cerebral blood flow CBF according to the
ohmic equationPa and Pv
are arterial and venous blood pressure respectively, which are parameters
external to the model. Cerebral blood flow (CBF) in the model means the volume
of blood which flows through a unit volume of tissue in unit time.
G is taken to be a function of a typical
“radius” r of the resistance vessels
according to the Poiseuille law:KG is simply a constant of
proportionality. r is determined by the balance of forces(1)where Te and
Tm are, respectively, the elastic and muscular
forces developed in the vessel wall, both functions of the radius, and
Pic is extravascular pressure (assumed to be
constant).
(Pa+Pv)/2 is an
average intravascular pressure. Following [29] the elastic tension
is given an exponential dependence on radius:(2)Here σe0,
Kσ, r0
and σcoll are parameters, while
h is the vessel wall thickness, set by conservation of wall
volume according to the equation:(3)h0 represents wall thickness when
vessel radius is r0.
The muscular tension is given by(4)Tm has a bell-shaped dependence on
radius, taking value Tmax at some optimum radius
rm. rt and
nm are parameters determining the shape of
the curve. Maximum muscular tension Tmax is a
crucial quantity, and is affected by all stimuli which cause changes in vascular
smooth muscle tension. To this end it is useful to define a dimensionless
quantity μ which represents the level of regulatory
input, givingTmax0 is a constant and
kaut is a control parameter, normally set to
1, but which can be lowered to simulate loss of a vessels ability to respond
actively to stimuli. μ varies between a minimum value
of μmin and a maximum value of
μmax. The level of regulatory input
depends on the level of stimuli capable of producing a response in vascular
smooth muscle. These stimuli are combined into a dimensionless quantity
η which determines μ via
a sigmoidal function:A single compartment with these functional responses was found in
preliminary simulations to be able to reproduce experimentally observed steady
state responses well. Further details are presented in the results and in Text S1.
In the model, four quantities are capable of producing direct or indirect
responses in vascular smooth muscle and hence affecting
η: arterial blood pressure, oxygen levels (taken for
simplicity to be mitochondrial oxygen levels), arterial CO2 pressure
PaCO2, and demand, which we represent as a dimensionless parameter
u. In its action within mitochondria, u may be
identified with the ADP/ATP ratio, while in its effect on blood flow it can be
seen as the level of the substrates connected with neurovascular coupling.
u is introduced in order primarily to simulate, via a
single parameter, the events occurring during functional activation. In order to
construct η, we define four quantities and vu. These are essentially
Pa, [O2], PaCO2
and u, respectively, passed through first order filters, in
order to represent possibly different time constants associated with each of
these stimuli:(5)The time constants τx control
how long it takes for each stimulus to have a vasoactive effect. Given that
blood flow regulation in response to a single stimulus often involves multiple
processes occurring on different time scales (for example direct and metabolic
effects of hypoxia), use of a single time constant for each stimulus is
necessarily an approximation. η is chosen to be linear
in all stimuli:The parameters RP,
RO, RC and
Ru represent the sensitivities to changes in
the different stimuli while vx,n represents the
normal value of vx, so that at normal values of all
stimuli η = 0, and
hence
μ = (μmin+μmax)/2.
Collapsing the complexity of the biology into a single quantity
η will necessarily have some pitfalls. However for
our purposes here, the simple form of η is
sufficient.
Knowledge of oxygen levels in blood is necessary both in order to interpret
haemoglobin related NIRS signals, and also in order to calculate oxygen
transport to tissue. It is conceptually simplifying to consider oxygen binding
sites on haemoglobin as the chemical of interest, with concentration four times
the concentration of haemoglobin. Thus oxyhaemoglobin concentration will refer
to the concentration of filled oxygen binding sites on haemoglobin.
Arterial oxyhaemoglobin concentration [HbO2,a] is
calculated from arterial saturation SaO2 and total haemoglobin concentration in
arterial and venous blood [Hbtot] (assumed constant) via
[HbO2,a] = SaO2[Hbtot].
A quantity JO2 can be defined as the rate of oxygen
flux from blood to tissue (in micromoles O2 per ml tissue per
second). A key requirement is that total O2 supplied to the tissue is
matched by oxygen delivery. This requirement is encoded in an equation(6)[HbO2,a] and
[HbO2,v] are the arterial and venous concentrations
of oxygenated Hb respectively. From the venous oxyhaemoglobin level we can
calculate a venous saturation
SvO2 = [HbO2,v]/[Hbtot].
The concentration of oxyhaemoglobin will clearly vary along the capillary bed.
Defining a typical capillary oxygen saturation
ScO2 = (SaO2+SvO2)/2 we can use this
to calculate a typical capillary oxygen concentration(7)φ is the concentration of dissolved
oxygen giving half maximal saturation, while nh is
the Hill exponent of the dissociation curve. Clearly choosing this form for
dissolved oxygen ignores possible complications arising from the Bohr effect
(see [36] for example).
By choosing a simplified form for the level of capillary oxygen, we run the risk
of miscalculating oxygen delivery. An example of a more complete treatment using
a distributed model can be found in [37]. In order to
investigate the possible errors introduced by this simplification a distributed
model was solved numerically and the true average capillary oxygen concentration
compared to that calculated from Equation 7. The results are presented in
Section D of Text S1. The approximation causes consistent overestimation of
capillary oxygen concentration introducing an error of approximately 2.5 percent
in normal circumstances. During severe ischaemia this error can grow to 6
percent. In order to minimise model complexity, we accept this level of error in
the current model.
The process by which oxygen is supplied to the mitochondria is assumed to be
diffusive occurring at a rate(8)where [O2] is the mitochondrial
oxygen concentration, and DO2 is the diffusion
coefficient. In order to ensure that arterial oxygen supply can never exceed
tissue oxygen delivery (and thus avoid venous oxygenation becoming negative) we
do not allow the value of supply to exceed
CBF[HbO2,a], i.e. we set
JO2 = min{DO2([O2,c]−[O2]),CBF[HbO2,a]}.
More details on this crude methodology for modelling a process which properly
requires PDE modelling are given in Appendix C of [32]. For in
vivo simulations where oxygen saturation may decrease
significantly, in order to avoid non-smooth behaviour we use the smooth
approximation to the function , choosing ε in this case to be
CBFn[HbO2,a,n]/10.
Equations 6–8 collectively serve to determine the values of
[HbO2,v],
[HbO2,c], [O2,c] and
JO2 and need to be solved simultaneously.
A key variable measurable using NIRS is tissue oxygen saturation (TOS), the
average saturation level of blood in the brain for which an absolute value can
be obtained. This can be expressed as a value between 0 and 1 or as a
percentage, and in the equations below we choose the former. In addition,
changes in tissue oxy-, deoxy-, and total haemoglobin concentration (as
distinguished from blood concentrations), termed ΔHbt, ΔHbO2 and
ΔHHb respectively and measured in μmol(l
tissue)−1 can be calculated.
In order to calculate TOS, we need only the relative volumes Volart
and Volven (and no value for the fractional volume of blood per unit
brain volume). Ignoring the capillaries, which are assumed to have small volume,
we getNext we assume that Volart is proportional to
r2 so that where Volart,n and rn
are the normal values of Volart and r. Dividing the
expression for TOS through by the normal arterial volume, Volart,n,
and defining normal arterio-venous volume ratio
AVRn = Volart,n/Volven,
then gives:In order to define the other NIRS quantities we require some
estimate of absolute blood volume in the tissue. So we define a parameter
Volblood,n in (ml blood)(ml tissue)−1, and
get the tissue concentrations of total, oxy- and deoxy-haemoglobin in
μmol(l tissue)−1 as, respectively:The factor of 1000 arises from conversion from mM to
μM, while division by 4 occurs because of our
definition of Hb as binding sites on haemoglobin.
Multiplication by Volblood,n is to convert to tissue concentrations.
NIRS signals ΔHbt, ΔHbO2 and ΔHHb are thenwhere Hbtn, HbO2n and HHbn are
normal values of Hbt, HbO2 and HHb.
The second key component of the model is a basic submodel of mitochondrial
dynamics centred in particular on the oxidation state of the CuA
centre in cytochrome c oxidase. The inspiration for this model
comes from the detailed models of [33] and [34], and
the abstract model in [38]. However, in order to minimise model size,
many of the processes in [33] and/or [34] have been omitted:
in particular phosphate and ADP/ATP transport, and the adenylate kinase and
creatine kinase reactions. Further, the behaviour of complexes I-III has been
lumped into a single process. On the other hand somewhat more detail has been
included in the treatment of complex IV (cytochrome c oxidase)
with a view to more accurate information on the redox state of the
CuA centre. It is worth mentioning that the simplifying assumption of
a single site of oxidative metabolism ignores the diverse roles of neurons and
astrocytes in brain energy metabolism.
Two redox centres in cytochrome c oxidase are identified
explicitly, CuA, and the terminal electron acceptor cytochrome
a3 (henceforth termed cyta3). Each of these centres
can exist in either an oxidised or a reduced form. A reducing substrate
transfers electrons (directly or indirectly) to CuA, which in turn
transfers its electrons to cyta3. Finally cyta3 transfers
its electrons to oxygen. These three electron transfers, which we will refer to
as reaction 1, reaction 2 and reaction 3, occur at rates
f1, f2 and
f3. These rates are taken to be the rates of
transfer of four electrons between substrates. They are accompanied
by the pumping of protons across the mitochondrial membrane, and hence both
create and are affected by the proton motive force Δp (also termed PMF,
discussed below). The structure of this submodel is shown in Figure 2.
From here on, we represent the concentration of oxidised and reduced
CuA by CuAo and CuAr respectively. Similarly
oxidised and reduced cyta3 are represented by a3o and
a3r respectively. The total concentrations of CuA and
cyta3 in mitochondria are assumed constant at some value
cytoxtot.
The proton motive force Δp has both a chemical and an electrical
component and has the formHere ΔΨ is the mitochondrial inner membrane
potential, pHo is pH in the intermembrane space assumed to be a
constant or controllable parameter.
Z = RT/F
where F is the Faraday constant, R is the
ideal gas constant, and T is the absolute temperature. The
dynamics of ΔΨ are discussed below.
Protons move across the mitochondrial membrane in both directions. A quantity
p1 of protons are pumped out during the
reduction of four CuA centres, and p2 are
pumped out during their oxidation, and p3 are pumped
during the final oxidation of cyta3.
ptot = p1+p2+p3
is thus the total number of protons pumped out of the mitochondria during the
reduction of one molecule of O2. The value of
p1, and hence ptot,
will depend on the reducing substrate.
The protons pumped out of mitochondria during electron transfer return into the
mitochondria via leak channels at rate Llk, and via
processes associated with ATP production (i.e. through Complex
V, and during ADP/ATP and phosphate translocation) at a rate
LCV. Thus the total return of protons into
the mitochondria occurs at rate
L = LCV+Llk.
Following [33] the leak rate is exponentially dependent on Δp:Llk0 and
klk2 are parameters controlling the
sensitivity of the leak current to changes in Δp.
kunc is a control parameter, normally set to 1,
used to simulate the effect of adding uncouplers to the system. It is only
altered during simulations of the simplified model of isolated mitochondria
described below.
LCV depends on both Δp and the demand
u. The formis chosen. If we identify the demand parameter u
with an (appropriately rescaled) ADP/ATP ratio, we see that this form is similar
to that for the rate of complex V in [33]. It is also
qualitatively similar to the form in [39] despite the
apparent complexity of the form in that reference. The parameter Δp
CV
0 is the value of Δp at which, given normal demand,
LCV goes to zero.
kCV controls the sensitivity of the rate to changes
in Δp. rCV controls the relative sizes of
maximal and minimal rates of LCV. If
nA protons enter the matrix for every molecule
of ADP phosphorylated, the actual rate of ADP phosphorylation is
LCV/nA. The
current consensus value of nA is given as 4.33 in
[40]. Note that because of differences in the
constructions of the two models, the parameter nA
has a somewhat different meaning to its counterpart in [33].
Following the methodology in [35],[41], the rate of change
of ΔΨ depends only on the flows of protons across the membrane
and is given byCim is the capacitance of the
mitochondrial inner membrane.
We now return to reactions 1, 2 and 3 with rates f1,
f2 and f3. For
simplicity each of these rates refers to the transfer of four electrons. The
processes associated with rates f1 and
f2 are assumed to be reversible. Assuming first
order kinetics for f1 giveswhere k1 and k−1 are the forward
and backward rate constants for the reaction. Although the details of how the
rate constants change with changes in Δp are not known in advance, the
equilibrium for the reaction can be set from energetic principles: Associated
with f1 we have a free energyThe important quantity E1 is
discussed further in Section C of Text S1. Setting
ΔG1 = 0
determines the equilibrium constant of the reaction Keq1, givingTo allow for inhibition by changes in the proton motive force,
k1 is set aswhere k1,0 is the value of k1 at normal
Δp. Since demand or experimental set-up may influence the redox state of
the initial reducing substrate k1,0 is not a constant (details in
Section C of Text S1). The exponential term reflects inhibition of the forward
rate by Δp, and the strength of this inhibition is controlled by the
parameter ck1. The backward rate constant
is then determined from the equilibrium constant:
A very similar process can be used to set f2. Again,
forward and backward rate constants k2 and k−2
are assumed, givingThis time the free energy isgiving the equilibrium constant Keq2k2 is then set asck2 controls the effect of
changes in Δp on k2. The backward rate constant is simplyReaction 3 is assumed to be irreversible, and its rate
f3 is set as(9)
The quantities c3 and Δp30 are parameters
controlling the sensitivity of f3 to Δp.
From the above form it is possible to calculate an apparent second-order rate
constant for the reaction taking place at zero PMF as(10)Values of this parameter can be experimentally measured [42] and
the measured values are used to determine the value of k3 in the
model.
As f3 is the rate of oxygen consumption it is used to
calculate the crucial model output:(11)In order to simplify the model we have assumed that control of
cytochrome c oxidase is via Δp alone, ignoring the fact
that changing ΔpH and ΔΨ can have different effects on
cytochrome c oxidase turnover [43].
The NIRS ΔoxCCO signal can be identified as the change, in
μM, in the tissue concentration of oxidised
CuA. In order to model this quantity, we defineThe factor of 1000 is to convert from mM to
μM, while multiplication by
Volmit—mitochondrial volume as a fraction of tissue
volume—converts from mitochondrial to tissue concentration.
Apart from the model described above, in order to set parameters and compare
model behaviour to experimental data a simpler submodel is also constructed.
This model will be referred to as the simplified model while
the model described above will be referred to as the full
model. The simplified model is designed to simulate in
vitro experiments on mitochondrial solutions, and so omits a number of
processes in the full model. A schematic of this model is shown in Figure 3.
The key differences between the simplified mitochondrial model and the full model
are that all processes and feedback involving blood flow are removed.
Mitochondrial O2 becomes a control parameter rather than a model
output, and the reducing substrate is not automatically assumed to be NADH, but
may be chosen to be other substrates such as succinate or TMPD. The simplified
model can also model experimental data involving uncouplers: These are
molecules, generally protonophores, that uncouple oxygen consumption from
oxidative phosphorylation, allowing rapid electron transfer with no ATP
synthesis. Data from experiments such as that in [20] can then be used
for model parameter setting or model validation.
We intend our model to be able to reproduce standard, well-understood experimental
phenomena; however, we also wish to use it to gain insight into areas where the
physiology and biochemistry underlying the changes in the ΔoxCCO signal are
poorly understood, especially quantitatively. To this end we have explored the
behaviour of the model under a range of conditions.
The steady state response of cerebral blood flow to changes in blood pressure
gives rise to “autoregulation” curves with blood flow being
insensitive to changes in blood pressure around the physiological value [44]–[47]. This is
obviously key behaviour that our model must be able to reproduce. Steady state
responses of cerebral blood flow to other stimuli, in particular PaCO2, are also
well characterised experimentally [48]. The model steady
state blood flow responses to changes in blood pressure and CO2
levels are plotted in Figure
4.
The pressure autoregulation curve is consistent with experimental curves (e.g.
the autoregulation curve in [44] constructed from data in [46],[47]) and modelled
curves (e.g. using the model in [29]). Data from these studies was used to set
model parameters as described in Section E of Text S1.
The value of RC has been set so that model steady
state response to changes in PaCO2 is consistent with published data [48].
Data from a hypercapnia study described below suggests that the magnitude of
this response may vary between individuals.
Functional activation provides a repeatable challenge giving rise to discrete
changes in metabolic demand, which can be assumed to be primarily cerebral.
Since its inception in 1993 [49]–[52], the study of
functional activation by NIRS (fNIRS) has rapidly become one of the main drivers
in the development of NIR technology for monitoring the human brain. Yet there
have been few studies focusing on the ΔoxCCO signal, despite its
potential to inform on the critical question of neurovascular coupling. In 1999,
a paper reported on oxidation of ΔoxCCO during fNIRS [15].
Despite a number of attempts to dismiss this result as an optical artefact, the
basic finding has resisted such explanations [53]. However, whether
the oxidation can be explained physiologically (effect of increased oxygen
delivery) or biochemically (effect of increased ATP turnover) is not clear.
In order to shed light on such questions, functional activation was simulated in
the model, via a step up in the demand parameter u. A ten
second activation was simulated by running the model at normal parameter values
for 10 seconds, followed by a 10 second increase in u, followed
by a further ten seconds at baseline. The responses of various quantities are
plotted in Figure 5.
As expected, the increase in blood flow more than compensates for the increase in
CMRO2 so that TOS goes up. The ratio of changes in blood flow to
changes in CMRO2 is consistent with the data in [54] where a ratio of
2∶1 is typical, although higher values are reported in [55].
Also clear from the data is that at normal parameter values an increase in
demand causes oxidation of CuA, and hence an increase in the
ΔoxCCO signal consistent in direction, but smaller in magnitude (by
about 50 percent) than the typical traces in [5]. Below we show
that, perhaps surprisingly, this effect is not primarily dependent on an
increase in blood flow and blood oxygenation.
The behaviour of the other NIRS signals—ΔHbO2, ΔHHb and
ΔHbt—during functional activation is plotted in Figure 6. Changing the time
constant associated with demand (τu) affects
the shape of the response, and the magnitude of a slight initial increase in
deoxygenated haemoglobin before it starts to drop.
Both the levels and direction of change of the haemoglobin signals are comparable
with previous experimental data [24], although the magnitudes predicted are
somewhat higher than reported in [5].
Consistent with the analysis in [38], both the size and the
direction of ΔoxCCO change in response to functional
activation are sensitive to a number of model parameters including the baseline
PMF and values of the standard redox potentials. One interesting question is
whether the effect is driven solely by the increase in cerebral blood flow
associated with functional activation. A simple way to test this is by
abolishing the response of blood flow to demand by setting
Ru = 0. This
reduces the ΔoxCCO increase (by about 40 percent) but does not abolish
it (results not shown).
In this light it is interesting to run an analogous simulation involving a step
up in demand on the simplified mitochondrial model. Such a change can be
identified with a transient increase in the ADP/ATP ratio in an in
vitro situation. As in the in vivo case, there was a
small but significant oxidation of CuA. To see whether this oxidation
is a robust response to activation, the level of activation was varied so that
CMRO2 varied between 80 percent and 170 percent of baseline. The
results of both simulations are plotted in Figure 7.
As is clear from Figure 7,
increased demand oxidises CuA even in the simplified model where
there is no change in oxygen level. Qualitatively similar results are obtained
when an increase in demand is replaced with uncoupling. These results suggest
the important conclusion that the change in the ΔoxCCO signal during
functional activation is primarily associated with changes in proton motive
force rather than being slaved to changes in oxygen levels. The ΔoxCCO
signal thus appears to encode information about cerebral metabolic state
independent of that contained in the other NIRS signals.
It is also interesting to note this work supports the conclusion of [55]:
That in the physiological range, an increase in CBF is not required for the
observed increase in CMRO2 to take place. In order to verify this,
the full model was run with
Ru = 0 so that
demand had no effect on blood flow. Again, significant increases in
CMRO2 – up to about 45 percent – could occur. The
relationship between oxygen levels and CMRO2 was also consistent with
data in [55] as shown in Figure 8.
Understanding the response of the ΔoxCCO signal to changes in oxygen
concentration is central to understanding much experimental data. Yet the
details of this response are controversial, even when measured during in
vitro experiments in cells and mitochondria. Partly this arises
from the technical difficulty of making measurements at low oxygen
concentrations (see [56] for a lively discussion of this from one
author). In particular, debate has centred around the Km for oxygen
consumption, which is known to be a complex function of cell metabolism [57].
Even simple models suggest that there is no need for standard Michaelis-Menten
type behaviour of consumption rate with oxygen levels [58]. Apart from the
uncertainties in the behaviour of consumption when oxygen concentration is
dropped, there are also uncertainties about how mitochondrial redox states
change in this situation. Again the quantitative response cannot be
heuristically predicted, and there is contradictory data in the literature [59],[60].
We used our simplified model to explore some of these questions. There are very
few reliable papers reporting on changes in the CuA redox state with
oxygen; therefore we focussed on a key paper that reported on cytochrome c redox
state changes [20], which we have shown is likely to be in close
redox equilibrium with CuA during enzyme turnover [61].
Here we show that our model is capable of reproducing quantitatively key results
from [20]. In Figure 9 the behaviour of redox state of cytochrome c and the
equivalent data for CuA in the model are presented. There is good
agreement between the experimental and modelled data. The figure caption gives
details of the simulation.
The apparent Km for oxygen of mitochondrial oxygen consumption is
quoted as 0.8 μM in [33], consistent
with values in [20]. The behaviour of CMRO2 as
[O2] is lowered in the simplified model is
illustrated in Figure 10.
Details of the simulations are presented in the figure legend.
For the coupled mitochondria, half-maximal CMRO2 occurs at a little
less than 1 μM O2. For the uncoupled
mitochondria half-maximal CMRO2 occurs below 0.1
μM O2. (In order to calculate the
Vmax—and hence Km—values in the case of the
coupled mitochondria, larger values of oxygen than shown were needed. As with
the model in [58], the graph does not fit a simple
Michaelis-Menten curve well. In the uncoupled case the graph was blown-up for
very low oxygen values in order to determine the Km value.) The model
values are consistent with the results in [20]. It should be noted
that the low value of u (high phosphorylation potential) used
in these simulations was essential to get the marked lowering of apparent
Km during uncoupling. Without this choice, the Km for
coupled mitochondria is also very low, suggesting that experimental results of
this kind might be sensitive to experimental details such as the levels of ADP
supplied.
In [58] we showed that the lowering of the Km
for oxygen during uncoupling can be achieved assuming that the effect of
uncoupling is to inhibit the reverse reaction during which electrons are
transferred from cyta3 to CuA. However in the model presented in that
paper the lowering in Km was not accompanied by any increase in flux.
As shown in the graphs above our new model can simultaneously achieve an
increase in flux and a drop in the Km for oxygen.
Obtaining the qualitative behaviour shown in Figure 9, the quantitative match in Figure 10, and the qualitative
behaviour during functional activation in [5] and [24] was
achieved by varying the six model parameters which control the response of
reaction rates to Δp: i.e. Δp30, c3,
ck1,
ck2,
LCV,0,
rCV and Δp
CV
0. This is discussed further in Section C of Text
S1.
As NIRS-derived parameters report on oxygen delivery and consumption in the
brain, there is obviously wide interest in the effect of hypoxia on the NIRS
signals. Indeed hypoxia is by far the most common in vivo NIRS
challenge, especially in animal models. It is also amongst the most
controversial, with different mathematical algorithms leading to different
conclusions about the relationship between the haemoglobin-based NIR signals and
that of ΔoxCCO [22], [62]–[64]. Even with a
single algorithm [65] different physiological explanations have
been proposed for the changes during hypoxia (large decrease in oxCCO from
baseline) and immediately post-hypoxia (small increase in oxCCO from baseline).
Currently the debates in this area have revolved around the physics of making the
measurements (choice of wavelengths, effect of multiple tissue layers on light
transport, etc.) Moreover, the systems studied have not always been identical
(animal models versus humans and newborn versus mature), raising the possibility
of differences in the underlying biochemistry and physiology. Therefore an
analysis of how our model behaves during hypoxia, and how variations in the
model parameters affect the relationship between the NIR signals, is clearly
important, being independent of measurement concerns and allowing an exploration
of possible effects of physiological variation.
The dynamic and steady state responses of modelled NIRS signals to hypoxia were
explored. In the first simulation a one minute drop in arterial oxygen
saturation from 96 percent to 80 percent was carried out. The results are
plotted in Figure 11.
Following hypoxia there is an increase in blood flow leading to a partial
restoration of TOS (and to a lesser extent ΔoxCCO) during the hypoxia.
This behaviour is connected with the rapidity of the drop in arterial oxygen
saturation and so in simulations of real hypoxias (see next section) this
adaptation is unlikely to be observed. Both TOS and ΔoxCCO show an
overshoot associated with the hyperaemia following reoxygenation, consistent
with some experimental observations [65].
In [66] data on the relationship between ΔHbO2
and ΔoxCCO during hypoxia is presented. In order to test the model
behaviour in this situation, a steady state simulation (as in the production of
steady state curves above) was carried out. The results of this simulation are
plotted in Figure 12.
In [66] a very clear biphasic relationship was reported
between ΔHbO2 and ΔoxCCO. At normal parameter values, although
the model does predict increased sensitivity of ΔoxCCO to oxygen levels
at lower oxygen levels, the biphasic relationship is slight (Figure 12A). Interestingly,
lowering both demand (and hence baseline CMRO2) and normal blood flow
leads to a considerably more marked nonlinearity in the relationship (Figure 12C). This simultaneous
change in demand and normal flow leads to a normal TOS of about 60 percent
consistent with that calculated from the absolute oxy- and deoxy-haemoglobin
values in [66].
This leads to some interesting questions. In both of the simulations above,
ΔoxCCO has an approximately linear relationship with CMRO2
(Figure 12B and 12D),
and so any significant drop in ΔoxCCO implies that arterial oxygen
supply can no longer match demand – an event we can term metabolic
failure. The simulations indicate that the threshold for metabolic failure can
be more or less sharp depending on the normal matching of oxygen supply and
demand for an individual. They raise the possibility that the relationship
between ΔHbO2 and ΔoxCCO during hypoxia may depend on
differences between species, age, and possibly individual, with some individuals
being more vulnerable to hypoxia. This may have important implications for
clinical management of patients in neurocritical care.
In the future we intend to challenge our model to reproduce a wide variety of
in vivo data sets. Here we present preliminary results in
this direction. First we compared our model output to experimental data from
subjects undergoing the most common challenge used to provoke responses in the
oxCCO signal – cerebral hypoxia. The data is from a study described in
[67]. Modelled and measured TOS and ΔoxCCO
signals for a subject undergoing a hypoxic challenge are presented in Figure 13.
The stimuli were a series of drops in inspired oxygen and consequent drops in
arterial oxygen saturation. Experimentally measured inputs to the model were
SaO2, PaCO2 and mean arterial blood pressure. All inputs were down-sampled to 1
Hz. The baseline value of the ΔoxCCO signal has been brought to zero,
and in order to remove high frequency noise the data has been filtered using a
5th order low pass Butterworth filter with a cut-off frequency of 0.1 Hz (Matlab
Mathworks Inc.)
In spite of the known inter-subject and regional variability in TOS, both
baseline TOS and changes in TOS are predicted well for this subject by the
model. The model seems to slightly underestimate ΔoxCCO signal changes,
although given the level of noise in the experimental data the extent of this is
not clear.
As a test of the model's behaviour in the context of changes in arterial
CO2, NIRS data from healthy subjects monitored while undergoing
moderate hypercapnia, described in [68], was compared
with model predictions. In this study, the only NIRS signal monitored was TOS.
There was wide variation in baseline TOS between subjects, corresponding to
natural variability in blood flow and CMRO2, but more importantly to
the fact that the arterio-venous ratio in the region of tissue queried can have
high variability. In all cases the modelled and measured data were qualitatively
comparable before any attempt to optimise model parameters. However a good fit
to the data could be obtained by varying two parameters: Normal arterio-venous
ratio AVRn, and RC, the sensitivity of blood flow to
PaCO2. Despite the fact that information is often not clearly
visible in the data (see Figure
14A, for example), in all cases but one, optimisation gave positive
values for RC, in other words, the model was able to
detect a positive cerebrovascular reactivity to CO2 in the
data—a fact which is potentially of clinical importance ([69]
for example). Two examples of data-sets before and after fitting are presented
in Figure 14.
Overall, preliminary comparisons between modelled and measured in
vivo data are encouraging. A future task will be to compare further
data from these studies and other in vivo studies with model
outputs.
A basic model of the control of cerebral blood flow and the behaviour of various
NIRS signals has been presented. The model is relatively simple, containing very
few dynamic variables, but nevertheless preliminary simulations show that it is
capable of reproducing basic expected behaviours, and matching experimentally
measured data. One important conclusion from these simulations is that the
ΔoxCCO signal contains information above and beyond what is available
from the other NIRS signals. This in turn gives more hope of achieving the
ultimate aim: Real time reconstruction from NIRS data of underlying
physiological events of clinical importance.
So far, several model parameters have only been set heuristically, and comparison
with measured data has not been systematic. The immediate next stage is to
explore systematically the effects of model parameters on important model
behaviours, for example on the Km for oxygen during hypoxia and the
direction of the ΔoxCCO signal during activation. Once key outputs are
identified it will be possible to carry out a sensitivity analysis of the kind
carried out in [34]. Parallel to identifying how model behaviour
is sensitive to parameter values, is the need to identify which parameters are
liable to show variability between individuals, or between health and pathology.
Some of our observations in these directions are presented in Text S1.
Once these parameters have been identified, optimisation of the kind described
in Figure 14 can focus on
setting these parameters from an individual's data.
A number of limitations of the model have been pointed out in the text. The
limitations we consider most serious are:
By running sensitivity analyses and comparisons with experimental data it will
become clear which of these limitations affect model behaviour appreciably,
enabling us to refine the model as necessary. The process of gathering data
needed to help validate the model is ongoing. Once the model is well validated
it should be possible to integrate its use into the normal NIRS measurement
process, greatly enriching the value of the measured data.
|
10.1371/journal.pgen.1000823 | Environmental and Genetic Determinants of Colony Morphology in Yeast | Nutrient stresses trigger a variety of developmental switches in the budding yeast Saccharomyces cerevisiae. One of the least understood of such responses is the development of complex colony morphology, characterized by intricate, organized, and strain-specific patterns of colony growth and architecture. The genetic bases of this phenotype and the key environmental signals involved in its induction have heretofore remained poorly understood. By surveying multiple strain backgrounds and a large number of growth conditions, we show that limitation for fermentable carbon sources coupled with a rich nitrogen source is the primary trigger for the colony morphology response in budding yeast. Using knockout mutants and transposon-mediated mutagenesis, we demonstrate that two key signaling networks regulating this response are the filamentous growth MAP kinase cascade and the Ras-cAMP-PKA pathway. We further show synergistic epistasis between Rim15, a kinase involved in integration of nutrient signals, and other genes in these pathways. Ploidy, mating-type, and genotype-by-environment interactions also appear to play a role in the controlling colony morphology. Our study highlights the high degree of network reuse in this model eukaryote; yeast use the same core signaling pathways in multiple contexts to integrate information about environmental and physiological states and generate diverse developmental outputs.
| Baker's yeast forms smooth round colonies when grown in favorable conditions. When starved for one or more nutrients, yeast can alter its growth pattern to produce complex structures consisting of numerous interacting cells. One mode of growth, the colony morphology response, produces visually striking, lacy colony architectures. We describe both conditions that induce this morphology and also genes and pathways that are required for the response. We demonstrate that low levels of carbon combined with abundant nitrogen trigger complex colony formation. Using a candidate gene approach coupled with genome-wide mutagenesis, we identified genes involved in the production of complex colony morphology. Many of these genes are components of either a MAP kinase cascade or the Ras-cAMP-PKA pathway, two well-studied signaling pathways that are conserved across eukaryotic organisms. Yeast use these pathways to mediate cellular responses to changes in their environment. We observe shared characteristics between complex colonies and biofilms, which are organized communities of microorganisms with relevance to human health and human infrastructure, making colony morphology a candidate model for understanding how microorganisms interact to form complex structures.
| Baker's yeast, Saccharomyces cerevisiae, is most often described as a simple, unicellular organism. Despite this perception, S. cerevisiae displays a surprising array of behaviors, many of them involving complex interactions between cells. Under nutrient rich conditions, S. cerevisiae grows via “yeast form,” mitotic growth, rapidly dividing and forming smooth, round colonies on solid media. Limitation of one or more key nutrients can trigger a variety of developmental responses. For example, nitrogen starvation of diploid cells induces pseudohyphal growth, which is characterized by elongated cells, agar invasion and unipolar budding, where mother and daughter cells remain attached [1]–[3]. Haploid invasive growth, a similar behavior, is observed in haploid cells grown under dextrose limitation [4], or in the presence of various alcohols [5]–[7]. Nitrogen starvation combined with a non-fermentable carbon source induces sporulation and meiosis [8]–[11].
A number of yeast developmental responses result in multicellular structures. For example, biofilm mat formation is induced by growth on solid media with low agar and dextrose concentrations [12]. The combination of plating on hard agar followed by UV irradiation has been shown to trigger the growth of multicellular, macroscopic stalks [13]. Cell-cell adhesion is a necessary component of these responses and is induced by several different stresses including carbon and nitrogen starvation and changes in ethanol concentration and pH [14]. Recent work suggests a quorum sensing mechanism in S. cerevisiae based on the autostimulatory aromatic alcohols phenylethanol and tryptophol. This quorum sensing mechanism has been shown to enhance filamentous growth, and presumably contributes to other developmental responses as well [15].
In addition to the developmental responses described above, S. cerevisiae can form colonies consisting of complex, organized, macroscopic structures (Figure 1). We refer to the induction of this phenotype as the “colony morphology response.” The determinants and function of the colony morphology response are poorly understood in yeast. Complex colonies produce an extensive extracellular matrix that is absent from simple colonies [16], and it has been proposed that complex colonies help protect yeast cells against a hostile environment [17]. It has been observed that starvation results in the reorganization of yeast colonies at the cellular level [18], and there is evidence that budding patterns and distributions of cell shape are different in complex colonies than simple colonies [19]. Microarray expression analysis comparing a strain with a complex colony phenotype and a strain with smooth colonies, derived from the first by passaging on rich media, found numerous differences in their transcriptional profiles [16]. However, it is impossible to tell which of these changes are cause, which are effect, and which are unrelated to the colony morphology response.
The colony morphology response is a promising system for the study of simple multicellular developmental processes because it involves cell-cell communication, cellular differentiation and specialization, and cell-adhesion. While the mechanisms involved in the development of complex yeast colonies are unlikely to be evolutionarily related to the developmental pathways regulating multicellularity in metazoans, S. cerevisiae offers the opportunity to explore the principles underlying multicellular differentiation in an extremely tractable model system. As a “facultative” multicellular behavior of a unicellular organism, complex colony formation raises interesting questions of cooperative behavior and the repeated evolution of multicellularity across the tree of life [20]. Similar colony morphologies are observed in many undomesticated bacteria [21]. This gross similarity at the macroscopic scale begs the question of whether such structures represent convergent, adaptive solutions that microbial lineages have evolved to deal with similar environmental challenges.
In this report, we define key environmental and genetic determinants of complex colony morphology in S. cerevisiae. By studying the phenotypes of a genetically diverse panel of S. cerevisiae isolates under a large number of growth conditions we have determined that fermentable carbon source limitation plus an abundant nitrogen source are the key nutritional signals for inducing complex colony morphology. We show that the complex colony response requires the filamentous growth MAP kinase (FG MAPK) cascade and Ras-cAMP-PKA signaling and that mutations at the RIM15 locus exhibit synergistic epistasis with components of these pathways. We also demonstrate that ploidy and mating type quantitatively contribute to the intensity of colony morphology and that genotype-by-environment effects are common for this trait.
We studied eight strains of S. cerevisiae (BY4743, BY4739, MLY40α, MLY61a/α, YJM224, YJM311, OS17, NKY292) under a variety of growth conditions (Table S1) in order to determine the most important environmental triggers for complex colony morphology (CCM). This strain panel was chosen to include common laboratory strain backgrounds - S288c (BY4743 [diploid] and BY4749 [haploid]), SK1 (OS17 [diploid] and NKY292 [haploid]), and Σ1278b (MLY61a/α [diploid] and MLY40α [haploid]) - as well as a distillery strain (YJM224 [diploid]) and a clinical isolate (YJM311 [diploid]). Σ1278b and SK1 are standard backgrounds for studying yeast development (sporulation in SK1, filamentous growth in Σ1278b) and their inclusion here facilitates comparisons between developmental processes. We varied the conditions of growth along five major axes: carbon source type and concentration, non-carbon nutrient concentration, media water content, media hardness (agar content), and temperature. Growth was monitored daily for six days, and each plate was scored for colony morphology (Figure 2). This survey showed that induction of colony morphology is primarily carbon source dependent, with the strongest effects induced by reduced dextrose (1% dextrose w/v) and non-fermentable carbon sources (isopropanol, ethanol, acetate). Increasing dextrose concentration (4% Dextrose YEPD) inhibits the colony morphology response, providing further evidence that carbon source limitation is a primary trigger for CCM. In contrast, media water content and hardness had little if any effect on CCM induction. The only obvious effect of temperature was slow growth at lower temperature, which prolonged the time course of colony development.
We further investigated the impact of carbon availability on CCM induction by growing the same strains on YEPD plates containing a range of dextrose concentrations, from 2% (standard YEPD) to 0.0625% (Figure 3). We observed two major trends in this experiment. First, the lowest concentrations of dextrose caused the fastest induction of CCM. On lower dextrose concentrations CCM is observable as early as day two for some strains (Figure 3). Second, there is strain-to-strain variation in dextrose sensitivity. By day six most CCM competent strains exhibit the phenotype on 1% dextrose (MLY40α, OS17, NKY292, and YJM311) and even weakly on 2% dextrose (NKY292), while others (YJM224) required a dextrose concentration of 0.5% or less to induce the colony morphology response. At the lower end of the dextrose concentrations tested, colonies were smaller at each time point, presumably because they exhausted all available carbon, or the low levels of carbon induced growth regulation. At the lowest dextrose concentrations some strains failed to demonstrate the strain specific colony morphotypes observed at intermediate concentrations, likely because of growth limitations.
Other nutrients also play a role in the complex colony response. Reducing yeast extract and peptone to half of the normal YEPD levels inhibits complex morphology, and doubling these nutrients induces it (Figure 2). We suspected that nitrogen might be the key nutrient causing this effect. To test this hypothesis we assayed colony morphology on synthetic media (SC) with and without the addition of glutamate, a preferred nitrogen source [22],[23]. None of the strains tested exhibited complex morphologies on 0.5% Dextrose SC (SCLD), but when the synthetic media is supplemented with 50mM glutamate (SCLD+Glu), some strains developed complex morphologies like those observed on YEPLD, while others developed intermediate morphologies (Figure 4 and Figure S1). The most glucose sensitive of the strains in our survey (YJM224) displayed only simple morphology on the glutamate supplemented SCLD media. Higher levels of glutamate (200mM) resulted in little if any additional changes in colony morphology (data not shown).
Because there are significant pleiotropic interactions between developmental pathways in yeast [24] we hypothesized that the signaling and regulatory pathways controlling the colony morphology response would show some degree of overlap with those regulating other developmental responses, such as pseudohyphal growth, haploid invasive growth, and sporulation. To test this, we assayed colony morphology phenotypes in a panel of knockout mutants of genes known to be involved in developmental processes. This panel consisted of over 150 strains representing more than 50 different gene knockouts in MATa, MATα, and MATa/MATα strains of two lineages of the Σ1278b background. Wild-type diploid Σ1278b shows simple colony morphology in our assays while haploid Σ1278b shows strong complex morphology (see section on ploidy below). We identified thirteen haploid loss-of-CCM mutants and four diploid gain-of-CCM mutants (Table 1). We found that some gene-knockouts behaved differently in the different lineages of the Σ1278b background. For example, the tpk3Δ/ tpk3Δ diploid mutants exhibit a gain of CCM in the “Heitman” Σ1278b background [25], but not in the Sigma2000 background [26]. This variation is likely due to small genetic differences between these strains (see below) resulting from distinct histories of strain construction [27]. In some cases we observed differences in the phenotypes of gene-knockouts between MATa and MATα strains (Table 1, Figure S2 and Figure S3). In addition to the four diploid mutants listed in Table 1, we observed that a hog1Δ/hog1Δ mutant had a gain-of-CCM when grown on YEPD, YEPLD, YEPHD, and YEPEthanol (Figure S4). This pattern of induction suggests that crosstalk between various signal transduction pathways, which has been observed to cause inappropriate responses to environmental signals [28]–[30], can also induce complex colony morphology as well.
In order to gain a more comprehensive understanding of the genes and pathways affecting colony morphology phenotypes, we carried out a transposon mutagenesis screen using the mTn7-mutagenized genome library created by Kumar et al [31]. This screen identified seven additional genes exhibiting loss-of-CCM mutant phenotypes: YTA7, RSC1, RGT1, RRT12, TRM9, ELP4, and PET122. Most of these genes have been previously described as affecting developmental pathways. Both ELP4 and TRM9 are members of the tRNA modification elongator complex. Other members of the elongator complex are required for filamentous growth and elp2Δ mutants show reduced biofilm mat formation [32]. Fischer et al showed that deletion of RSC1 impairs FLO11 expression and hence leads to a loss of invasive and pseudohyphal growth [33]. YTA7 is involved in chromatin silencing and maintains a barrier between heterochromatin and euchromatin upstream of the silent HMR locus [34]. In other screens, YTA7 mutants have been found to have a loss of “fluffy” colony morphology [35] and decreased filamentous growth [36]. RGT1 encodes a glucose responsive transcription factor and mutations in this gene are known to cause sporulation defects, though this may result from decreased cell size in these mutants [37]. RRT12 (OSW3) encodes a protein involved in the formation of a protective dityrosine coat required for spore wall assembly [38].
As described above, we observed phenotypic differences among knockout mutants in different lineages of the Σ1278b background, and in some cases we noted differences between MATa and MATα strains, particularly in the “Heitman” Σ1278b background. Because this variation was consistent between experimental replicates, we reasoned that the phenotypic variation we observed was due to mutations that accumulated in each lineage during laboratory domestication. We used SNP calls from high-throughput sequencing data (Magwene, in prep.) to identify heterozygous sites in the diploid strain MLY61a/α, created from a cross of MLY40α and MLY41a. We then predicted which of these sites were heterozygous for premature stop codons (relative to the predicted peptide sequences of the reference strain S288c). Among the heterozygous sites we identified was a nonsense mutation in RIM15, a G >T transversion at position 1216 that converts a Gly codon to an opal stop codon (rim15:1216G>T). Rim15p is a protein kinase shown to play a key role in mediating developmental responses to nutrient conditions [39],[40]. The wild-type RIM15 encodes a 1770aa long protein. The rim15:1216G>T allele encodes a truncated protein with a predicted length of 406aa, which includes two putative functional domains (PAS and zinc-finger) [39], but not the kinase domain (Figure 5A). We confirmed the presence of two distinct alleles in the Heitman lineage by sequencing a 312bp portion of RIM15 covering the polymorphic site, from MLY61a/α, MLY40α, MLY41a, and G85 (Sigma2000). This confirmed that MLY61a/α was heterozygous, MLY40α, bore the predicted rim15:1216G>T allele, and MLY41a encodes the full length (wild-type) RIM15. G85 is homozygous for the wild-type allele.
The MATα strain, MLY40α, reproducibly develops a subtly weaker form of the complex colony phenotype than does the MATa strain, MLY41a (Figure 5B, top). We predicted that this was due to a partial or complete loss of Rim15p function. To test this we compared the colony morphology of XPY90a and XPY90α (rim15Δ::HygB derivatives of MLY41a and MLY40α respectively) [41] with that of MLY41a and MLY40α. As predicted, the rim15Δ mutants (Figure 5B, bottom) exhibited a colony morphology phenotype very similar to that of MLY40α and decreased relative to MLY41a (compare top and bottom rows of Figure 5B). We also noted differences between MATa and MATα strains for several of the deletion mutants we tested (Figure 5C and 5D). We predicted that these differences reflected epistatic interactions between RIM15 and the gene knocked out, such that a gene deleted in MLY41a was the expected single knockout, whereas the same deletion in MLY40α was effectively a double-mutant with rim15:1216G>T. To test this we crossed XPY5a (MATa, tpk2Δ) with XPY90α (MATα, rim15Δ) and MLY179α (MATα, mga1Δ) with XPY90a (MATa, rim15Δ) and analyzed how colony morphology segregated in tetrads relative to mating type and the gene deletions. The results of these crosses indicate the following: 1) both mutations at the RIM15 locus (rim15Δ and rim15:1216G>T) interact epistatically with mutations at the TPK2 and MGA1 loci such that the degree of colony morphology loss is greater than the sum of the single mutants (rim15Δ, tpk2Δ < rim15Δ or tpk2Δ and rim15Δ, mga1Δ < rim15Δ or mga1Δ); 2) the rim15:1216G>T allele may maintain some functionality because the degree of CCM reduction observed in mutants with this background are typically milder than those for comparable mutants in the rim15Δ background and; 3) there is still an effect of mating type on the degree of colony morphology independent of the RIM15 locus. These findings are illustrated in Figure 5B–5D. Results of the crosses are thus consistent with a model of synergistic epistatic interaction between RIM15 and other genes involved in colony morphology.
In addition to nutritional determinants, we observed a role for ploidy in the colony morphology response. Several strains that have simple or mild colony morphologies as diploids (MLY61a/α and OS17) exhibit strong colony phenotypes as haploids (MLY40α and NKY292) (contrast Figure 6C and 6E with Figure 6D and 6F). To further explore the colony morphology differences between isogenic haploids and diploids, we constructed haploid derivatives of a clinical isolate (YJM311) that exhibits a strong CCM phenotype as a diploid. We observed variation in colony morphology among the haploid derivatives of this strain, presumably due to allelic heterozygosity in the parental strain, but many displayed a morphology similar to that found in other haploid strains (compare Figure 6H with Figure 6D and 6F).
In order to confirm the role of ploidy in the colony morphology response we tested a set of isogenic haploid, diploid, triploid, and tetraploid strains [42] for colony morphology phenotypes in the Σ1278b.. We found an inverse correlation between ploidy and colony morphology; strains with ploidy of 2N and greater showed mild or no signs of complex colony morphology (Figure S5). Here as well mating type has a weak but noticeable affect on colony morphology independent of ploidy. The diploids heterozygous at the MAT locus (the normal state for diploids; Figure S5E) have simple morphology, while those homozygous for MAT have colonies that are somewhat elaborated (Figure S5B).
During our survey of growth conditions, we observed that colony morphology exhibits genotype-by-environment (G×E) effects. To provide a framework for study of G×E interactions we defined six morphotype classes: spokes (with weak concentric rings in this case) (Figure 1A), concentric rings (Figure 1B), lacy (Figure 1C), coralline (similar to lacy, but the cable-like structures are more angular, and tend to have more height) (Figure 1D), mountainous (possibly a variation on spokes) (Figure 1E), and irregular (which includes a wide range of forms that have no obvious regularity) (Figure 1F). For example, YJM311 grown on YEPLD media has a “lacy” morphotype (Figure 1C). The same strain grown on YEPEthanol, YEPIsopropanol, or YEPAcetate (Figure S6A, S6B, S6C) has a morphology that closely resembles a tangle of string (a variation of the lacy morphotype). On galactose, sucrose, and 1% agar YEPD the same strain exhibits the spoke morphotype, although each media induces a distinct version of the spoke morphotype (Figure S6D, S6E, S6F).
Having identified the key signals for the colony morphology response, we expanded our survey to include all 35 S. cerevisiae strains from the Saccharomyces Genome Resequencing Project (SGRP; [43]). Our goal was to determine the prevalence and diversity of complex colony morphologies and to identify strains of interest for future work. Of these thirty-five strains, by day six of growth, thirteen exhibited non-smooth or stronger colonies (anything beyond a smooth, shiny colony) on at least one media type. For most of these, this was simply a bumpy or textured colony surface, but six of these thirteen had at least “signs of CCM” (score of two or greater) (Figure S7).
In common with other developmental switches in yeast, the complex colony morphology response is induced by nutritional signals. Fermentable carbon source limitation coupled with an abundant nitrogen source appears to be the key trigger.
Taking our results together with information on other developmental responses sheds light on how S. cerevisiae responds to variable nutritional environments (Figure 7A). Haploid invasive growth, like complex colony morphology, is induced by dextrose limitation [4]. What seems to distinguish the two is the availability of other nutrients, particularly nitrogen. CCM competent strains grown on low dextrose synthetic media do not generally exhibit complex morphology. However, supplementing this synthetic media with glutamate is sufficient to induce the colony morphology response in most competent strains. In contrast, nitrogen availability seems to have little effect on haploid invasive growth [4].
Our findings also suggest a link between complex colony morphology and S. cerevisiae biofilm formation [12]. Like complex colony morphology, reduced dextrose is a trigger for biofilm development, and biofilms exhibit gross structural features resembling some of the colony structures we have observed [12],[44]. Cellular level organizational changes observed in starving colonies [18] might help explain how starvation signals result in macroscopic changes in both colony and biofilm structure.
The emerging picture of yeast development suggests that S. cerevisiae uses the core elements of two key signaling pathways, a MAP kinase cascade and a Ras-cAMP-PKA pathway, in multiple contexts [1],[3],[45]. The colony morphology phenotypes we observed in knockout strains implicate both of these pathways as playing key roles in regulating colony architecture (Figure 7B).
The filamentous growth/mating MAPK cascade (consisting, in part, of the kinases Ste20p, Ste11p, and Ste7p) regulates mating, filamentous and invasive growth, and cell wall integrity, in response to pheromone, nutrient limitation and osmolar stress respectively [46]. The mating and filamentous growth pathways both involve the transcription factor Ste12p, which induces expression of mating genes by binding pheromone response elements (PREs), and dimerizes with Tec1p to bind filamentous response elements (FREs) in the promoters of filamentation genes. Dig1p and Dig2p inhibit activation by Ste12p at PREs and by the Ste12p/Tec1p heterodimer at FREs [47].
Because multiple signals flow through the same core kinases of the MAPK cascade, several mechanisms are employed to prevent incorrect output. Knocking out genes in the cascade can disrupt this “insulation,” resulting in crosstalk between the pathways [28]–[30]. Such crosstalk is observed in hog1Δ mutants, which can be induced to mate by osmolar stress [29]. We observe similar crosstalk in the regulation of colony morphology. The diploid hog1Δ/ hog1Δ mutant exhibits colony morphology on low dextrose, high dextrose or alcohol containing media (Figure S4).
The crosstalk observed in MAPK cascade mutants complicates interpretation, but the loss of CCM in ste20Δ, ste11Δ, ste7Δ, ste12Δ, and tec1Δ mutants demonstrates that the MAPK cascade plays a key role in the regulation of colony morphology. We observed no gain of CCM in a diploid ste12Δ/ste12Δ, dig1Δ/dig1Δ, dig2Δ/dig2Δ triple mutant strain, but we did find a gain of CCM in the diploid tec1Δ/tec1Δ, dig1Δ/dig1Δ, dig2Δ/dig2Δ triple mutant. Our interpretation of this result is that when Dig1p/Dig2p repression of Ste12p is relieved, Ste12p is capable of activating a set of Tec1p independent targets, as has been show previously [48], and that this subset of targets affects colony morphology.
Our identification of ELP4 and TRM9 in the mutagenesis screen further argues for an important role of the MAPK cascade in regulating complex colony morphology. Abdullah and Cullen recently demonstrated a role for the elongator complex and other tRNA modification proteins in the MAPK dependent regulation of filamentous growth [32]. Elongator affects this pathway via starvation dependent induction of the signaling mucin gene MSB2, which interacts with Cdc42 to activate MAPK signaling [49]. Our independent identification of elp4Δ and trm9Δ mutants in this study adds to the evidence for a role for elongator and other tRNA modification complexes in regulating yeast development via the MAPK pathway.
In addition to the MAP kinase cascade, the colony morphology response also requires a functional Ras-cAMP-PKA pathway. Mutants that inhibit this pathway exhibit an attenuation of colony morphology, while those that up-regulate cAMP levels and/or PKA activation show an increased expression of complex morphology in diploid backgrounds.
A ras2Δ haploid mutant shows a loss of CCM consistent with similar decreases in biofilm formation and pseudohyphal growth observed for ras2Δ mutants [44],[50]. We also confirmed the observation of Halme et al. [51] that deletion mutants of IRA2 exhibit an increased colony morphology phenotype. Ira2p promotes Ras inactivation by stimulating GTPase activity, and treatment of cells with glucose destabilizes Ira2p, allowing active Ras proteins to induce cAMP production by adenylate cyclase [52].
There are three catalytic subunits of yeast PKA, Tpk1p, Tpk2p, and Tpk3p. Previous studies [41],[53],[54] have demonstrated distinct developmental and physiological roles for each of these subunits. For example, Tpk2p promotes filamentous growth and expression of Flo11p while Tpk1 and Tpk3p inhibit filamentous growth [41],[54]. Similar to these previous studies, we observed distinct effects of the PKA subunits on the colony morphology response. We found a loss of CCM in haploid tpk2Δ mutants as well as in tpk1Δ, tpk2Δ double mutants. The tpk2Δ, tpk3Δ double mutant showed a mild decrease in CCM. In diploids the tpk3Δ/tpk3Δ single mutant showed a background dependent increase in colony morphology. The tpk1Δ/ tpk1Δ, tpk3Δ/tpk3Δ double mutant also showed an increase in colony morphology. We did not observe a definite colony morphology phenotype in haploid or diploid TPK1 mutant strains or diploid TPK2 mutants. The opposite phenotypes of TPK2 and TPK3 mutants can be explained by a model put forth by Pan and Heitman [41] that suggests Tpk3p (and possibly Tpk1p) act in a negative feedback loop that attenuates cAMP levels. A candidate target for this feedback interaction via Tpk3p is the low-affinity phosphodiesterase Pde1p [55]. Our interpretation of this model and the mutants described above is that an active cAMP-PKA pathway is required for the development of complex colonies. Mutations that lead to a decrease in cAMP and/or PKA activation (ras2Δ and tpk2Δ) also decrease complex colony morphology and those that increase cAMP levels (ira2Δ and tpk3Δ) promote the development of complex colonies.
Given that a good nitrogen source seems to be a requirement for complex colony morphology, it is perhaps surprising that we observed a loss of CCM in a gln3Δ mutant. Gln3p is a transcriptional activator that activates “nitrogen starvation genes,” genes repressed by preferred nitrogen sources such as glutamate and ammonium. Under good nitrogen conditions, Gln3p is sequestered in the cytoplasm by Ure2p. Nitrogen deprivation leads to dissociation of Gln3p from Ure2p, Gln3 then localizes to the nucleus [23]. However, ours is not the first study to observe unintuitive results with respect to the effects of nitrogen catabolite repression pathway mutants on yeast development. Lorenz and Heitman [56] found that both a gln3Δ/gln3Δ mutant and a ure2Δ/ure2Δ mutant are defective in pseudohyphal growth. These results suggest that a balance between Ure2p and Gln3p may be necessary for appropriate response to nitrogen levels.
We find that ploidy has a major effect on colony morphology phenotypes. In some strains, this is manifested as a decrease in colony morphology in diploids relative to haploids; in others, there is simply a change in the stereotyped colony morphotype with ploidy. For example, colonies of Σ1278b haploids strains develop complex morphologies within six days, whereas diploid strains take much longer [19]. It has been proposed that this ploidy difference in colony morphology is linked to the ploidy specific expression of FLO11 [19],[42],[57]. The role of ploidy in the colony morphology response is another link between colony morphology and biofilm formation, which is also stronger in haploids [12]. There is presumably also a connection to the ploidy specificity of filamentous growth [58]. Pseudohyphal growth is a behavior of diploids starved for nitrogen, whereas the similar haploid invasive growth is induced by fermentable carbon limitation [4].
The crosses we carried out using rim15 mutants demonstrate that some of the mating type differences we observed in the Heitman Σ1278b lineage were the result of polymorphism for a loss-of-function allele in the RIM15 locus (rim15:1216G>T). This allele, present in the MATα background, was associated with a weaker CCM phenotype. However, after breaking this linkage, we still find residual CCM variation that segregates with mating type. MATα strains consistently exhibit a weaker version of the CCM phenotype than do matched MATa strains in the Heitman background, regardless of the allelic state of RIM15. We observe a similar direction of difference between MATa and MATα in the Fink Σ1278b lineages.
The flocculin Flo11p is known to be involved in several developmental processes, including filamentous growth [59] and biofilm formation [12]. There is a great deal of previous evidence of a role for FLO11 in colony morphology. For example, FLO11 was shown to be required for the “wrinkled” colony morphology observed in Ira- mutants [51], and insertion of a wild “flor” allele of FLO11 into a laboratory-domesticated strain induces the formation of “compact fluffy colonies” [35]. Finally, FLO11 is expressed at higher levels in a strain with complex morphology than a strain with simple morphology, but at very low levels in both [16]. Our finding that haploid flo11Δ strains fail to form complex colonies is consistent with these observations. The key stimuli we identify here, glucose and nitrogen, are both known to influence the expression of FLO11 [59],[60]. However, high levels of FLO11 expression are clearly not the sole determinant of colony morphology, since FLO11 is upregulated in diploid cells grown on SLAD (low nitrogen, high glucose) [59]. Growth on SLAD triggers pseudohyphal growth, but not the complex colony response.
Rim15p is a protein kinase that is thought to play a central role in the integration of nutrient signals [39]. RIM15 was first identified in a screen for mutants defective in the expression of genes expressed early in meiosis [61]. Subsequent studies [40],[62] have demonstrated that Rim15p helps to regulate entry into G0 (stationary phase) in response to nutrient depletion, particularly glucose, by regulating the expression of a large number of stress responsive genes. Current models [39],[63],[64] posit that Rim15p integrates signals from at least three major nutrient signaling pathways, the Ras-cAMP-PKA, Sch9, and TOR pathways. Rim15-dependent effects on transcription are mediated by the transcription factors Msn2, Msn4, and Gis1 [65].
We identified and analyzed a loss-of-function mutation in RIM15 (rim15:1216G>T) that contributes to variation in colony morphology phenotypes among lineages of the laboratory strain Σ1278b. Our results support a model of genetic interactions in which RIM15 mutations have a modest effect on colony morphology by themselves, but can exhibit significant epistatic interactions in combination with mutations at other loci. The SNP we observed is also a strong candidate as a contributor to subtle colony morphology differences between the Heitman Σ1278b lineage and the Sigma2000 lineage. This mutation may also contribute to differences in related developmental responses, such as pseudohyphal growth, that have been noted by other investigators [27].
Since glucose limitation causes hyperphosphorylation and nuclear accumulation of Rim15p [62], and glucose limitation is also a strong inducer of complex colony morphology, we hypothesize that the CCM defects we observe in RIM15 mutants are due to a failure to trigger the upregulation of stress responsive genes via Gis1 and Msn2/4. However, the mutant phenotypes also point to the existence of one or more RIM15 independent pathways, since RIM15 mutants do not show a complete loss of colony morphology, even when combined with knockouts at other loci. One possibility is that FLO11 expression is necessary but not sufficient to induce robust colony morphology, and that Rim15p signaling might be needed to amplify or intensify the strength of the CCM response in a FLO11 independent manner.
What role, if any, does the complex colony response play in yeast ecology? It has been proposed that complex morphologies help to protect against a hostile environment [17], and the observation that some strains switch to simple morphologies after a small number of passages on rich media (i.e. auspicious conditions) may support this hypothesis [16]. It has been observed that starvation results in reorganization of yeast colonies at the cellular level [18], and there is evidence that budding patterns and distributions of cell shape are different in complex colonies than simple colonies [19]. Extensive extracellular matrix is produced by complex colonies, and is absent from simple colonies [16]. The role that we demonstrate here for RIM15 in mediating colony morphology helps to more clearly link colony morphology to stressors such as oxidative stress [65] and calorie restriction [64], where Rim15p plays an important role in mounting transcriptional responses.
Colony morphology is a phenotype that is ripe for further research. The work presented herein provides a foundation, in terms of signals and pathways, for future studies of the developmental circuitry underlying the complex colony response. While we have found important genetic intersections between colony morphology and other developmental pathways, there is clearly not a complete overlap. We found no clear change in colony morphology in many of the knockout strains we tested that are known to have altered filamentous growth. Conversely, we have identified a number of genes, such as RRT12 and RIM15, that are known to affect sporulation, but have never been shown to have filamentous growth phenotypes.
The key cellular factors that contribute to the morphogenesis of complex colonies are largely undefined. Factors such as strength of adhesion, bud location, cell shape, spatially and temporally variable rates of cell division and cell death, secretion of extracellular matrix, and other such variables must contribute in some way to establishing and maintaining colony architecture during colony growth. Future studies that exploit genetic variation among strains along with mutants and cellular reporters will help to unravel this fascinating morphogenetic process.
Complex colony morphology, together with mating, filamentous growth, biofilm formation and sporulation, represent outputs of a complex decision-making machinery that integrates information on internal cell state, nutrients, potential mating partners, and various environmental stresses. A major challenge moving forward will be to better understand how simple eukaryotes such as yeast are able to correctly discriminate between different combinations of signals and how they are able to generate a diversity of responses given that the same core signaling pathways are used in different contexts.
YEPD and Hartwell's Complete (HC) media, were made as described in Burke et al. (2000). YPD+G418 and YPD+G418+HygB contained 200mg/L geneticin. YPD+HygB and YPD+G418+HygB contained 300mg/L hygromycin B. YEPGalactose, YEPSucrose, YEPAcetate, YEPEthanol, YEPIsopropanol are the same as YEPD, except with 2% of the named carbon source (e.g. galactose in YEPGalactose) substituted for 2% dextrose. Modified YEPD media were made in the same manner as YEPD with changes as noted: 1% agar YEPD; 4% agar YEPD; 0.5% yeast extract 1% peptone YEPD; 2% yeast extract 4% peptone YEPD; 4% dextrose YEPD; 1% dextrose YEPD; 0.5% dextrose YEPD (YEPLD); 0.25% dextrose YEPD; 0.125% dextrose YEPD; 0.0625% dextrose YEPD. For “wetted” media, 400 µl of water was added to each plate and allowed to absorb; “dried” media was treated by incubation at 40C for two days. Modified synthetic complete (SC) media were made according to Kaiser et al. [66], with the following changes: 0.5% Dextrose SC (SCLD); 0.5% Dextrose SC, 50mM L-Glutamic acid monosodium salt monohydrate (SCLD+Glu); 0.5% Dextrose SC -uracil, 50mM L-Glutamic acid monosodium salt monohydrate (SCLD-Ura+Glu).
All strains used in this work are listed in Table S2. Strains used are of diverse origin, including laboratory strains as well as clinical, distillery isolates.
To generate haploid derivatives of the homothallic diploid YJM311, the HO endonuclease was knocked out by transformation with the HO-poly-KanMX4-HO plasmid [67]. Knockouts were confirmed by PCR of the HO locus, then sporulated and tetrads were dissected. Haploid gene knockout strains PMY566, PMY568, PMY570, PMY572, PMY575, PMY577, PMY579, PMY581, PMY583, PMY585, and PMY589 were derived from diploids [26] by sporulation and tetrad dissection.
The environmental conditions tested are detailed in Table S1. Cells were plated with a targeted density of 20 or 60 cfu/plate. Several of the strains used in this study form flocs and/or aggregates of incompletely budded cell clusters. In order to accurately determine titers to plate a consistent number of cells, cultures were washed, then incubated for 15 minutes at room temperature in deflocculation buffer (90 mM mannose, 20 mM citrate, (pH 7.0), 5 mM EDTA) [68], briefly sonicated, then counted by hemocytometer. In addition to, or instead of this spreading procedure, some assays of colony morphology were conducted by pinning a small amount of yeast cells from a colony or water suspension directly to the assay plate.
For the initial survey of growth conditions, most strain-by-condition combinations were tested at two plating densities: 20 cfu/plate (results shown here) and 60 cfu/plate (data not shown). Results were similar for both plating densities. The strain-by-condition combinations not replicated are ones that showed no CCM: neither of the S288C derivatives (BY4743 and BY4739), were replicated; the wetted, dried, and room temperature conditions were also not replicated.
Once carbon limitation was determined to induce the colony morphology response, we found that YEPD with 0.5% Dextrose (Yeast Extract, Peptone, Low Dextrose - YEPLD), to be nearly optimal, strongly inducing the response while allowing sufficient colony growth to permit development of characteristic morphology (Figure 3). This medium was therefore used as a standard for subsequent experiments.
For the treatment screen, all plates were scored for colony morphology every day from day one to day six. Haploid derivatives of YJM311 were scored on day six. Mutant strains were scored on day 6 and compared to parental wild-type colonies.
We developed a qualitative method of scoring colony morphology using a scale from zero to five, based on the complexity and definition (depth) of morphology structures. While this framework is subjective, all scoring was performed by a single individual to ensure consistency. Scores were determined based on the survey of all the colonies on a plate, rather than a single colony (although for almost all plates, the colonies on a plate all had very similar morphology). The numerical scores have the following meanings: (0) No colonies or microcolonies; (1) Simple colony morphology; (2) Hints of colony morphology; (3) Weak or early colony morphology; (4) Strong colony morphology; (5) Very strong colony morphology (Figure S8). In summary, colonies that have no signs of CCM, but have a non-smooth surface texture receive a score greater than one but less than two. Colonies that have some signs of CCM receive a score of two or greater but less than three. Colonies that have definite morphology receive a score of three or greater.
Genome-wide transposon-mediated mutagenesis was carried out following the methods of Kumar and Snyder [69], with modifications as noted, using an mTn7 mutagenized S. cerevisiae genome library generated by Kumar et al. [31]. Briefly, individual pools of mutagenized library were digested with Not I to linearize, then transformed [70] into PMY574. The transformation reactions were plated onto SCLD-Ura+Glu, to simultaneously select for transformants and induce colony morphology. Colonies displaying a loss of complex morphology relative to PMY574 were picked and pinned to YEPLD to confirm the colony morphology phenotype. DNA was extracted from loss-of-morphology mutants using the DNeasy Blood & Tissue Kit (Qiagen), following the supplementary protocol for yeast DNA. Transposon insertion locations were identified by two-step PCR (ST-PCR) [71]. Primers mTn [69] and THG.SEQ [72] were used as ST-PCR primer 1 and primer 3 respectively to amplify from the “left” end of mTn7, and primers mTn7_5895R (GCACTGTTTTTATGTGTGCGATA) and mTn7_6007R (GCCGTTTACCCATACGATGT) were used as ST-PCR primer 1 and primer 3 respectively to amplify from the “right” end of mTn7. Primers 2 and 4 were as described [71]. Primers THG.SEQ and mTn7_6007R were used for sequencing ST-PCR products from the left and right ends, respectively. Finally, sequencing reads were BLASTed against the S. cerevisiae genome in order to locate their position within the genome.
Genes identified by mutagenesis were confirmed for colony morphology phenotypes by construction of knockout mutants in the PMY574 and PMY575 backgrounds. Primers used for gene deletion and deletion confirmation were based on primer sequences generated by the Saccharomyces Genome Deletion Project [73], however the UP_45 and DOWN_45 ORF specific oligonucleotides were joined with primers specific for the pRS400 plasmid series, and were used to amplify the URA3 fragment from pRS406 [74]. Transformants were selected on SC –uracil, then assayed for colony morphology phenotype by growth on YEPLD.
XPY5a was crossed with XPY90α to generate diploids heterozygous for deletions at the RIM15 and TPK2 loci. MLY179α was crossed with XPY90a to generate diploids heterozygous for deletions at the RIM15 and MGA1 loci. Diploids were selected by growth on YPD+G418+HygB, then sporulated and tetrads were dissected. Segregation of the RIM15, TPK2, and MGA1 alleles was determined by assaying growth of segregants on YPD+HygB, YPD+G418, and YPD+G418 respectively. Mating type of segregants was determined by crossing with AAY1017 and AAY1018, then assaying for growth on SD. Colony morphology phenotypes of segregants were assayed by growth on YEPLD.
|
10.1371/journal.pcbi.1002344 | Adjusting Phenotypes by Noise Control | Genetically identical cells can show phenotypic variability. This is often caused by stochastic events that originate from randomness in biochemical processes involving in gene expression and other extrinsic cellular processes. From an engineering perspective, there have been efforts focused on theory and experiments to control noise levels by perturbing and replacing gene network components. However, systematic methods for noise control are lacking mainly due to the intractable mathematical structure of noise propagation through reaction networks. Here, we provide a numerical analysis method by quantifying the parametric sensitivity of noise characteristics at the level of the linear noise approximation. Our analysis is readily applicable to various types of noise control and to different types of system; for example, we can orthogonally control the mean and noise levels and can control system dynamics such as noisy oscillations. As an illustration we applied our method to HIV and yeast gene expression systems and metabolic networks. The oscillatory signal control was applied to p53 oscillations from DNA damage. Furthermore, we showed that the efficiency of orthogonal control can be enhanced by applying extrinsic noise and feedback. Our noise control analysis can be applied to any stochastic model belonging to continuous time Markovian systems such as biological and chemical reaction systems, and even computer and social networks. We anticipate the proposed analysis to be a useful tool for designing and controlling synthetic gene networks.
| Stochastic gene expression at the single cell level can lead to significant phenotypic variation at the population level. To obtain a desired phenotype, the noise levels of intracellular protein concentrations may need to be tuned and controlled. Noise levels often decrease in relative amount as the mean values increase. This implies that the noise levels can be passively controlled by changing the mean values. In an engineering perspective, the noise levels can be further controlled while the mean values can be simultaneously adjusted to desired values. Here, systematic schemes for such simultaneous control are described by identifying where and by how much the system needs to be perturbed. The schemes can be applied to the design process of a potential therapeutic HIV-drug that targets a certain set of reactions that are identified by the proposed analysis, to prevent stochastic transition to the lytic state. In some cases, the simultaneous control cannot be performed efficiently, when the noise levels strongly change with the mean values. This problem is shown to be resolved by applying extra noise and feedback.
| There have been numerous experiments conducted on a wide range of organisms such as prokaryotic [1]–[3] and eukaryotic [4], [5] cells including mammalian cells [6], [7], to study gene expression noise. The noise originates from randomness in biochemical processes involving in transcription-translation, shared synthesis-degradation mechanisms [8], the cell cycle [9], [10], and other unidentified processes. Stochastic gene expression can lead to significant phenotypic cell-to-cell variation. For example, the stochasticity can help cells survive in stress environment [11]–[13] or determine the fate of viruses between latency and reactivation by randomly switching the two states [14], [15]. In metabolic networks, noise in enzyme levels causes metabolic flux to fluctuate and eventually can reduce the growth rate of host cells [16].
Although the measured noise is often explained by mathematical models [1]–[7], a systematic analysis on parametric control of noise has been lacking. This is attributed to the fact that noise propagation through pathway connections generates correlations between the pathway species [17], which make analysis difficult. Most noise control analyses have been focused on identifying the analytical structure of the noise propagation [17]–[19]. As the system size increases, the mathematical structure, however, becomes highly intractable. There have been some efforts to describe noise propagation in a modular way [18]. However, complicated feedback and feedforward structures in real biological networks hamper modular noise analysis.
Here, we are concerned with control of noise in biological systems such as gene regulatory networks and metabolic networks. In particular, we are interested in independent (orthogonal) control of noise and mean levels. For example, noise can stochastically switch one gene expression state to another via stochastic switching. This phenomenon was investigated in the expression of ComK that regulates DNA uptake in Bacillus subtilis [12]. The study used orthogonal control of noise to show that the reduction in the expression noise decreases the switching to competence [12]. Similarly, one can study how stochastic viral decisions [15] are made by independently changing the noise and mean levels of viral gene expression. Their individual contributions can be compared and used for identifying noise control schemes. This could eventually provide an efficient way to prevent viral activation. Here, we provide a systematic mathematical analysis method for simultaneous control of noise and mean levels and apply it to a number of well known biological examples.
We approach this control problem numerically by quantifying the parametric sensitivity of noise characteristics at the level of the linear noise approximation [20]. Our numerical approach, which we name stochastic control analysis [21], is practical in interpreting noise control experiments and computationally efficient and scalable in system size. Based on our analysis method, ‘active’ control of noise is proposed to manipulate the noise. We pursue various control schemes, such as independent control of mean and noise levels (such control will be called orthogonal), control of multiple mean and noise levels with certain ratios, and control of system dynamics of noisy oscillations. Active noise control can be applied to modify natural gene regulatory networks and improve their noise-related phenotype, and furthermore to design and construct gene regulatory networks for better performance by exploiting noise. It can be applied not only to gene regulatory networks but also to other biological systems such as metabolic networks [16].
In addition, we make a connection between noise control and network structure, and propose the mechanisms that could enhance the efficiency of orthogonal control. In a certain class of metabolic networks [22], probability distribution functions of each metabolite concentration were shown to be statistically independent of other species at the stationary state. The same result was also found in zero-range processes [23] in physics, complex balanced systems [24] and Jackson queueing networks [25] in mathematics. This independence was shown to be rooted to a certain network structure satisfying Feinberg's deficiency zero theorem [24]–[26]. We will show that when such species independence occurs, the orthogonal control of mean and noise levels is not possible, but that the application of extrinsic noise or feedback could help achieve orthogonal control.
For the purpose of noise control, we introduce stochastic sensitivities [21] called control coefficients (CCs) similar to the control coefficients in metabolic control analysis (MCA) [27]–[29]. These coefficients quantify the response of a system () from one stationary state to another due to a parameter perturbation (), mathematically defined by(1)The system parameters can include reaction rate constants [21], and the system responses include the mean and noise levels of concentrations and the temporal correlations of the concentrations (i.e., autocorrelations [30]).
CCs have been widely used in MCA for metabolic networks in the deterministic framework [27]–[29]. Here we use CCs to control noise in stochastic systems [3], [21]. Since noise can be considered a response of continuous perturbations in system parameters, the attributes of the dynamical response of the system (such as the period and amplitude of oscillations) [2], [3], [6], [31]–[34] can be deduced from noise characteristics, such as autocorrelations [30]. Thus, stochastic CCs also can be used to control system dynamics.
The noise level is defined as variance (covariance) divided by mean square (product of two mean values). We compute the noise levels and auto-correlations at the first level of approximation (see Methods) such that the noise level is assumed to be small enough that the rate laws can be linearized. From the computed noise levels and auto-correlations, we obtain the CCs (see Methods) to indicate where and by how much the system parameters are controlled.
In deterministic classical control theory [35] and MCA [36]–[39], the orthogonal control of system variables (flux and concentrations) has been studied. Here, we mainly consider orthogonal control in the stochastic regime to independently control mean and noise levels of concentrations. The noise level is often strongly anti-correlated with its mean level; for example, when a molecular species degrades with a first order reaction and is synthesized at a constant rate, the concentration level follows the Poisson distribution, where the variance is equal to the mean value, i.e., the noise level is equal to the inverse of the mean value. Thus, orthogonal noise control typically requires two or more parameters to be perturbed. In addition, the noise level shows non-local correlations between different species of molecules due to noise propagation [17], [40]. This also implies that a set of multiple parameters may need to be controlled simultaneously. Taking into account these points, we present a systematic non-local method for orthogonal control using the control coefficients.
We introduce a control vector(2)which is defined in an control parameter space. By the definition of the control coefficients, the inner product between and a parameter perturbation vector gives the change in the system variable due to the perturbation:By denoting the parameter perturbation vector by , the above equation becomes:(3)When parameters are perturbed in the direction of , a system variable (concentration mean or noise level) will increase. When are perturbed in one of the perpendicular directions to , the system variable does not change (one particular direction is . This corresponds to MCA-like summation theorems [21]).
For example, consider the following synthesis-degradation process:(4)where is a constant synthesis rate and a degradation rate constant. These two parameters are considered the control parameters: . We aim to reduce the noise level of without changing its mean level. At the stationary state, the mean synthesis rate equals the mean degradation rate: . Therefore, the mean level at the stationary state becomesThe noise level becomes , since the probability distribution of satisfies the Poisson distribution function and one of its properties is that the variance of is equal to the mean level of . Therefore, the noise level can be obtained asThe control vectors for the mean and noise level can be calculated by using the definition of CCs, Eq. (1):When the parameters are perturbed in the perpendicular direction of :with a non-zero real number, the mean level does not change [21]. However, since the noise level is the inverse of the mean value, the noise level does not change, either [21]. This is because the control vector for the noise level is anti-parallel with that of the mean value. Therefore, when a species concentration satisfies the Poisson distribution function, its orthogonal control is impossible. The appearance of the Poisson distribution is known to be generalized for a certain class of mass-action networks that satisfy complex balance [24]. We will show later that the application of extrinsic noise and feedback onto these networks enable orthogonal control.
In the last section, we saw a simple system, where we could not achieve orthogonal control. This begs the question, what networks can be controlled. This section describes how to answer this question and in addition, if controllable, how to determine the direction of parameter perturbations.
Consider that the vector of system variables , represented by , that is to be changed by percentage amounts via parameter perturbations . Once control coefficients are computed, the parameter perturbations can be obtained by solving Eq. (3). The unit vector of , denoted by , indicates the direction of control.
In the case of orthogonal control considered in the system (4), the mean level of (denoted by ) was aimed to be fixed, and its noise level (denoted by ) to be decreased, here for example by 3% . These system variables were controlled by perturbing and . Thus, Eq. (3) can be written in the following matrix form:This equation has no solution for , meaning that the desired control cannot be achieved and is overly-constrained. When the desired control is given by and (not an orthogonal control case), the control can be, however, achieved in various ways. Eq. (3) becomes simplified to . There are infinite number of solutions and Eq. (3) is then called degenerate.
In degenerate cases, we need to determine the direction of control that requires the minimum amount of change in system parameters for a given change in system variables. Mathematically, Eq. (3) can be solved for , where the norm () is minimized, by using the Lagrange multiplier method (see the Methods). is normalized to obtain the direction of control .
This section focuses on orthogonal control between two system variables, noise level and mean value . We aim to reduce the concentration noise level with its mean level fixed.
We consider single-promoter gene expression systems to show orthogonal control of noise and mean expression levels. Yeast promoter GAL10 [11], [41] and HIV-1 long terminal repeat (LTR) promoter [42] show significant gene expression noise that mainly originates from transcriptional bursting [11], [42]: Once chromatin structure is remodeled, RNA polymerase II enzymes, while waiting for the remodeling, can continue the transcription elongation process in a bursting manner [4], [11], [42]–[44]. This phenomenon has been modeled as a two-state model describing stochastic gene activation and deactivation [4], [42], [44] (cf. [11] and see Fig. 2b):where and denote inactive and active states of a promoter and the functions that are placed above or below the arrows are reaction rates, not constants.
Here we identify which parameter control scheme is optimal for noise and mean level orthogonal control. We constrain ourselves to the case that two parameters can be controlled for each experiment. For all possible two-parameter combinations, control efficiency and strength are computed, and the parameter combination leading to the best efficiency and strength is identified as the most optimal control scheme.
We can also apply our analysis to control dynamics. Temporal noise correlations have been used to understand the topology of gene networks and their dynamical properties, such as E. coli CRP-GalS-GalE feedforward related to galactose metabolism [3], HIV Tat-mediated positive feedback [6], and cell damage response of p53-Mdm2 [32]. Thus, sensitivity analysis on the temporal correlation can provide a method for controlling the attributes of the dynamics. We consider the cell damage response of p53-Mdm2 and its stochastic model presented in [32] (Fig. 6a). The model describes the stochastic fluctuations in p53 and Mdm2 by using Langevin equations with Gaussian white noise (Text S1), and provided successful explanations on sustained noisy oscillations in p53 and Mdm2 under DNA damage [32]. We apply the CCs for the autocorrelation to control the amplitude and period of the oscillations.
The autocorrelation of p53 shows damped-oscillations (Fig. 6c and d), implying potential sustained noisy oscillations. Here it is aimed to increase the oscillation amplitude or period. First, consider amplitude controls. An amplitude increase can be reflected in the autocorrelation as an increased vertical separation between troughs and peaks. For such an increase, the computed CCs at and hr (corresponding to the trough and peak; Fig. 6c) need to be large same-sign values. This control does not belong to orthogonal control since both the trough and peak heights need to increase together, and can be mathematically described by using Eq. (3):where both and are real same-sign values with similar magnitude, and indicates the value of autocorrelation at time .
We consider one-parameter controls, and then the inner products in the above equations become number products, indicating that and are real same sign values with the similar order of magnitude. This is well satisfied by the control coefficients corresponding to . Thus, we decreased by 50% and this led to a visible increase in the p53 oscillation amplitude (Fig. 6c). Experimentally, p53 effective degradation, , was reduced by introducing the small molecule Nutlin3 that inhibits p53 from binding to Mdm2 [50], [51] (the Mdm2-p53 complex shows enhanced degradation) and the oscillation amplitude was found to increase without affecting the period [51].
Second, consider period controls. The period increase causes the stretch-out of the autocorrelation in -axis. This implies that the CCs at and hr (corresponding to and in , which decreases and increases when the sine function shifts to the right, respectively) need to be large opposite-sign values, respectively (Fig. 6b). Mathemtically, Eq. (3) is expressed aswhere and are real opposite-sign values with similar magnitude.
For one-parameter control, the above equations indicate that and are real opposite-sign values with similar magnitude. Both the controls on and were found to be the best case. When one of these parameters was decreased to its 10% levels of the original value, a significant increase in the period was obtained (Fig. 6d). The decrease in or causes the ATM level to decrease and experimentally this can be achieved by decreasing -irradiation intensity [52]. (For cases without the irradiation, and can be considered to vanish, resulting in a second-order linear model in [32].) Our analysis based on control coefficients showed successful control on noisy oscillation. This can serve as an important tool for analyzing the parameter dependence of stochastic dynamics, particularly when an analogous deterministic counterpart does not exist.
In this section, we will investigate the relationship between noise control and network structure. To show the applicability of our analysis to other systems, we will consider metabolic networks. It has been known that noise at enzyme levels causes metabolic flux to fluctuate and eventually to reduce the growth rate of host cells [16] due to nonlinearity in the system and noise propagation [40], [53] from the enzyme to the pathway. Here, we consider linear metabolic pathways (Fig. 7) and aim to reduce the noise level of the end product ( in Fig. 7) without altering its mean level. One of the enzymes () is considered explicitly and is used to supply extrinsic noise to the metabolic network. If such orthogonal noise reduction is achieved, the decrease in the growth rate that would have occurred due to the noise propagation can be suppressed.
Here, we show that feedback in the metabolic network and noise propagation [17], [40] from enzyme fluctuation [16] play important roles in enabling orthogonal control. We will limit control parameters to for ease of comparison between the original network and its variants.
First, consider the metabolic network under a constant enzyme level (located in the first step) and without any feedback (Network A in Fig. 7). It is known that at the stationary state, the probability distribution function of the whole system takes a product form and that inter-species covariance vanishes [23]–[25], resulting in the cancellation of the net effect of noise propagation [22]. This cancellation is related to network structure; the product form distribution was derived for mass-action systems (and some non-mass-action systems) [24] that satisfy the deficiency zero theorem [26]. This theorem is only dependent on the network structure. Furthermore, it was shown that each individual concentration distribution satisfies the Poisson distribution function [23], [24]. This indicates that the mean and noise levels are inversely related and that their control vectors are anti-parallel. Therefore, orthogonal control of the mean and noise levels cannot be achieved for any metabolites: , , and . We verified this by computing the control vectors; for example, the control vectors for were obtained in the parameter space of :These vectors are anti-parallel, so control efficiency becomes zero. This fact implies that low control efficiency can be predicted by examining stoichiometry and topology.
Second, consider an end-product inhibition: negative feedback from to the synthesis of (Network B in Fig. 7). The covariances between metabolites were computed by using Eq. (8). The covariances between and and between and were found not to vanish. This implies that the stationary state does not take a product form distribution and that the Poisson distribution does not appear, either (the deficiency zero theorem [26] does not apply here, unless the mechanism of the feedback is expressed in terms of chemical reactions). Therefore, the control vectors will be no longer anti-parallel, providing the possibility of orthogonal control. The control vectors were computed:The control efficiency was significantly increased to 0.47, when compared with Network A.
Third, we consider the enzyme fluctuations in in the absence of negative feedback (Network C in Fig. 7). For this system, the product form distribution does not hold since the system is not weakly reversible (Text S1) [26] and the deficiency zero theorem does not apply. Noise originating from can be observed in metabolite fluctuations. The control vectors for were computed:The control efficiency was further increased to 0.72, when compared with Network B.
Finally, we allow both the noise propagation from and the end-product inhibition. The control vectors were computed:The control efficiency was decreased to 0.41, when compared with Network B and C. This is because the signs of the second and third elements of are opposite for Network B and C.
In the metabolic networks we consider, the application of extrinsic noise in or the end-product inhibition significantly enhanced the control efficiency. This implies that in the case when orthogonal control cannot be performed with a high efficiency, perturbations in the network structure such as stoichiometry and topology can enhance the control efficiency. The result presented here, however, may not be directly applicable to gene regulatory networks, since gene expression processes occur in cascades of transcription and translation and thus they are not weakly reversible (similar to the case of the Network C).
This section describes a computational protocol for iterative noise reduction. Since our analysis is based on differential sensitivities, infinitesimal perturbations can be continuously applied along the perturbation direction quantified by , to achieve a finite-size perturbation. At the first level of approximation, the finite but small enough size of perturbations can be applied iteratively. We consider again the previous metabolic network models.
We performed the noise reduction control in the following sequence.
We compare two cases with and without iteration. First, we performed a single large perturbation: . The noise was decreased by 36% (), and the mean level by 11% (). This non-negligible change in the mean level is due to the fact that the size of the perturbation is large enough that our analysis based on differential sensitivity becomes inaccurate. Second, we performed a series of small but finite perturbations: with 5 iterations by repeating the procedure (1)–(3). The noise level was significantly reduced by 50% (), with a minor mean level decrease of 1.4% (), as shown by the change in the probability distribution functions of (Fig. 8C).
The protocol we describe is mathematically equivalent to a first order Euler approximation to find the parameter trajectory satisfying the control aim , since the next parameter values are determined by the slope () calculated at the current parameter values. The mean values deviate from the desired constant level on the order of magnitude of : One Euler step updates parameters from to , causing the mean value, here denoted by , to change from to , where the second term in the right hand side vanishes since was set to be perpendicular to . Therefore, the magnitude of the change is of the order of .
In this paper we describe a systematic method for orthogonal control of noise and mean levels and provided its applications. In addition to these examples, our work can also be useful in synthetic biology.
In synthetic biology, biological organisms are engineered via design and construction of new useful biological functions that do not exist in nature. In synthetic gene regulatory networks (gene circuits), the signals are often considered the concentrations of transcription factors. Their copy numbers can be so low that their fluctuations are significant, meaning that the signals can be very noisy [54]. This causes cell-to-cell variability in gene expression levels and potentially their related phenotypes at the population and individual levels. In addition, the noise, both extrinsic and intrinsic, can propagate through a synthetic network [17], possibly preventing the predictable modular construction of circuits. From an engineering perspective, gene circuits have been designed and constructed based on the concept of modularity [33], [34], [55]–[60], to ensure predictable behavior when combining modular circuits. The reliability and predictability can be enhanced via simultaneous control of mean and noise levels by increasing signal-to-noise ratios and by suppressing unwanted noise propagation.
Noise control can also be used to improve gene circuit function. The properties of gene circuit components such as input-output responses can be engineered by exploiting noise. For example, noise can improve the sensitivity in a system response with respect to an input change via stochastic focusing [40], [53]. The noise can also help input signals be reliably transferred to output signals at a certain optimal level of intrinsic or input noise via stochastic resonance [61], [62]. These beneficial effects can be readily realized when the noise and mean levels can be independently controlled to their optimizing values.
For the p53 study, a frequency-domain analysis can be performed as an alternative approach. We can apply a Fourier transformation on Eq. (9), obtain its power spectral density, and compute control coefficients for the spectral density. The magnitude of the main spectral peak can be examined to quantify the oscillation amplitude, and the frequency corresponding to the main spectral peak can be used to determine the oscillation period. The reason that an autocorrelation function was used instead of its Fourier transform was that the numerical computation of the autocorrelation and its corresponding control coefficients can be performed without matrix inversion. Thus, it is computationally more efficient compared to using the spectral density, although control schemes for changing the period and amplitude might be more complex.
Our analysis is based on sensitivity to infinitesimal parameter changes and this was the reason that for the PHO5 promoter study the control parameters changed significantly depending on the specific parameter values, where the system responses to the parameter changes became highly nonlinear. Our approach can be used, however, as a first level of approximation for such cases, although the control schemes may not necessarily be the best ones. A global picture of controllability can be obtained by computing the sensitivities for various parameter values and determining the landscape of the sensitivities over the parameter space. If the landscape is flat, the proposed analysis can be applied to finite-size perturbations.
Our analysis is also based on the linear noise approximation [20]. The validity of this approximation needs to be verified on a case-by-case basis: The approximation depends on how large the noise levels in concentrations are compared to the (quasi-)linear region of the non-linear reaction rate functions. The strength of the noise level depends on how noise from the upstream network is propagated into the non-linear reaction rate functions. This means that the validity of the linear noise approximation crucially depends on noise propagation and the upstream network as well as the downstream non-linearity of reaction rate functions. Therefore, the linear noise approximation needs to be tested on a case-by-case basis. For the test, one possibility would be to use computer simulation for an exhaustive search.
Our analysis method can be applied to more complex networks than the systems previously considered. For example, consider the experiment on yeast cells performed in [63], where the expression noise of a reporter protein was controlled via transcriptional negative feedback. The reporter gene expression showed highly-sigmoidal dose-response in the absence of feedback, but it was linearized with the introduction of the feedback [64], [65]. The linearized dose-response led to smaller fluctuations in the response, when the input dose is centered around the sigmoidal region. Our analysis may be applied, for example, to increase/decrease the region of the linear dose-response by computing control coefficients for the mean levels of the reporter at two or more different input doses (e.g., three doses: , , and ) and by setting desired changes in the responses (, , and ), and by solving Eq (3) for :with and . For more complex control, where multiple mean and noise levels are controlled simultaneously, Eq. (3) can be used again to identify control schemes computationally.
In summary, we have proposed a numerical analysis method for adjusting noise-related phenotype by controlling system parameters of mathematical models. The analysis quantifies which parameters need to be controlled by how much, with scaled non-dimensional values. In addition, we proposed how to improve control efficiency by changing network structure when control efficiency is weak. We have shown that MCA-like summation theorems exist and that the analysis can be applied to stochastic biological systems such as gene regulatory and metabolic networks and not only for statics but also for dynamics.
We consider stochastic reaction systems described as continuous time Markov processes. Stochastic fluctuations in concentrations caused by random reaction events are assumed to be small enough that the reaction law can be linearized with respect to the mean values for the study of the fluctuations. Such assumption is called the linear noise approximation [20]. Under this approximation, the covariance matrix can be computed by solving the Lyapunov equation (also known as the fluctuation dissipation relationship [17], [66]):(8)with the Jacobian matrix and the diffusion matrix [17]. We compute noise levels () from :where is the temporal average concentration level of the -th species at the steady state. The autocorrelations are defined asThe autocorrelations can be computed by solving the following ordinary equation [31]:(9)for all , where is equal to . From Eq. (8) and (9), the noise levels and the autocorrelations can be computed numerically and analytically.
The CCs for the noise levels and the autocorrelations can also be computed from Eq. (8) and (9). For mathematical simplicity, we will denote the matrix component of by , with and representing vectors. The Lyapunov equation (8) is invariant under parameter perturbations from one steady state to another corresponding to before and after the perturbation:This can be expanded by using the chain rule:(10)where we have used and means the change in the concentration covariance matrix due to the change in , which defines an un-scaled CC for . and can be also expanded by applying the chain rule:(11)(12)where is an un-scaled control coefficient for mean concentration (for notation simplicity, instead of ). Under the linear noise approximation, concentration mean levels are obtained by using deterministic rate laws, neglecting noise propagation to the reaction rates [40]. Thus, the un-scaled CC can be obtained as in the deterministic MCA [27]–[29]:(13)where is a reduced stoichiometry matrix [67]. Equation (13) is substituted to Eqs. (11) and (12) and the resultant equations to Eq.(10), to numerically estimate the un-scaled CCs for , i.e., .
Next, we obtain CCs for noise level. The noise level is defined asThe un-scaled CCs for the noise level is expressed by applying the chain rule:(14)By substituting Eq. (13) and the computed to Eq.(14), the un-scaled CCs for the noise level, i.e., can be estimated and then converted to the scaled version:Next, we obtain CCs for autocorrelation functions. Equation (9) is invariant under parameter perturbations:Since can be estimated by using Eq. (9) and is equal to , un-scaled CCs for () can be obtained by solving the above equation. This un-scaled CCs can be converted to the scaled version:A MATHEMATICA file is provided for the estimation of CCs for noise levels in Text S2.
The Lagrange multiplier method will be used to obtain the direction of parameter perturbation for orthogonal control of two system variables, and , where is increased but remains fixed: and . For non-degerate cases, can be obtained by solvingwhere is a control vector for a variable . If the above equation is degenerate, the most optimal parameter perturbation needs to be identified. The solution can be considered optimal, if the net amount of parameter perturbations – the norm of – is smallest among all possible solutions. We introduce Lagrange multipliers and and the Lagrange function :and solveThe solution of the first equation,(15)is substituted in the second and third equations, which can be solved to obtain and :withBy substituting these equations to Eq. (15), we finally obtain the optimal paramter perturbation:This perturbation, , is normalized to obtain the direction of control, which is the same as that of the control vector expressed with Eq. (5) (since , and is negative if is considered as a noise level that is aimed to be reduced).
When the control vector for is small, does not change significantly when the parameter perturbation is directed even toward that of the control vector. This means that the orthogonal control can be effectively performed over a much wider set of parameter perturbations, not just limited to the perpendicular plane to the control vector for . The norm of the control vector for indicates the percentage change in caused by a unit parameter perturbation directed along the control vector for . As the value of the norm decreases, the degeneracy of the orthogonal control increases.
Mathematically, when the mean value is allowed to change up to a certain tolerance level (tol) under a unit parameter perturbation, the perpendicular plane can be expanded up to a certain angle () from the plane (Fig. 9a), which can be determined as follows:(16)This expanded perpendicular space (colored in Fig. 9) means that the control efficiency and strength need to be re-defined: The control vector for the noise level is projected on the expanded perpendicular space, and for the most efficient control, projected on the closest one. Thus, the control efficiency and strength are re-defined by replacing the angle to the minimal angle from to the expanded perpendicular space (see Fig. 9b–d):(17)
|
10.1371/journal.pntd.0000851 | Pre-Clinical Assays Predict Pan-African Echis Viper Efficacy for a Species-Specific Antivenom | Snakebite is a significant cause of death and disability in subsistent farming populations of sub-Saharan Africa. Antivenom is the most effective treatment of envenoming and is manufactured from IgG of venom-immunised horses/sheep but, because of complex fiscal reasons, there is a paucity of antivenom in sub-Saharan Africa. To address the plight of thousands of snakebite victims in savannah Nigeria, the EchiTAb Study Group organised the production, testing and delivery of antivenoms designed to treat envenoming by the most medically-important snakes in the region. The Echis saw-scaled vipers have a wide African distribution and medical importance. In an effort to maximise the clinical utility of scarce antivenom resources in Africa, we aimed to ascertain, at the pre-clinical level, to what extent the E. ocellatus-specific EchiTAbG antivenom, which was designed specifically for Nigeria, neutralised the lethal activity of venom from two other African species, E. pyramidum leakeyi and E. coloratus.
Despite apparently quite distinctive venom protein profiles, we observed extensive cross-species similarity in the immuno-reactivity profiles of Echis species-specific antisera. Using WHO standard pre-clinical in vivo tests, we determined that the monospecific EchiTAbG antivenom was as effective at neutralising the venom-induced lethal effects of E. pyramidum leakeyi and E. coloratus as it was against E. ocellatus venom. Under the restricted conditions of this assay, the antivenom was ineffective against the lethal effects of venom from the non-African Echis species, E. carinatus sochureki.
Using WHO-recommended pre-clinical tests we have demonstrated that the new anti-E. ocellatus monospecific antivenom EchiTAbG, developed in response to the considerable snakebite-induced mortality and morbidity in Nigeria, neutralised the lethal effects of venoms from Echis species representing each taxonomic group of this genus in Africa. This suggests that this monospecific antivenom has potential to treat envenoming by most, perhaps all, African Echis species.
| Snakebite is principally a health concern of rural poor communities. The high snakebite risk of subsistence farming and paucity of effective antivenoms in sub-Saharan Africa means that many communities remain unacceptably vulnerable to snakebite mortality and morbidity. There is therefore a compelling need to maximise the utility of the snakebite therapies that are available. To address Nigeria's severe snakebite problem, the government funded a collaboration of ministry officials, antivenom manufacturers and academics (the EchiTAb Study Group) to produce, test and deliver antivenom. Accordingly, we prepared EchiTAbG, an antivenom specific for envenoming by the saw-scaled viper (E. ocellatus) which is responsible for 80% of snakebite deaths in Nigeria. Since E. ocellatus is widely distributed across the West African savannah, EchiTAbG offers considerable therapeutic promise in many countries in the region. Since other Echis species represent public health concerns elsewhere in Africa, the objective of this study was to examine the pre-clinical intra-generic venom-neutralising efficacy of EchiTAbG. Our results suggest that EchiTAbG (Nigeria registration: A6-0078) has pan-African efficacy against Echis envenoming indicating that costly investment in region-specific antivenoms therefore may not be required. This represents an important progression to minimise development costs and maximise the delivery of snakebite therapy for the continent.
| The rural communities of sub-Saharan Africa suffer the multiple burdens of low economic status, inadequate access to effective health care and the debilitating effects of numerous infectious and parasitic diseases. The subsistence agriculture livelihood, non-mechanised farming techniques, remote locations and proximity of homes to farms/grain stores all contribute to the fact that these communities also suffer a disproportionally high snakebite mortality rate [1]. Extrapolations from recent global snakebite incidence and mortality data [2] reveal that while the percent lethality of snakebite in Latin America is 1.8% (2,300 deaths; 129,000 incidences) it is 7.6% in sub-Saharan Africa (32,000 deaths; 420,000 incidences). These rather crude data-extrapolations are presented simply to emphasise the point that circumstances in sub-Saharan Africa make snakebite a more life-threatening event than elsewhere.
While socioeconomic issues at the community level and per capita government expenditure on health at the national level certainly contribute to this disparity [1], an important additional explanation is the relative scarcity of antivenom in Africa [3]–[6]. IgG antivenom is the most effective treatment of systemic snake envenoming. However, its manufacture from sera of venom-immunised horses or sheep means that antivenom is a more expensive therapy (US$100/vial in S Africa) than many other non-subsidised medicines administered in sub-Saharan Africa. As described in the cited and related literature, the (i) relatively high cost of antivenom, (ii) its restricted efficacy to the species of snake whose venom was used in its manufacture and (iii) factors relating to commercial manufacturing incentives have all combined to severely limit the availability of antivenom in Africa. There is therefore a compelling need to maximise the clinical utility of effective antivenoms that are becoming available in the region.
In response to the crisis in antivenom supply affecting Nigeria, the EchiTAb Study Group (a collaboration between the Nigerian Federal Ministry of Health, antivenom manufacturers in Costa Rica (Instituto Clodomiro Picado) and Wales (MicroPharm Ltd) and academics in the Liverpool School of Tropical Medicine and University of Oxford) has organised the production, pre-clinical testing [7], [8], human efficacy testing [9] and delivery of two new antivenoms for Nigeria, EchiTAb-Plus-ICP and EchiTAbG.
The saw-scaled viper, Echis ocellatus, is responsible for most snakebite-deaths in Nigeria [10]–[12]; the other medically-important snakes are the puff adder, Bitis arietans, and the spitting cobra, Naja nigricollis [13], [14]. EchiTAb-Plus-ICP is a new equine polyspecific IgG antivenom developed in Costa Rica to treat envenoming by all three snake species [7]. In view of the very high number of E. ocellatus bites in Nigeria and the severe haemorrhaging and incoagulable bleeding experienced by systemically envenomed victims, an additional ovine IgG antivenom, EchiTAbG (MicroPharm Ltd) was manufactured in Wales for the treatment of E. ocellatus envenoming and has been registered (A6-0078) by the Nigerian medicines authority, NAFDAC. Since E. ocellatus is widely distributed across the West African savannah (Figure 1), EchiTAbG offers considerable therapeutic promise in many countries in the region. With the intent of maximising the clinical contribution of this new antivenom and cognisant that (i) other Echis species represent public health concerns in East (E. pyramidum) and North-East (E. coloratus) Africa (Figure 1) and (ii) that their venom protein family composition is not dissimilar to E. ocellatus [15], the objective of this study was to examine the pre-clinical intra-generic venom-neutralising efficacy of EchiTAbG.
To achieve a more complete understanding of the immunology underpinning the cross-specific venom neutralising potential of EchiTAbG, we performed a series of assays to determine the IgG titre, specificity and relative avidity of sera raised in sheep immunised with venom from species representing each taxonomic group of the Echis genus in Africa [16] as follows: (i) E. pyramidum leakeyi (Kenya) representing the pyramidum complex which includes E. leucogaster and E. p. pyramidum, (ii) E. coloratus (Egypt) representing this species and E. omanensis, (iii) E. ocellatus representing this species and E. jogeri. To assess the geographic limitation of the exercise we included venom from a non-African saw-scaled viper (E. carinatus sochureki, United Arab Emirates) in all the experiments, including provision of species-specific antisera.
Snakes were maintained in the Herpetarium at the Liverpool School of Tropical Medicine. Venom was extracted from wild-caught specimens of E. ocellatus (Nigeria), E. coloratus (Egypt), E. p. leakeyi (Kenya) and E. c. sochureki (UAE) on several occasions, pooled, frozen, lyophilised and stored at 4°C prior to reconstitution in phosphate-buffered saline (PBS).
Antisera were generated against venom from E. p. leakeyi, E. coloratus and E. c. sochureki using protocols identical to that used in the production of the E. ocellatus-specific antivenom, EchiTAbG. Six sheep (two per venom) were initially immunised with 0.5mg of venom emulsified with Freund's Complete Adjuvant followed by boosting immunisations of 1.0mg of venom emulsified with Freund's Incomplete Adjuvant every 28 days. To maximise seroconversion, venom immunisations were administered subcutaneously in six sites close to the major draining lymph nodes in the neck and groin. Blood samples were taken 14 days after immunisation.
Once the anti-venom IgG had reached a plateau (16 weeks, personal communication, MicroPharm Ltd) one litre of blood was taken, allowed to clot, centrifuged and the sera stored at −20°C. IgG was extracted by the addition of caprylic acid (Sigma, UK) to a final concentration of 5% and stirring vigorously for two hours to precipitate non-IgG proteins. The suspension was centrifuged at 13,000 rpm (in a microcentrifuge) for 60 min and the supernatant IgG dialysed with three changes of 20 mM sodium phosphate buffer (pH 7.4). All the IgG preparations were formulated to the same concentration, 30mg/ml in PBS, and stored at −20°C. IgG generated against E. ocellatus venom and the E. ocellatus antivenom EchiTAbG were obtained from MicroPharm Ltd.
Note: to avoid possible confusion between the antivenoms resulting from commercial manufacture and the anti-Echis species IgG antivenoms prepared in our laboratory – the latter have been termed IgG antisera.
All animal experimentation was conducted using protocols approved by the University of Liverpool Animal Welfare Committee and performed under licenced approval of the UK Home Office.
The reduced SDS-PAGE profiles reveal intra-generic variation in molecular mass and quantitative representation of the venom proteins present in the four Echis venoms (Figure 2A). However, immunoblotting of the four venoms with each of the four Echis species-specific IgG antisera demonstrated that the intensity of immunoreactivity of each IgG antiserum to proteins of the homologous venom was matched by that to the three heterologous venoms (Figure 2B–E). Indeed, such was the intensity of the cross-species IgG immunoreactivity that we were unable to detect any protein in the SDS-PAGE venom profiles that was not reactive with each of the four Echis species-specific IgG antisera. Furthermore, the immunoblots revealed the existence of more venom proteins than was apparent from the SDS-PAGE. We interpret this analysis as illustrating that while intra-generic differences in the size of the venom proteins exist, these proteins were likely to be size-variants of the same protein families.
The immunogenicity of each Echis venom was assessed using the EPT ELISA assay to determine the IgG titre of each of the four Echis species-specific IgG antisera to each Echis venom (Figure 3A–D). The overall plateau and then decline of IgG titre after successive IgG dilution was strikingly similar for each of the IgG antisera; as reflected in the assigned EPT IgG titres (Table 1). The slightly slower decline of the E. coloratus venom-antisera profiles against each of the venoms (except the E. c. sochureki venom, Figure 3D) suggested that E. coloratus venom is perhaps more immunogenic than the other Echis venoms. Similarly, because the E. p. leakeyi IgG antisera showed the most rapid decline in IgG titre against its homologous and the heterologous venoms, it could be surmised that E. p. leakeyi venom is the least immunogenic. However, these differences were minor and we therefore caution against assigning much immunological significance to these observations.
Considerable intra-generic immunological cross-reactivity was noted from the immunoblotting and ELISA assay when the venoms were in ‘reduced’ or ‘native’ states (respectively). We next wished to examine this immunological cross-reactivity using a technology also using ‘native’ venoms but in a manner that would likely present the venom proteins to the IgG in a different configuration - and perhaps better reflecting the in vivo situation - than that achieved in the ELISA assay. We therefore prepared small scale CnBr-activated venom-affinity columns for each Echis venom and measured the amount of each Echis species-specific IgG that remained bound to the column after extensive washing of the column (Table 2). This new assay to measure venom-antivenom interactions revealed, in each case, that highest binding occurred between the homologous combination of venom and IgG antisera. This was the first assay to indicate that there are intra-generic differences in Echis venoms with immunological significance.
To determine the strength of venom-antivenom binding in the presence of ammonium thyiocyanate (which disrupts protein-protein interactions), we used the Relative Avidity ELISA assay to examine the titre of the four Echis species-specific IgG antisera to each of the venoms in the presence of increasing amounts of the chaotrope (Figure 4). Consistent with the results of the small scale venom affinity assay, the venom-antivenom interactions least affected by the chaotrope, therefore exhibiting the strongest binding, were between the homologous combinations of venom and IgG antisera. The E. ocellatus IgG antisera exhibited the strongest binding and the E. c. sochureki IgG antisera showed the overall weakest binding to other Echis venoms (Figure 4).
The lethal effects (expressed as LD50s) of the four Echis venoms to mice, ranged from 9.81µg venom/mouse for E. coloratus to 15.10µg for E. c. sochureki. The 95% confidence limits indicate there is no significant difference between the venom lethalities of the four Echis species (Table 3).
Only slightly varying volumes of the E. ocellatus antivenom, EchiTAbG, was required to completely neutralise venom-induced lethality of the three African Echis species (E. ocellatus, E. p. leakeyi and E. coloratus; Table 3). Perhaps surprisingly in light of results of the small scale venom-affinity and Relative Avidity assays, EchiTAbG was more effective against E. coloratus venom than against its homologous venom, E. ocellatus. Importantly, EchiTAbG was ineffective at the maximal permitted volume of antivenom (100µl) against the Asian species, E. c. sochureki.
We were concerned with the apparently conflicting observations that although the E. c. sochureki venom exhibited the lowest toxicity (highest LD50 result), EchiTAbG was unable to neutralise its venom effects. We therefore performed an ED50 test with E. c. sochureki venom and its homologous IgG antisera (Table 3); the latter proving as effective against E. c. sochureki venom (54.42µl/mouse) as EchiTAbG against E. ocellatus venom.
The efficacy of EchiTAbG against the lethal effects of E. ocellatus venom noted in this study (an ED50 of 58µl antivenom/mouse: 1740µg antivenom/mouse) was substantially lower than that recently reported for this antivenom-venom combination (ED50 of 8µl) in a separate study [8]. The E. ocellatus venom LD50 results here and from the other study [8] are comparable and consistent with other publications [18], [20] indicating that the disparate ED50 results did not arise from batches of E. ocellatus venom with different toxicities. We have repeated the ED50 experiment with different batches of EchiTAbG antivenom and E. ocellatus venom in published [18] and unpublished experiments (data not shown) with results that confirm those obtained in this study.
The results of the ED50 assays demonstrate a lack of congruence between the results of in vivo pre-clinical tests and immunological assays. Notably, no single immunological assessment could be used to predict the pre-clinical efficacy of EchiTAbG. Only the Relative Avidity ELISA results indicated the potential ineffectiveness of EchiTAbG against E. c. sochureki venom (Figure 4). We interpret this as indicating that while all effective antivenoms require high levels of IgG titre, specificity and avidity [18], [21], these immunological characteristics cannot be used to predict antivenom efficacy.
Physicians throughout Africa are tasked with treating victims suffering life-threatening effects of envenoming that include systemic haemorrhage, coagulopathy, neurotoxicity and renal failure. Identifying the snake species is often difficult - making it problematic to select the most appropriate antivenom, which, in a resource-poor setting, are scarce and expensive. The problem is made more complex because the snake species responsible could be any of the following a) viper species (the Echis saw-scaled vipers; the puff adder, B. arietans; and several other Bitis species which, although incidences are rare can be potentially lethal), b) elapid species (the black-necked spitting cobra, Naja nigricollis; the Egyptian cobra, N. haje; the Mozambique spitting cobra, N. mossambica; the forest cobra, N. melanoleuca; species of the mamba genus including the black mamba, Dendroaspis polylepis and green mambas of East Africa, D. angusticeps and Central/West Africa, D. jamesoni and D. viridis) and c) the colubrid Boomslang, Dispholidus typus. Presumably in consideration of the above, the SAIMR polyspecific antivenom manufactured in South Africa (South African Vaccine Producer) includes venoms from many of the above snakes in its venom-immunisation mixture.
The EchiTAb Study Group's decision to manufacture an E. ocellatus-specific antivenom and a E. ocellatus, B. arietans and N. nigricollis polyspecific antivenom reflects the snakebite therapeutic needs of Nigeria. Epidemiological studies had identified these three species as being of greatest medical importance in the country [11]–[14], [22] and were the basis for deciding upon a polyspecific antivenom – EchiTAb-Plus-ICP produced in Costa Rica. The decision to manufacture an E. ocellatus-specific antivenom (EchiTAbG produced in UK) was based on (i) the unusually high E. ocellatus-bite incidence rate in Nigeria [11], (ii) the 10–20% fatality rate of untreated victims of E. ocellatus envenoming [10], [23] and (iii) that the 20 minute blood clotting test [24], [25] can be reliably used to distinguish E. ocellatus envenoming from other West African venomous species – facilitating the physicians antivenom-selection decision. Another consideration was that the dose-efficacy of monospecific antivenoms is typically greater than polyspecific antivenoms (the curative dose of EchiTAbG is one vial and three vials for EchiTAb-Plus-ICP) and thus monospecific antivenoms offer a more cost effective treatment option – providing that the snake species can be identified using either distinct symptomatology or species identification. The incidence of antivenom-induced adverse effects of both the EchiTAb Study Group antivenoms is low [8], perhaps because both are produced under sterile GMP conditions employing caprylic acid to select IgG from sera/plasma. EchiTAbG is now a registered medicine in Nigeria (A6-0078) and EchiTAb-Plus-ICP is currently in the process of registration.
In addition to providing the Nigerian Federal Ministry of Health with the monospecific EchiTAbG and polyspecific EchiTAb-Plus-ICP antivenoms, the EchiTAb Study Group has also provided ‘best-practice’ training of hospital physicians in (i) the clinical use of these antivenoms, (ii) treatment of adverse effects and (iii) surgical treatment of the tissue-necrotic effects of local envenoming. Training was also given in the use of snake-identification and symptomology of envenoming to assist in making the most cost-effective and clinically-effective antivenom-selection decisions. The EchiTAb Study Group also provided ambulances to improve the speed of antivenom treatment of envenomed victims in an effort to improve the clinical outcome. The EchiTAb Study Group considered this multi-faceted approach to snakebite treatment as the most effective means of addressing the variant needs of snakebite victims in the region.
EchiTAbG is being provided free to patients in two hospitals in Nigeria (Kaltungo, Gombe State and Zamko, Plateau) where admitting 30 snakebite victims a day is not unusual, particularly in the biannual rain seasons. While it is the intent of the EchiTAb Study Group to expand the geographical delivery of its antivenoms, the current scarcity of effective antivenom in the region has resulted in victims undertaking long and expensive journeys to attend these hospitals, with some victims reportedly travelling from as far as Cameroon in the East and Niger in the North-West (personal observation, AN and ND). These observations indicate the paucity of effective and affordable antivenom in West Africa where snakebite, and particularly E. ocellatus, is a medical problem in most countries (Burkina Faso [26], Mali [27] Côte d'-Ivoire [28], Ghana [29], Benin [30], Niger [31] and Cameroon [32]). EchiTAbG therefore offers a therapeutic benefit in many countries other than Nigeria for which it was designed.
Since the East and North-East African Echis vipers are also a public health concern, our objective was to determine, at the pre-clinical level, whether the efficacy of EchiTAbG against E. ocellatus could be extended to these other Echis species. Ideally, our pre-clinical assays would have been conducted on venoms from all the African Echis species but we could not justify the ethical and financial costs of such a large number of mice. Therefore, based upon the most comprehensive taxonomic study of the genus Echis [16], we selected a single species from each of the four distinct species complexes; (i) E. p. leakeyi (Kenya) as a representative of the pyramidum complex which also includes E. leucogaster and E. p. pyramidum, (ii) E. coloratus (Egypt) as a representative of this species and E. omanensis, (iii) E. ocellatus as a representative of this species and E. jogeri and (iv) E. c. sochureki (United Arab Emirates) as a representative of the Asiatic carinatus complex. Our earlier work on the venom gland transcriptomes of these representative Echis species revealed considerable intra-generic differences in the number of isoforms comprising the main Echis toxin groups (snake venom metalloproteinases, phospholipases A2, serine proteases, C-type lectins) [15]. However, each pathogenic toxin family was represented in all the Echis species [15] – a result suggesting the possibility that EchiTAbG would have cross-Echis species efficacy.
However, we were also aware of previous clinical failures of the ‘heterologous’ administration of Echis species-specific antivenoms [33], [34]. Consequently, and in line with WHO recommendations [17], we performed here a series of immunological assays examining the immunological venom cross reactivity of ovine IgG raised against each representative Echis species. The results of these assays, which were designed to measure IgG titre, specificity and relative avidity to venoms in reduced and native states, indicated a very considerable degree of immunological cross-reactivity of each species-specific IgG antisera to each Echis venom. However, EchiTAbG was ineffective in neutralising the in vivo lethal effects of E. c. sochureki. This indicates that while these immunological tests provide informative and comprehensive immunological profiles of an antivenom, they can not yet replace pre-clinical in vivo testing to indicate the efficacy of an antivenom.
The most important result of the study was that EchiTAbG neutralises the lethal effects of venom from East and North-East African Echis species (E. p. leakeyi and E. coloratus) with an efficacy equal to that it shows against E. ocellatus from West Africa. A recent study reports a similar potential for the other EchiTAb Study Group antivenom, EchiTAb-Plus-ICP [20]. While these pre-clinical results require verification in human clinical trials, they do indicate a wider than intended application for both EchiTAbG and EchiTAb-Plus-ICP. We believe this is vitally important to the sustained delivery of these new antivenoms, developed to resolve a crisis in antivenom supply to Nigeria, because their production is now vulnerable to the same fiscal insecurities that caused the antivenom crisis a decade ago. A greater market, through geographical expansion, should permit the application of economies of scale that hopefully will, sequentially, reduce costs to the purchasing ministries of health, increase demand and improve the delivery of these urgently needed life-saving therapies.
|
10.1371/journal.pntd.0007245 | Molecular and antigenic characterization of Trypanosoma cruzi TolT proteins | TolT was originally described as a Trypanosoma cruzi molecule that accumulated on the trypomastigote flagellum bearing similarity to bacterial TolA colicins receptors. Preliminary biochemical studies indicated that TolT resolved in SDS-PAGE as ~3–5 different bands with sizes between 34 and 45 kDa, and that this heterogeneity could be ascribed to differences in polypeptide glycosylation. However, the recurrent identification of TolT-deduced peptides, and variations thereof, in trypomastigote proteomic surveys suggested an intrinsic TolT complexity, and prompted us to undertake a thorough reassessment of this antigen.
Genome mining exercises showed that TolT constitutes a larger-than-expected family of genes, with at least 12 polymorphic members in the T. cruzi CL Brener reference strain and homologs in different trypanosomes. According to structural features, TolT deduced proteins could be split into three robust groups, termed TolT-A, TolT-B, and TolT-C, all of them showing marginal sequence similarity to bacterial TolA proteins and canonical signatures of surface localization/membrane association, most of which were herein experimentally validated. Further biochemical and microscopy-based characterizations indicated that this grouping may have a functional correlate, as TolT-A, TolT-B and TolT-C molecules showed differences in their expression profile, sub-cellular distribution, post-translational modification(s) and antigenic structure. We finally used a recently developed fluorescence magnetic beads immunoassay to validate a recombinant protein spanning the central and mature region of a TolT-B deduced molecule for Chagas disease serodiagnosis.
This study unveiled an unexpected genetic and biochemical complexity within the TolT family, which could be exploited for the development of novel T. cruzi biomarkers with diagnostic/therapeutic applications.
| Chagas disease, caused by the protozoan Trypanosoma cruzi, is a lifelong and debilitating neglected illness of major significance in Latin America, for which no vaccine or adequate drugs are yet available. Identification of novel biomarkers able to transcend the current limits of diagnostic and/or therapeutic assessment methods hence surfaces as a main priority in Chagas disease applied research. In this framework, we herein undertook a thorough biochemical and antigenic characterization of T. cruzi TolT surface antigens. Our results unveil an unexpected complexity within this family, with at least 12 polymorphic TolT genes in the T. cruzi CL Brener reference strain genome. According to structural features, TolT deduced molecules could be split into three robust groups that show differences in their structural features, expression profile, sub-cellular distribution, post-translational modification(s) and antigenic structure. Overall, we show that TolT molecules are conspicuously expressed by both major mammal-dwelling stages of the parasite, and that they are differentially recognized by the immune system in Chagasic patients and in T. cruzi-infected mammals. Our findings are discussed in terms of the evolution and possible structural/functional roles of TolT molecules, as well as in terms of their applicability in Chagas disease serodiagnosis.
| With ~6 million people already infected and ~100 million at risk of infection, Chagas disease constitutes the most important parasitic disease and leading cause of infectious cardiomyopathy in Latin America [1]. Migratory trends of infected populations from endemic areas to Europe, North America, and the Western Pacific have also led to the spreading of this illness, which is now recognized as an emerging threat to global public health [2]. Trypanosoma cruzi, the etiological agent of Chagas disease, is a protozoan parasite that transitions between vertebrates (including humans) and blood-sucking triatomine vectors, with different developmental stages involved in each host. Within the insect, two major developmental forms can be observed: replicative epimastigotes in the midgut and metacyclic trypomastigotes in the hindgut [3]. The latter forms bring the infection into mammals when deposited on the skin or mucosa along with the excreta of the bug during blood-feed. Following cell invasion, parasites differentiate into rounded amastigote forms [3]. Along this transformation, the parasite undergoes remarkable physiological and morphological changes [4], including the complete disposal of its flagellum [5]. After several rounds of replication and just before disruption of the parasite-laden cell, amastigotes differentiate back into non-dividing and highly motile bloodstream trypomastigotes, which disseminate the infection within the mammal and may be eventually taken up by the triatomine during a bloodmeal.
Following a 30–60 day-long acute phase, strong and parasite-specific immunity is elicited in T. cruzi-infected people [6]. However, the parasite ability to quickly invade a wide variety of cell types, and the concurrent deployment of multiple elaborated evasion systems turn this immune response only partially effective [6,7]. In this context, the surface coat of bloodstream trypomastigotes fulfills a key dual purpose: to interact with host cell receptors prior to parasite internalization, and to provide protection against mammalian host-derived defense mechanisms [7]. This coat is composed of densely packed glycosylphosphatidyl inositol (GPI)-anchored glycoconjugates, which are usually coded by large, polymorphic, and developmentally regulated gene families [8]. In quantitative terms, the most important trypomastigote coat glycoproteins are mucins, Gp85/trans-sialidases (TS) and mucin-associated surface proteins (MASPs), all of which distribute over the entire parasite cell body, the flagellum, and even the flagellar pocket [8].
In 1990, a novel type of T. cruzi trypomastigote antigen was identified [9]. This antigen turned out to display homology to bacterial TolA proteins [10], and was accordingly designated TolT (TolA-like protein from T. cruzi). Instead of showing a broad surface distribution, TolT localized exclusively to the trypomastigote flagellum [9], apparently in the part of this structure in contact with the parasite body. Western blot analysis showed that TolT actually consisted of ~3–5 different molecules with sizes between 34 and 45 kDa [9]. All of them however collapsed to a single species upon treatment with endoglycosidase H, suggesting they corresponded to identical and/or highly similar polypeptides undergoing differential glycosylation. Subsequent immunological screenings led to the identification of three genes (termed TolT 1–3) in the T. cruzi Esmeraldo strain, which were arranged in tandem following a head-to-tail disposition [11]. Two of these genes, TolT1 and TolT2 were identical at the nucleotide level, and showed 98.9% sequence identity with respect to TolT3 [11]. The recurrent identification of peptides showing slight variations to TolT 1–3 deduced sequences in recent proteomic surveys however hinted at an underestimated TolT complexity [12–16].
The complete DNA sequence of the T. cruzi CL Brener reference clone was released in 2005, and it is represented by two datasets of contigs, each corresponding to one parental haplotype, which are referred to as ‘Esmeraldo-like’ or ‘non-Esmeraldo-like’ [17]. The CL Brener genome revealed a highly repetitive structure, which corresponded to a marked expansion of transposable elements, satellite DNA, and large multigene families including the above mentioned mucins, TS and MASPs, usually organized in tandems [17]. These features, together with CL Brener hybrid nature resulted in a highly fragmented genome assembly [17]. The CL Brener genome was subsequently followed by that of distinct parasite strains/clones [18,19], and by the genomes of phylogenetically related organisms [20–23]. More recently, third-generation sequencing technologies and bioinformatics allowed high-quality genome assembly of T. cruzi genomes [24–26]. This wealth of genetic information, along with the pressing need of novel T. cruzi biomarkers [27], prompted us to revisit TolT.
In this work, we show that TolT constitutes a larger-than-expected family of genes in T. cruzi, with at least 12 polymorphic members in the CL Brener reference strain and homologs in different trypanosomes. According to structural features, TolT deduced proteins could be split into three robust groups, all of them showing homology to bacterial TolA proteins, a biased amino acid composition, canonical signatures of surface localization and/or secretion, and trypomastigote flagellar surface localization. All of them were also found to be expressed in amastigote forms, with a TolT group-specific sub-cellular distribution. Thorough biochemical and immunological characterizations indicated that distinct TolT groups show additional differences in their post-translational modification(s) and antigenic structure. We finally used a recently developed fluorescence magnetic beads immunoassay to validate a recombinant TolT-B protein as an appealing reagent for Chagas disease serodiagnosis.
DNA sequences were compared using BLAST tool at the NCBI non-redundant DNA sequences databases at TriTrypDB (http://tritrypdb.org/tritrypdb/) and GeneDB (http://www.genedb.org/) using TolT1 (GeneBank accession number AF099099) sequence as query. Sequences showing an E value < 10−5 (~40% identity) were retrieved and their complete open reading frames (ORFs) were aligned using T-Coffee. After manual curation of the output, a preliminary phylogenetic tree was built using the Neighbor-Joining method. This tree allowed the definition of 3 robust groups, termed TolT-A, TolT-B and TolT-C. The complete ORF of one representative member of each group (TolT-A: TcCLB.508767.20, TolT-B: TcCLB.510433.20, TolT-C: TcCLB.506815.20) was further used to perform ‘iterative’ screenings, using the same conditions as stated above. The final phylogram (made upon DNA sequences) is the consensus tree of 1,000 bootstrap replicates and was graphically modified for presentation using iTOL. In addition, the deduced polypeptide of each representative member was used to search for similar sequences in the protein databases at TriTrypDB and GeneDB. Identification of signal peptides (SP) and GPI-anchoring signals was done using the online servers SignalP 4.0 and PredGPI, respectively. Post-translational modifications were predicted using NetPhos 3.1, NetNGlyc 1.0, NetOGlyc 4.0 and CSS-Palm 3.0. Homology to TolA was evaluated by independently querying the bacterial database of UniProt (http://www.uniprot.org/blast/) under default conditions with each predicted TolT product. Logos were generated using WebLogo (http://weblogo.berkeley.edu/logo.cgi).
CL Brener developmental forms were obtained and purified as described [28]. Briefly, epimastigotes were grown at 28°C in brain-heart tryptose medium supplemented with 10% (v/v) heat-inactivated fetal calf serum (FCS). Cell-derived trypomastigotes (henceforth trypomastigotes) and extracellular amastigotes were harvested from the supernatant of Vero cells (ATCC) grown at 37°C and 5% CO2 in minimal essential medium (MEM) supplemented with 10% (v/v) FCS, 0.292 g/L L-glutamine, 100 IU/mL Penicillin and 100 μg/mL Streptomycin (all from GIBCO Laboratories).
T. cruzi genomic DNA from CL Brener epimastigotes was purified as described [29]. Gene amplifications were obtained by PCR using 1–10 ng of DNA as template, recombinant Taq DNA Polymerase (Invitrogen), and the oligonucleotides detailed in S1 Table.
Different parasite developmental forms (4 x 108 of each) were homogenized in 1 mL of TRIzol reagent (Thermo), treated with DNAse I (Sigma), further partitioned in chloroform and centrifuged at 12,000 x g. The aqueous phase was recovered and RNA integrity was evaluated by 1% agarose gel electrophoresis. RNA was precipitated with 1 mL of 2-propanol. First strand cDNA was synthesized from total RNA samples using Superscript II reverse transcriptase (Life Technologies). Briefly, total RNA was resuspended in RNAse-free H2O and used at a final concentration of 0.25 μg/μL (3 μg of RNA per reaction), 10 μM oligo-dT-anchor primer (S1 Table), and 10 mM dNTPs in the reverse transcriptase (RT) First Strand Synthesis kit (Sigma). RT reactions were diluted appropriately, and used as templates for Real-Time quantitative PCR (RT-qPCR) reactions using Kapa Sybr Fast Universal Kit (Biosystems) and primers (S1 Table) were designed using PerlPrimer software v1.1.21. To verify that the SYBR Green dye detected only one PCR product, all the reactions were subjected to the heat dissociation protocol after the final cycle of the PCR and sequenced [30]. Samples were tested against T. cruzi Calmodulin and Glyceraldehyde 3-phosphate dehydrogenase (TcGAPDH) as reference genes for data normalization (S1 Table). Each experiment was performed in triplicate for two independently generated sets of cDNA templates.
PCR amplicons corresponding to TolT molecules were cloned into the pGEM-T easy vector (Promega), and used to transform DH5α cells (Invitrogen). DNA sequencing was carried out at Macrogen. These amplicons were then digested with the indicated restriction enzymes (S1 Table) and cloned into a tailored version of pGEX-2T (GE Healthcare) [31]. The glutathione S-transferase (GST)-fusion protein bearing the repetitive domain of T. cruzi antigen 1 (Ag1, also known as FRA, JL7 or H49 [27]) has been described [32].
Soluble fractions of E. coli strain BL21-Codon Plus (DE3)-RP cultures induced for 3 h at 28°C with 0.1 mM isopropyl ß-D-thiogalactopyranoside (Fermentas) were purified by gluthatione-Sepharose chromatography (GE Healthcare) and dialyzed against PBS [32,33]. GST, GST-Ag1, and GST-TolT samples were quantified by Bradford reagent (Bio-Rad) and purity was assessed by Coomasie blue-stained SDS-PAGE. Purified GST-TolT proteins were injected into animals as described [34] to generate specific antisera (S2 Table). Antiserum to T. cruzi TSSA has been described [35].
For IIF assays, trypomastigote forms (~106) were harvested, washed in PBS, adhered to poly-L-lysine (Sigma)-coated cover-slips and fixed for 30 min in PBS containing 4% (v/v) p-formaldehyde (PBS-PFA). Parasites were blocked for 30 min in 5% (w/v) Bovine Serum Albumin (Sigma) in PBS (PBS-A) supplemented with 0.5% (w/v) saponin (Sigma) for permeabilization, and probed with the indicated antiserum diluted in PBS-A. After extensive washings with PBS-A, secondary Alexa Fluor-conjugated antibodies (Molecular Probes) were added. Nuclei were stained with DAPI prior to montage in FluorSave reagent (CalBiochem). To evaluate the reactivity of T. cruzi intracellular stages 10,000 Vero cells were plated onto round coverslips, let stand overnight and infected with 1 x 106 CL Brener trypomastigotes per coverslip as described [35]. After 5 h, trypomastigotes were removed and cells were extensively washed and incubated in DMEM 10% (v/v) SFB. At 72–120 h post-infection, cells were washed with PBS, fixed and processed for IIF as above. Images were obtained with a Nikon Eclipse 80i epi-fluorescence microscope coupled to a DS-Qi1 CCD camera, and processed using ImageJ. For co-localization analyses, trypomastigotes fixed and adhered to cover-slips as above were blocked with PBS supplemented with 3% (w/v) BSA and 2% (v/v) horse serum (PBS-AHS) and incubated with rat anti-TolT-A and mouse anti-TolT-B sera (both diluted 1:100 in PBS-AHS). Following extensive washes, secondary antibodies were added for 1 h at 1:1,000 dilution in PBS-AHS. Images were obtained with an Olympus IX-81 microscope attached with a FV-1000 confocal module. Co-localization analysis was assessed using the Co-localization Finder plugin from ImageJ.
CL Brener trypomastigotes (2 x 108) were resuspended in 1 ml of ip buffer (PBS supplemented with 0.1% (v/v) Triton X-100 and a protease inhibitor cocktail (Sigma-Aldrich)). The extract (input fraction) was added with 6 μl of TolT-A antiserum raised in rat and incubated overnight at 4°C under constant agitation. This mixture was then added with 150 μl of protein G-Sepharose (GE Healthcare) previously washed with 500 μl of ip buffer, and incubated for 3 h at 4°C under constant agitation. The tube was centrifuged at 600 x g for 1 min. The supernatant (flow-through fraction) was removed, resin washed five times in 1 ml each of ip buffer, and stripped at 100°C for 5 min in SDS-PAGE loading buffer: 50 mM Tris-HCl pH6.8, 1% (w/v) SDS, 0.01% (w/v) bromophenol blue, 10% (v/v) glycerol and 50 mM dithiothreitol (DTT).
To achieve non-reducing conditions, the reducing reagent (DTT) was omitted from the SDS-PAGE loading buffer (see composition above). Samples were mixed with the indicated loading buffer and heated at 100°C for 3 min before being processed for Western blot as described [14]. Briefly, extracts from 1.5 x 107 parasites were resolved into SDS-PAGE (12.5% gels), transferred onto PVDF membranes (GE Healthcare), reacted with the indicated antiserum followed by HRP-conjugated secondary antibodies and developed using enhanced chemiluminescence (Thermo).
Pellets containing 1–5 x 108 trypomastigotes were homogenized in 2 mL of GPI buffer [10 mM Tris/HCl, pH 7.4, 150 mM NaCl, 2% (v/v) Triton X-114, 1 mM PMSF and 1% (v/v) protease inhibitor cocktail (Sigma)] on ice for 1 h and centrifuged at 8,800 x g for 10 min at 0°C as described [36]. The supernatant (S1) was stored at −20°C for 24 h. The pellet (P1) was washed with 1 ml of buffer A (10 mM Tris/HCl, pH 7.4, 150 mM NaCl, 0.06% (v/v) Triton X-114 and 1 mM PMSF) and stored. S1 was thawed and submitted to phase separation at 37°C for 10 min followed by centrifugation at 3,000 x g for 3 min at room temperature. The upper phase (S2) was collected and the detergent-rich phase was re-extracted with 1 ml of buffer A. The upper phase (S3) was collected, and the detergent rich phase was extracted with 1 ml of buffer A, homogenized, incubated for 30 min at 0°C and centrifuged at 18,000 x g for 10 min at 0°C. The pellet (P2) was washed with 1 ml of buffer A and stored, whereas the supernatant (S4) was submitted to a new phase separation. The upper phase (S5) was collected and the lower detergent-rich phase, enriched in GPI-anchored proteins, was taken as the GPI fraction (GPI). For PI-PLC treatment, 5 x 107 trypomastigotes were treated with or without 0.1U of recombinant PI-PLC from Bacillus cereus (Thermo) for 30 min at 4°C. Normal morphology and motility was controlled by microscopic observation before and after the incubation time. Following PI-PLC treatment, parasite pellets and supernatants were separated by centrifugation and fractions were processed for Western blot.
Trypomastigote pellets (3 x 108) were homogenized in 500 μL of ConA buffer [50 mM Tris/HCl, pH 7.4, 150 mM NaCl, 1% (v/v) Triton X-100, 0.1% (v/v) Nonidet P40, 0.1% (w/v) sodium deoxycholate, 5 mM Cl2Ca, 5 mM Cl2Mg, 5 mM Cl2Mn, 1% (v/v) inhibitor protease cocktail (Sigma), and 1 mM DTT] and processed as described [36]. After clarification, parasite extract was fractionated overnight at 4°C onto 100 μL of ConA–Sepharose (GE Healthcare), and retained glycoproteins eluted with 50 μL of SDS-PAGE loading buffer. Flow-through and ConA-bound fractions were analyzed by Western blot with different antisera. For N-glycosylation analysis, trypomastigotes were lysed in 1X Glycoprotein Denaturing Buffer (BioLabs), boiled for 10 min and extracts corresponding to 2.5 x 107 parasites were treated with 2,000–3,000U of Endoglycosidase H for 1 h following manufacturer’s procedures (BioLabs) and analyzed by Western blot.
Synthesis, screening and data analysis of high-density T. cruzi-derived peptide microarrays have been described [31,37].
GST-fusion proteins were dissolved in carbonate buffer (pH 9.6) at 10 μg/mL. Flat-bottomed 96-well Nunc-Immuno plates (Nunc, Roskilde, Denmark) were coated overnight at 4°C with 100 μL of the antigen solution, washed 3 times with PBS containing 0.05% (v/v) Tween 20 (PBS/T), and blocked for 1 h with 4% (w/v) skim milk in PBS/T at 37°C [33]. The plates were washed 3 times with PBS/T prior to the addition of serum samples diluted 1:500 in blocking buffer. Following incubation for 1 h at 37°C and washings with PBS/T, HRP-conjugated goat IgG to species-specific IgG (all from Sigma) diluted 1:5,000 in blocking buffer was added to the plates, and incubated at 37°C for 1 h. The plates were next washed and incubated with 100 μL of freshly prepared 0.5 mM 3,3’,5,5’-tetramethylbenzidine (Sigma) in citrate-phosphate buffer (pH 4.2) containing 0.2% (v/v) hydrogen peroxide. The reaction was stopped with 50 μL of 2 M sulfuric acid, and the absorbance at 450 nm was read. Each sample was assayed in triplicate, unless otherwise indicated.
Synthesis and coating of superparamagnetic microbeads with purified GST-fusion proteins were performed as described [32]. Functionalized beads (0.5 μg of each antigen in 20 μL of beads per reaction) were incubated with human serum samples (1:100 dilution), washed three times and bound antibodies were detected with DyLight 650-conjugated goat anti-human IgG antibodies (1:1,000 dilution, Jackson ImmunoResearch Laboratories). After washing three times, fluorescence was directly determined using a plate fluorometer (DTX880 Multimode Detector, Beckman Coulter). Incubation of coated beads with serum samples and conjugate antibodies were carried out for 5 min each, at room temperature without agitation. Sample and conjugate antibody dilutions, as well as washes between incubation steps were performed with PBS containing 0.2% (v/v) Tween 20. Washes were done using a magnetic rack without the need of centrifugation. Results of the FMBIA were expressed as percentage of reactivity of the mean fluorescence units of a standardized, positive control serum included in each assay run [32].
Serum samples from T. cruzi-infected subjects have been described [32,33,38], and were obtained from the Laboratorio de Enfermedad de Chagas, Hospital de Ninos "Dr. Ricardo Gutierrez". All procedures were approved by the research and teaching committee and the bioethics committee of this institution, and followed the Declaration of Helsinki Principles. Written informed consent was obtained from all individuals (or from their legal representatives), and all samples were decoded and de-identified before they were provided for research purposes. Chagasic patients were coursing the chronic stage of the disease without cardiac or gastrointestinal compromise. Serum samples were analyzed for T. cruzi-specific antibodies with the following commercially available kits: ELISA using total parasite homogenate (Wiener lab, Argentina) and indirect hemmaglutination assay (IHA, Polychaco, Argentina). The participating subjects currently live within the urban limits of Buenos Aires, an area free of vector-borne parasite transmission, though they (or their parents) were born and raised in endemic areas from Argentina or neighbor countries, where most likely acquired T. cruzi infection. Serum samples from healthy individuals that gave negative results in the aforementioned tests were obtained from different blood banks: Fundación Hemocentro Buenos Aires (Buenos Aires, Argentina), Hospital de Enfermedades Infecciosas ‘Dr. Francisco Javier Muñiz’ (Buenos Aires, Argentina), Hospital Italiano de Buenos Aires (Buenos Aires, Argentina) and Hospital Municipal ‘Dr. Diego E. Thompson’ (San Martín, Buenos Aires, Argentina).
The Institutional Review Board of UNSAM has evaluated the current project and considered that it complies with the Basic HHS Policy for Protection of Human Research Subjects requirements to be included in the ‘exemption 4', because it involved the use of de-coded and de-identified human serum samples obtained from sera repositories. The protocol of animal immunization followed in this study was approved by the Committee on the Ethics of Animal Experiments of the Universidad Nacional de San Martín (IACUC/CICUAE N° 08/2015), and all the procedures were carried out according with the recommendations of the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health.
Searches in the kinetoplastid genomic databases at TriTrypDB and GeneDB were carried out using the complete ORF of T. cruzi TolT1 as bait [11]. A total of 43 different sequences showing significant similarity (> 40% nucleotide identity) to TolT1 were retrieved; 38 sequences from different isolates of T. cruzi (CL Brener, Sylvio X-10, Dm28c, Esmeraldo) and 5 sequences from phylogenetically related protozoa such as the bat parasite T. cruzi marinkellei (4 sequences) and the human parasite Trypanosoma rangeli (1 sequence). Further searches were carried out using different TolT deduced protein sequences as query, which allowed for the identification of an additional TolT molecule in the reptilian parasite Trypanosoma grayi (DQ04_05721021). No TolT-related sequence was found in the genus Leishmania or in strict salivarian trypanosomes, i.e. trypanosomes that develop in the salivary glands of the insect vector such as Trypanosoma brucei, Trypanosoma congolense, Trypanosoma evansi and Trypanosoma vivax. A phylogenetic tree based on Neighbor-Joining method allowed the definition of 3 main groups of TolT-related sequences (termed TolT-A, -B and -C), which were supported by significant bootstrap values (Fig 1).
TolT-A sequences showed ~96–100% identity to TolT1 and included the 3 original sequences from the T. cruzi Esmeraldo strain [11], 5 full-length sequences from the T. cruzi CL Brener clone (TcCLB.508767.20, TcCLB.506617.10, TcCLB.506617.20, TcCLB.504157.130 and TcCLB.511109.10) and one partial sequence, also from CL Brener (TcCLB.506617.5) bearing C-terminal truncation (Fig 1). As originally described in the Esmeraldo strain [11], TolT-A sequences in CL Brener were arranged in a head-to-tail assembly of 3 members, which mapped to chromosome 23 (Fig 2). A fourth gene, TcCLB.504157.130, could not be linked to this cluster most likely due to CL Brener genome assembly deficiencies. Similar analyses performed on T. cruzi TCC, a hybrid strain from the same evolutionary lineage than CL Brener with a high-quality genome assembly [26], strongly supported the inclusion of TcCLB.504157.130 within the CL Brener TolT-A cluster found in chromosome 23. Moreover, they suggested the existence of additional TolT-A members in such CL Brener cluster that may have collapsed during genome assembly (Fig 2). TcCLB.511109.10 was annotated as a pseudogene due to a single nucleotide deletion that originated a frame shift mutation within the deduced N-terminal signal peptide (S1 Fig). Downstream from this mutation, however, TcCLB.511109.10 sequence was > 96% identical to the remaining TolT-A genes (S1 Fig), suggesting that it constitutes either a very recent pseudogene or, most likely, a sequencing error. Indeed, the orthologous TCC gene (tcc_111_157, Fig 2) was 100% identical to TcCLB.511109.10 except for this single nucleotide deletion. The TolT-A group also included 5 sequences from T. cruzi Dm28c, 1 from T. cruzi Silvio X-10 and 2 from T. cruzi marinkellei (Fig 1).
The TolT-B group displayed 76–90% identity to TolT1 and comprised 2 full-length sequences (TcCLB.510433.20 and TcCLB.504277.20) and 2 partial sequences bearing N-terminal truncations (TcCLB.504277.11 and TcCLB.508767.10) in the CL Brener genome (Fig 2). TolT-B sequences mapped to a single cluster present in chromosome 35 except for TcCLB.508767.10, which localized immediately downstream to TcCLB.508767.20 within the ‘Esmeraldo’ haplotype of the TolT-A genomic cluster (Fig 2). The TCC genome supported this particular disposition and, again, suggested the existence of additional TolT-B genes in the CL Brener chromosome 35 cluster (Fig 2). TolT-B also included 5 sequences from T. cruzi Sylvio X-10 and 7 sequences from T. cruzi Dm28c (Fig 1). Interestingly, TcCLB.510433.20 and TcCLB.504277.20 were almost identical sequences except for their predicted C-termini, where they become highly divergent at both nucleotide and amino acid sequences (S2 Fig). Because of these differences, the TcCLB.504277.20 deduced protein was predicted to lose the GPI-anchoring signal. The identification of orthologous genes in Sylvio X-10 (TcSYL_004013) and TCC (tcc_94_104) strongly argued that ‘GPI-less’ variants are not the result of genome assembly problems but rather genuinely diversified TolT molecules (S2 Fig).
The TolT-C group was composed by just two alleles of a single gene (TcCLB.504277.30 and TcCLB.506815.20) within the CL Brener genome (Fig 2). This locus mapped immediately downstream of the TolT-B genomic cluster present in chromosome 35; and a quite similar disposition was also observed in TCC (Fig 2). TcCLB.506815.20 was annotated as a MASP pseudogene. However, manual inspection allowed us to identify a partial TolT-B gene, a putative intergenic region and a complete TolT-C gene within this sequence (Fig 2 and S3 Fig). The TolT-C group also included sequences retrieved from T. cruzi Dm28c, T. cruzi Sylvio X-10 and T. cruzi marinkellei (Fig 1).
The main proteins encoded by TolT-A, TolT-B and TolT-C bore 310, 305/284 and 344 amino acids, respectively, with predicted molecular masses of ~30.9–37.1 kDa and predicted pI of 7.1–8.2. All of them displayed sequences that constitute canonical signatures of surface localization and/or secretion, including a predicted N-terminal signal peptide (SP) and a C-terminal GPI-anchoring signal (except for the ‘GPI-less’ TcCLB.504277.20) (Fig 3A). In addition, different algorithms predicted the existence of palmitoylation signals in TolT-A and TolT-B ‘canonical’ proteins (Fig 3A). The deduced TolT polypeptides were characterized by a high content of Ala (18.8 to 24.4%), Glu (9.0 to 11.9%), Leu (8.1 to 9.9%), Lys (6.3 to 10.4%) and Arg (4.9 to 8.1%), with these residues not evenly distributed throughout the entire proteins (Fig 3A). In addition, TolT-A and TolT-B products displayed a Cys-X7-Cys3 motif (where X means any residue) in their predicted SP (Fig 3A). Interestingly, this motif is also present in the SP of T. cruzi mammal-dwelling-expressed mucins (TcMUC, [28,39]), suggesting that TcMUC and TolT molecules may have a common origin or, more likely, that this motif may have been selected for the improved expression and/or post-translational processing of surface-associated molecules in such parasite forms (see below).
The predicted mature TolT molecules, i.e. upon processing of the SP and, if present, the GPI-anchoring signal displayed 17–41 potential sites for phosphorylation, 2–5 sites for N-glycosylation, 27–36 sites for O-glycosylation, and 2 strictly conserved Cys residues (Fig 3A). The only recognizable and unifying feature was the presence of two Ala-rich regions, which are predicted to fold into an αhelix-enriched secondary structure (Fig 3A). These regions, along with the interconnecting sequence, displayed structural and marginal sequence similarity to the central domain of bacterial TolA proteins. TolT-A and TolT-B products, in addition, bore a particular region towards the mature C-terminus, which is highly enriched in Arg and Glu residues (Fig 3A).
Amino acid alignments highlighted very few and minor intra-group polymorphisms among CL Brener TolT proteins, the only exception being the C-terminus of the ‘GPI-less’ variant encoded by TcCLB.504277.20 (S4 Fig). When comparing between groups, the genetic drift of TolT-C proteins and the high level of sequence conservation among TolT-A and TolT-B members were also evident (S4 Fig). Regarding the latter issue, it is worth noting that the amino acid identity value between TolT-A and ‘canonical’ TolT-B deduced products was not homogeneous. Rather, and as schematized in Fig 3B, this value was maximal along their predicted SP and C-terminal region but dropped down significantly towards their predicted mature N-terminal region (from residues ~40 to 180).
Samples of total RNA from different T. cruzi CL Brener developmental forms were purified and the relative expression of representative members of each TolT group evaluated by RT-qPCR. As shown in Fig 4, TolT-A transcripts were the most abundant, followed by TolT-B and TolT-C. Abundance of TolT mRNAs was significantly decreased in epimastigotes, particularly for TolT-C, for which transcript expression was barely detectable (Fig 4). When comparing between mammal-dwelling stages, i.e. trypomastigotes vs amastigotes, and somehow at odds with early steady-state transcriptome analyses [40], no significant differences were observed for either TolT group (Fig 4).
We next produced antisera against recombinant, GST-fusion proteins spanning sequences of different TolT molecules (S2 Table). To minimize the extent of possible cross-recognition between TolT-A and TolT-B molecules, specific antisera were raised against sequences from their most divergent, mature N-terminal regions (Fig 3). We also generated an antiserum to the TolT-A and TolT-B conserved C-terminal region (henceforth TolTA/B antiserum, S2 Table). As for the most divergent TolT-C group, we generated an antiserum towards a GST-fusion protein spanning most of its central region (S2 Table). These antisera were used to determine expression of TolT products along the parasite life cycle. IIF assays on T. cruzi-infected cells indicated that TolT-A and particularly TolT-B variants were significantly more expressed in trypomastigotes as compared to amastigotes (Fig 5A). The TolT-C antiserum, on the other hand, yielded similar signals in both parasite stages (Fig 5A). TolT-C signals, in turn, were weaker than those recorded for TolT-A and TolT-B, and should be thus compensated for presentation.
As originally reported for ‘TolT’ [9], TolT-A, TolT-B and TolT-C molecules localized to the part of the flagellum in contact with the parasite body. This was verified in both intracellular and extracellular trypomastigotes (Fig 5A and 5B). It is however worth noting that TolT-C molecules displayed an apparent continuous distribution whereas TolT-A and TolT-B proteins yielded a more punctuated labeling pattern (Fig 5B). Confocal images strongly supported TolT-A and TolT-B discontinuous distribution (Fig 5C). Most importantly, they revealed only minor co-localization between TolT-A and TolT-B signals (Fig 5C; Pearson’s R correlation coefficient of co-localization = -0.3).
In addition to the trypomastigote flagellum, some TolT-A antisera (but not all of them) also labeled a discrete region towards the posterior end, i.e. the parasite pole opposed to the site of emergence of the (vestigial) flagellum of amastigote forms (Fig 5A). This was consistent with Western blot data, showing the presence of TolT-A-reactive species, though in significantly lesser amounts as compared to trypomastigote ones, in amastigote forms (Fig 5D). Interestingly amastigote- and trypomastigote-expressed TolT-A molecules displayed different electrophoretic mobility, suggesting differences in their post-translational processing (Fig 5D). TolT-B and TolT-C antisera also labeled discrete internal region(s) of amastigote forms by IIF assays (Fig 5A). At variance with TolT-A, TolT-B signals accumulated towards the anterior pole of the amastigote whereas TolT-C products accumulated at both amastigote tips, and also at a compartment juxtaposed to the kinetoplast DNA (kDNA, Fig 5A). Amastigote-expressed TolT-B and TolT-C species could not be detected by Western blot (Fig 5D), most likely due to differences in the sensitivity of both methods. Neither antiserum displayed specific reactivity towards epimastigote stages (Fig 5D). Immunoprecipitation assays further demonstrated that the generated antisera were TolT group specific (Fig 5E).
To evaluate in silico predictions (Fig 3A), intact CL Brener trypomastigotes were firstly treated with PI-PLC, which specifically cleaves T. cruzi GPI anchors, and the supernatant and pellet fractions were analyzed by Western blot. As shown in Fig 6A, addition of PI-PLC caused the disappearance of TolT-A-, TolT-B- and TolT-C-reactive bands from the parasite pellets, and their concomitant appearance in the supernatant fractions. Interestingly, a minor fraction of TolT-A and TolT-C, but not TolT-B molecules could not be solubilized by PI-PLC (Fig 6A). These PI-PLC-resistant species may bear a different kind of acyl group or, alternatively, they may correspond to immature molecules that have not yet reached the parasite surface.
We next purified total GPI-anchored proteins from CL Brener trypomastigotes taking advantage of their preferential fractionation in Triton X-114. Aliquots corresponding to the different fractions were analyzed by Western blot. As shown in Fig 6B, TolT-A and TolT-B species were found in P1 (total parasite lysates), P2 (containing mostly membrane-associated molecules excluded from GPI- and sterol-rich micro-domains), and GPI fractions (containing mostly GPI-anchored proteins). Together with PI-PLC data, these findings confirmed that TolT molecules (at least a major fraction of them) are anchored to the trypomastigote plasma membrane through a GPI lipid motif. The finding of TolT-A and TolT-C PI-PLC-resistant species (Fig 6A), and of inter-group variations in the GPI/P2 abundance ratio (Fig 6B) unveiled certain heterogeneity in the way TolT molecules become associated to the parasite surface, and suggest that a minor fraction of them, particularly from TolT-A and TolT-C groups, use a different acyl group to achieve this issue.
We also analyzed whether the consensus N-glycosylation sites predicted in the deduced TolT products (Fig 3A) had an attached oligosaccharide in vivo. To that end, we initially carried out Western blot assays of CL Brener trypomastigote lysates upon fractionation on ConA lectin. As shown in Fig 6C, part of TolT-A-, TolT-B and TolT-C-reactive products were recovered in the ConA-bound fractions, indicating that at least a fraction of them indeed bear high-mannose type glycans. Trypomastigote lysates were next treated with endoglycosidase H. In line with original ‘TolT’ results [9], this treatment increased the electrophoretic mobility of TolT-A and TolT-B molecules (Fig 6D). Interestingly, the lower TolT-A-reactive band was not affected by endoglycosidase H treatment, suggesting that it may correspond to non-glycosylated species. Together with ConA-fractionation data, these findings suggest that at least part of the species from each TolT group undergo N-glycosylation in vivo.
To explore a possible structural role of Cys residues on the mature region of TolT molecules (Fig 3A), trypomastigote extracts were resolved in parallel on reducing and non-reducing SDS-PAGE and evaluated by Western blot. As shown, TolT-B molecules appeared to assemble into oligomers, which translated into the appearance of ~100 kDa species in non-reducing SDS-PAGE (Fig 6E). Considering the apparent molecular mass of TolT-B monomers (~35 kDa), the ~100 kDa species may likely correspond to trimers. In sharp contrast, solely monomeric species were observed for TolT-A and TolT-C molecules under non-reducing conditions (Fig 6E). TSSA, a well-characterized GPI-anchored molecule from the trypomastigote surface and devoid of Cys residues on its mature region [35,41] was used as control for these assays, and yielded solely monomeric species (Fig 6E). As expected, the ~100 kDa band was also revealed by the TolTA/B antiserum (Fig 6E). However, and since this antiserum recognized both TolTA (exclusively monomeric) and TolT-B (mostly trimeric), the ratio between trimers/monomers was shifted towards monomers (Fig 6E).
When expressed in a bacterial system, a TolT-B fusion molecule bearing both Cys residues 75 and 127 (GST-TolT-B 61–162) was able to assemble into multiple oligomeric forms (Fig 6F). GST-TolTB 97–162 (bearing solely Cys 127), on the other hand, yielded monomeric species and low amounts of a ~70 kDa species, likely a dimer, suggesting that the participation of both Cys residues is a pre-requisite in order to get trimeric and/or higher order aggregates. GST-TolT-B 61–103 (bearing solely Cys 75) was not able to dimerize under non-reducing conditions (Fig 6F), similar to GST-TolT-B 155–260 (bearing no Cys residue) used as negative control. These findings indicated that Cys 75 residue in GST-TolT-B 61–103 cannot engage into disulfide bond formation, likely due to structural constraints. Accordingly, in vitro incubation of GST-TolT-B 97–162 alone or in combination with GST-TolT-B 61–103 yielded the same profiling of high molecular mass species, i.e. solely the ~70 kDa band corresponding to the dimeric form of GST-TolT-B 97–162 (Fig 6F). Together, these findings strongly suggest i) that TolT-B, but neither TolT-A nor TolT-C molecules, are spontaneously assembled into trimers in vivo, on the surface of the trypomastigote; and ii) that trimeric TolT-B species are sustained by covalent inter-molecular disulfide bonds involving both Cys 127 and Cys 75.
TolT molecules were shown to elicit B-cell responses during T. cruzi infection in humans, and 3 full-length variants (2 ‘canonical’ TolT-B and 1 TolT-A member) were thereby included in a 16-recombinant protein-based, multiplexed assay for serodiagnosis of Chagas disease [42]. However, neither the fine antigenic structure of TolT molecules nor the impact of herein evidenced diversity on the TolT epitopic landscape was yet addressed. Three TolT-A (TcCLB.506617.10, TcCLB.504157.130 and TcCLB.508767.20), 2 ‘canonical’ TolT-B (TcCLB.510433.20 and TcCLB.504277.11), 1 ‘GPI-less’ TolT-B (TcCLB.504277.20), and 1 TolT-C (TcCLB.504277.30) sequences were firstly analyzed using high-density peptide microarrays [37]. Briefly, overlapping sequences with 1-amino acid residue offset were probed with IgG samples from different pools of chronic Chagasic sera. Arrays were processed firstly with normal human IgG to assess the background reactivity, and final antigenic profiles calculated by subtraction [37].
Oddly, TolT products displayed an overall very low reactivity when assessed by this approach. As shown in Fig 7A, solely TcCLB.504277.30 (TolT-C), TcCLB.504277.11 (‘canonical’ TolT-B) and TcCLB.508767.20 (TolT-A) could be annotated as ‘weak antigens’ in the context of the whole array (see also [37]). Moreover, each one of them actually displayed a single ‘antigenic peak’, i.e. a stretch of consecutive peptides yielding reactivities above the established cutoff (Fig 7A). The TolT-C antigenic peak encompassed the sequence 145AAVDADTAALAALLEVLQ, and was recognized by 2 out of 3 analyzed pools of sera (Fig 7A). In addition, TolT-C yielded a couple of negative antigenic peaks in one assay, suggesting that these sequences may be recognized by IgGs from healthy individuals (Fig 7A) [37]. TolT-A and TolT-B weak antigenic peaks encompassed the same sequence (TATRIQRTRPRVD), located on their C-terminal region, which was recognized by solely 1 analyzed pool of sera (Fig 7A). Large variations in the length of TcCLB.504277.11 and TcCLB.508767.20 deduced proteins, and hence in the relative location of the TATRIQRTRPRVD epitope within them, reflected that TcCLB.504277.11 was annotated as an N-terminal truncated protein (see Fig 2).
The antigenic profile of TolT proteins was next evaluated using recombinant proteins. To that end, a series of GST-fusion molecules were generated and purified from engineered bacteria. These molecules were used in ELISA tests to search for specific antibodies in serum samples from chronic Chagasic patients. As shown in Fig 7B, TolT-specific antibodies were indeed detected by this procedure in a fraction of assayed Chagasic sera and in none of the 19 normal serum samples (Fig 7B). Every TolT-positive serum recognized the C-terminal region conserved among TolT-A and TolT-B molecules (GST-TolT-B/A C-term protein, Fig 7B). Due to cloning/expression purposes, the TATRIQRTRPRVD sequence highlighted on the microarray assays (Fig 7A) was not included in the GST-TolT-B/A C-term protein, hence indicating the presence of additional B-cell epitope(s) in this molecule. In addition to the GST-TolT-B/A C-term protein, solely one sample reacted against the N-terminal region of TolT-A and two samples against TolT-C (Fig 7B). Western blot assays to a panel of TolT-B deletion mutants further stressed the significantly skewed recognition profile of anti-TolT antibodies elicited during T. cruzi infection in humans. As shown in Fig 7C, the recognition of four TolT-reactive Chagasic sera not included in our ELISA panel was also restricted to the conserved C-terminal region (residues 155–260, according to TolT-B). Quite similar results were obtained upon testing serum samples from T. cruzi-infected mice, rabbits and dogs by ELISA (Fig 7B).
The diagnostic performance of a GST-TolT-B fusion protein spanning most of its mature region (residues Q61 to R260, S2 Table) was evaluated by a recently developed FMBIA test, using 2 panels of serum samples obtained from non-infected individuals (n = 122) or from patients with chronic Chagas disease (n = 78). The latter was heterogeneous, and included people living and/or raised in different endemic areas from Argentina or neighbor countries, and hence most probably parasitized by different T. cruzi strains [43]. For comparison purposes, the same analysis was performed in parallel using GST-Ag1, a well-established Chagas disease serodiagnostic reagent [27]. For both antigens, a significant difference in the overall reactivity values between the negative and positive populations was obtained (P < 0.0001; Fig 8A). Most importanly, the area under the ROC curve for GST-TolT-B showed that this is a highly performant diagnostic classifier, with an area under the curve (AUC) value very similar to that of GST-Ag1 (0.9430; 95% CI, 0.9089–0.9772 for GST-TolT-B and 0.9742, 95% CI, 0.9477–1 for GST-Ag1) (Fig 8B). Plots of the diagnostic sensitivity and specificity of these assays as a function of the cut-off values (TG-ROCplot) indicated a cut-off value that concurrently optimizes both parameters of 17.4% for Ag1 and of 39.8% for GST-TolT-B. Overall, these results indicate that a recombinant, GST-fusion protein spanning most of the mature region of a ‘canonical’ TolT-B molecule, including its antigenic and conserved C-terminus, provides an appealing reagent for Chagas disease serodiagnosis.
Despite original proposals [9,11], we herein show that TolT constitutes a complex family of genes in kinetoplastids. According to phylogenomic data, it is apparently restricted to the branch of stercorarian trypanosomes, i.e. trypanosomes that develop in the hindgut of the insect vector, such as T. cruzi, T. cruzi marinkellei, T. grayi and T. rangeli. Even though T. rangeli is also able to develop in the triatomine salivary glands, taxonomic studies indicated it failed to group with other strict salivarian trypanosomes [44]. Solely in the T. cruzi CL Brener reference clone, we were able to find 12 TolT genes distributed in two chromosome clusters. Moreover, comparative analyses carried out in TCC, a closely related strain with a better-quality genome assembly and annotation [26], strongly suggest that such TolT gene dosage may be an underestimation. T. cruzi TolT genomic clusters comprise a discrete number of tandemly arranged genes from the same group, i.e. TolT-A genes in chromosome 23 cluster and TolT-B genes in chromosome 35 cluster. Within each cluster, TolT genes show minor polymorphisms among them. As extensively discussed, gene expansion by tandem duplication without further differentiation likely constitutes a kinetoplastid evolutionary strategy to increase protein yield in the absence of transcriptional regulation [45]. Indeed, mRNA and protein expression data roughly correlate with the estimated gene dosage for each TolT group. Interestingly, an additional and ‘different’ TolT gene is found immediately downstream of both TolT clusters, strongly suggesting that it evolved by mutation accumulation on a previously duplicated copy from another group. In the case of TolT-C genes, which display rather low similarity to the remaining T. cruzi TolT genes (~47% identity to any TolT-A gene), this particular genomic disposition was one pivotal criterion for their inclusion within the TolT family.
Overall, the most parsimonious hypothesis integrating these findings suggests that TolT emerged in an ancestor from trypanosomes, early after the divergence of the salivarian branch [46]. The acquisition of its embedded bacterial TolA-like motif may have occurred either by horizontal gene transfer, as previously proposed for other trypanosomatid molecules [47,48] or by convergent evolution. Whatever the case, the original TolT sequence likely underwent subsequent events of gene duplication followed (or not) by diversification (and eventually pseudogenization), thus leading to the formation of a rather complex family of genes. A rather similar evolutionary path, characterized by remarkable expansion and diversification, seems to have been followed by several T. cruzi gene families coding for surface molecules involved in the interaction with the mammalian and/or vector hosts [49].
TolT deduced proteins show a biased amino acid composition and molecular signatures of surface localization and/or secretion such as cleavable SP, glycosylation, and lipid modification. Based on topology predictions and experimental data, it could be inferred that, upon maturation in the secretory pathway, TolT molecules become tethered to the outer leaflet of the flagellar membrane via their C-termini. TolT membrane anchor most likely occurs post-translationally, by the addition of a GPI lipid moiety early upon their entry to the secretory pathway [50]. However, and as mentioned, our data is compatible with the possibility that a minor fraction of them, particularly those from TolT-A and TolT-C groups, use an alternative acyl group (i.e. palmitoyl) to achieve this issue. The only exception to the topological model proposed above would be TcCLB.504277.20, which loses its predicted GPI-anchoring signal due to focalized mutation accumulation. As shown for T. cruzi surface mucins, recombinant GPI-less variants (i.e. deletion mutants lacking the GPI-anchoring signal) accumulate in the endoplasmic reticulum, with only a minor fraction being ultimately released to the medium as anchorless products [51]. Further studies, currently underway, will be required to address TcCLB.504277.20 biochemical properties and sub-cellular distribution.
A unique and unifying feature of TolT molecules is that they bear similarity to the central region of TolA proteins. These are integral membrane molecules dedicated to maintain outer membrane stability in different bacteria such as Escherichia coli and Pseudomonas aeruginosa [10,52,53], being used also for the uptake of several filamentous phages and bacterial toxins called colicins [54]. TolA molecules are anchored to the inner bacterial membrane via their N-terminal domain, and display a central region made up essentially of alanine-rich stretches that show very stable helix conformation. Importantly, the TolA alanine-rich central region seems to play mainly a structural role, allowing the projection of the functional C-terminal region across the bacterial periplasm [54]. Though not experimentally proven, different algorithms predict that the TolA-like motif present in T. cruzi TolT molecules also encompasses a very long and unique αhelix, which likely adopts a rod-like structure on the trypomastigote membrane. In such a case, the acquisition of the TolA-like motif on the TolT central region may have been selected for as a crafty strategy to ensure the protrusion, and hence maximize the exposition, of the outermost (and variable) mature N-terminal region. Variations on this theme, leading to the projection of functional domains across the parasite glycocalix have been proposed for other T. cruzi surface molecules [8,55,56]. Most importantly, the overall topological model predicts a functional role for the TolT mature N-terminal regions; which according to their sub-cellular distribution may be related to the interaction between the flagellum and the trypomastigote body [57].
Original IIF assays carried out by Saborio et al showed that ‘TolT’ (presumably a TolT-A molecule according to our current classification) localized to the trypomastigote flagellum surface, apparently in the part of this structure in contact with the parasite body [9]. Here we assessed this sub-cellular localization for every TolT product, suggesting that despite their amino acid and biochemical variations all of them share the targeting signals responsible for this selective trafficking. Most interestingly, TolT-A and TolT-B products distribute in discrete foci along the surface of the trypomastigote flagellum. This is consistent with recent finding showing a biased lipid composition for the flagellar membrane of trypanosomatids, which apparently promotes the accumulation of GPI- and other kinds of acyl-anchored proteins into lipid-raft-like structures [58]. Moreover, the punctuate and non-overlapping pattern observed for TolT-A and TolT-B molecules builds upon our hypothesis of the trypomastigote membrane as a highly organized structure made up of multiple and discrete nanoscale domains bearing different protein composition [8,14]. Inter-molecular disulfide bonds leading to the formation of TolT-B homopolymers may also contribute to the formation/organization of these particular domains. Alternatively, disulfide bonds may have a rather classical structural role as a TolT-B quality control system in the endoplasmic reticulum [59] or, as shown for other protozoan surface antigens, in the undermining of the mammalian host immune response [60]. In this sense, it should be emphasized the lack of antibody response to TolT N-terminal regions. Our discoveries also raise the interesting possibility that the state of extracellular reduction-oxidation reactions on the vicinity of the trypomastigote flagellum may regulate the polymerization status of TolT-B molecules in vivo, which in turn may affect their yet-to-be-addressed functional and/or signaling properties [61–63]. In vivo studies using site-specific mutants and defined conditions should help to clarify these issues.
In addition to trypomastigotes, TolT products are also expressed on amastigote forms. Most interestingly, amastigote-expressed molecules accumulate on intracellular compartments. TolT-A (and TolT-C) molecules accumulate in discrete regions towards the posterior end of the amastigote, which may correspond to degradative organelles described in T. cruzi insect-dwelling forms [64]. TolT-B (and TolT-C) molecules, on the other hand, are likely retained in the flagellar pocket, the organelle that contributes to the traffic of GPI-anchored proteins between the Golgi complex and the plasma membrane [65] whereas TolT-C molecules accumulate in an undefined compartment juxtaposed to the kDNA. Whether these intracellularly-displayed molecules correspond to immature proteins en route to the amastigote membrane and/or to recycled species targeted for degradation remains to be addressed.
Our immunological characterizations support TolT molecules as targets of the immune response during T. cruzi infections. Indeed, by using a recently developed FMBIA test we show that a recombinant, GST-fusion protein spanning most of the mature region of a ‘canonical’ TolT-B molecule exhibits quite similar diagnostic performance than a well-established T. cruzi antigen, included in commercial serodiagnostic tests [27]. Antibody recognition seems to be focused towards peptides from the TolT-A/TolT-B conserved C-terminus, independently of the evaluated mammalian species. These findings may be attributed to intrinsic antigenic issues, i.e. biased distribution of B-cell epitopes, to the over-representation (in molar terms) of the TolT conserved C-terminal region or to the in vivo molecular shielding of TolT N-terminal regions by structural constraints and/or post-translational modifications. The latter hypothesis is however not consistent with IIF-based data showing that these regions are readily available to antibodies on both trypomastigote and amastigote forms. The lack of correlation between the peptide chip- and recombinant protein-based approaches, which is unique among other tested T. cruzi molecules [31,33,37], suggest that B-cell epitopes from Tol-T molecules are not strictly linear in nature.
In summary, we have shown that TolT constitutes a complex family of genes in T. cruzi, which could be split into three robust groups displaying differences in their structure, sub-cellular distribution, post-translational modification and antigenic composition. The fact that these molecules i) are abundantly expressed on T. cruzi developmental stages that dwell in the mammalian host; ii) provide robust and reliable reagents for the improvement/development of novel diagnostic and/or epidemiologic applications (see also [42]); and iii) were shown to constitute appealing vaccine candidates [11] indicate that they constitute excellent targets for molecular intervention in Chagas disease.
|
10.1371/journal.pgen.0030173 | Evolution of Nova-Dependent Splicing Regulation in the Brain | A large number of alternative exons are spliced with tissue-specific patterns, but little is known about how such patterns have evolved. Here, we study the conservation of the neuron-specific splicing factors Nova1 and Nova2 and of the alternatively spliced exons they regulate in mouse brain. Whereas Nova RNA binding domains are 94% identical across vertebrate species, Nova-dependent splicing silencer and enhancer elements (YCAY clusters) show much greater divergence, as less than 50% of mouse YCAY clusters are conserved at orthologous positions in the zebrafish genome. To study the relation between the evolution of tissue-specific splicing and YCAY clusters, we compared the brain-specific splicing of Nova-regulated exons in zebrafish, chicken, and mouse. The presence of YCAY clusters in lower vertebrates invariably predicted conservation of brain-specific splicing across species, whereas their absence in lower vertebrates correlated with a loss of alternative splicing. We hypothesize that evolution of Nova-regulated splicing in higher vertebrates proceeds mainly through changes in cis-acting elements, that tissue-specific splicing might in some cases evolve in a single step corresponding to evolution of a YCAY cluster, and that the conservation level of YCAY clusters relates to the functions encoded by the regulated RNAs.
| Alternative splicing generates different mRNA isoforms from a single gene and thus increases the number of proteins a cell can produce. This is particularly important in the brain, which possesses a number of brain-specific splicing factors. In this study, we have looked at evolution of brain-specific splicing regulation by one such factor, Nova. Previous studies have identified ∼100 alternative exons that are regulated by Nova in mouse brain. We find that the Nova protein sequence changed little during vertebrate evolution from fish to human, whereas the RNA targets themselves have evolved significantly. Interestingly, the presence of conserved Nova binding elements in an RNA transcript in most cases correlates with conservation of brain-specific splicing. In addition, the evolution of Nova-dependent splicing relates to the functions encoded by the target RNAs, such that Nova-regulated splicing of RNAs encoding core roles such as synaptic adhesion, ion channel, and cytoskeletal proteins is on average more conserved than splicing of the RNAs encoding regulatory roles, such as transmembrane receptor and signal transduction proteins.
| Alternative splicing is believed to be one of the major mechanisms by which proteome diversity is generated in multicelullar organisms [1,2]. Initial estimates that 5% of human genes are alternatively spliced [3] have recently been revised, such that it is now believed that 40%–60% of human genes are alternatively spliced [4,5]. A large number of these alternative exons are spliced with brain-specific patterns, as shown by microarray studies [6,7]. Such tissue-specific splicing patterns require interactions between defined cis-acting sequences present in the vicinity of the alternative exons and trans-acting regulatory factors [8,9]. It is still an open question to what extent mutations in cis-acting sequences or in genes encoding trans-acting regulatory factors have contributed to the evolution of splicing regulation [10,11], or to the development of neurologic disease [12].
Analysis of the evolutionary conservation of alternative exons can provide evidence of the regulation and functional significance of the exons. In addition, since alternatively spliced exons and flanking intronic regions are generally more conserved than constitutive exons, analysis of evolutionary conservation can be used to identify de novo alternative exons and regulatory sequences [13]. Interestingly, analysis of alternative exon conservation between human and mouse transcriptomes revealed a relationship to their inclusion level: major exons, included at over 50%, were found to be highly conserved (98%), whereas minor exons, included below 50%, were poorly conserved (26%) [14,15]. However, if a minor alternative exon shows a tissue-specific splicing pattern, its conservation between human and mouse rises to the level of major exons [16]. In general, exons with tissue-specific splicing patterns have increased conservation and frame preservation relative to other alternative exons [16,17].
Nova is a brain-specific splicing factor, first identified as an antigen in a neurologic disorder termed paraneoplastic opsoclonus-myoclonus ataxia [18]. Nova binds to clusters of tetranucleotide YCAY motifs (YCAY clusters), which are often present in the vicinity of Nova-regulated alternative exons [19–22]. When these are positioned within 200 nucleotides of the splice sites, their position predicts whether Nova functions to inhibit or enhance alternative exon inclusion, following the rules of an RNA map. So far, biochemical studies and analysis of mouse genomic sequence have found 76 different YCAY clusters located within Nova-regulated pre-mRNAs [6,20–24]. Fifty-four of these clusters are located in genes that are expressed both in brain and liver. This allowed us to analyze the splicing pattern of Nova-regulated exons from brain and liver mRNA of different vertebrates and relate it to the conservation of YCAY clusters. We found that a brain-specific splicing pattern was conserved across species in all (24/24) tested cases where YCAY clusters were conserved. We also found seven cases lacking the brain-specific splicing pattern, which correlated with the absence of detectable YCAY clusters in the corresponding pre-mRNAs. Interestingly, in all of these cases, we also observed a loss of alternative splicing (i.e., a single isoform detected in brain and liver). This suggests that some RNAs might have acquired brain-specific splicing regulation in a single step by addition of the YCAY cluster, without the need for preliminary mutations that would create an alternative exon.
Recently, an RNA map was constructed to relate Nova-dependent splicing regulation in brain to the position of its binding site (YCAY cluster) on the pre-mRNA [24]. This RNA map defines how YCAY clusters at five major positions on pre-mRNA direct the action of Nova on the splicing of neighboring alternative exons. At two such positions, YCAY clusters act as splicing enhancers and at three positions as splicing silencers. Seventy-six such YCAY clusters were found within Nova-regulated pre-mRNAs, providing a tool to relate the conservation of YCAY clusters to the evolution of brain-specific splicing. Nova proteins show high conservation between different vertebrate species, with the amino acid sequence of Nova1 and Nova2 KH domains 94% identical between zebrafish and mouse orthologues (Figures 1A and S1). Moreover, we found that both Nova1 and Nova2 bind to similar YCAY clusters in vivo, using the cross-linking and immunoprecipitation (CLIP) method ([22,25], unpublished data). We also confirmed that Nova expression is restricted to the brain in chicken and zebrafish (Figure 1B).
Given the highly conserved nature of the Nova proteins, we hypothesized that the main driving force in the evolution of Nova-dependent splicing regulation in vertebrate species might be changes in YCAY clusters within pre-mRNAs. We analyzed YCAY clusters in the pre-mRNAs containing 49 Nova-regulated exons that were previously identified by splicing microarray analysis of Nova knockout mouse brain [6]. Using the same YCAY cluster score algorithm as in the previous study [24], we found that 30% of YCAY clusters were conserved at orthologous zebrafish positions (Table S1A). This algorithm, which detects clusters with a maximum distance of three nucleotides between YCAY motifs [24], was written to be very stringent, since limiting the analysis to the highest affinity Nova-binding sites was important for precision in defining the original RNA map. However, previous analyses of minimal functional Nova binding site and of in vivo Nova–RNA interactions via the CLIP method [19–23] have shown that in some cases more dispersed YCAY clusters can be functional, even though they may bind Nova with a lower affinity. Indeed, manual analysis of sequence alignments (Figure S2) shows that the orthologous sequences often contain more dispersed YCAY clusters than the mouse genome. To be able to take these YCAY clusters into account, we modified the algorithm for the purpose of this study to detect clusters with distance of up to nine nucleotides between YCAY motifs. Analysis of gene sequences orthologous to mouse showed that 41% of the pre-mRNAs contain a conserved YCAY cluster in zebrafish, 66% in chicken, 82% in opossum, and 94% in the human genome (Figure 2; Table S1B).
To assess whether the presence of YCAY clusters can predict conservation of brain-specific splicing in vertebrates as an independent variable, we used RT-PCR to analyze brain-specific splicing of orthologous exons in zebrafish and chicken. Seventy-six different YCAY clusters located within the Nova-regulated pre-mRNAs were assigned a cluster score ≥0.6 by analysis of mouse genomic sequence using the modified YCAY cluster score algorithm [24] (Table S2 and Materials and Methods). Fifty-four (54/76) of these clusters are located in genes that are expressed both in brain and liver (Table S3). We designed PCR primers for these genes using cDNA and EST data or alignment of orthologous alternative exons and flanking constitutive exons. We were able to design PCR primers and obtain RT-PCR data for 25 chicken orthologous exons and 14 zebrafish orthologous exons.
RT-PCR analysis showed a striking correlation between the presence of YCAY clusters and the conservation of brain-specific splicing. Together, we tested 39 independent splicing events (25 in chicken and 14 in zebrafish). In 24 cases, the YCAY cluster was conserved (YCAY cluster score ≥0.6 in the orthologous sequence), and RT-PCR analysis showed at least 20% change in exon inclusion between brain and liver in 23/24 cases (|ΔI|>0.20, p<0.05, Figure 3; Table 1; Figure S2, where ΔI is used as defined previously [6]). In one case (chicken efna5 gene), the difference in exon inclusion between brain and liver was smaller, but still significant (ΔI = −0.09 and p = 0.008). Position of the YCAY cluster in all cases correctly predicted the direction of alternative splicing, following the rules of RNA map [24]. Seventeen splicing enhancers predicted higher exon inclusion in the brain, whereas seven splicing silencers predicted lower exon inclusion in the brain relative to the liver.
In 15 cases, YCAY clusters were not conserved (YCAY cluster score < 0.6 in the orthologous sequence). In 7/15 cases, RT-PCR detected no difference in splicing patterns of brain and liver RNA (ΔI < 0.01), and in one case (Efna5 in zebrafish) the difference in splicing pattern was the reverse of that seen in mouse. However, a significant difference in the splicing pattern of brain and liver RNA (p < 0.01) was detected in the remaining 7/15 cases, and in each case the change in splicing (inclusion or exclusion) was in the same direction as that seen in the mouse (Figures 3 and S2; Table 1). Thus, in 8/15 cases, the absence of YCAY cluster correlated with lack or reversal of a brain-specific splicing pattern, whereas a brain-specific splicing pattern was conserved in seven cases despite the absence of YCAY cluster in the pre-mRNA.
Taken together, we found that brain-specific splicing was conserved in all (24/24) pre-mRNAs containing YCAY clusters, but only in 47% (7/15) of pre-mRNAs lacking YCAY clusters. To address the statistical significance of this observation, we defined our problem as a comparison of these two binomial samples (first with 24/24 and second with 7/15 conservation frequency), since no previous study is available to define the expected frequency of brain-specific splicing conservation between mouse, chicken, and zebrafish. This comparison shows that the presence of a conserved YCAY cluster in pre-mRNA significantly increases the likelihood of brain-specific splicing being conserved (p = 0.0001, one-sided Fisher exact test).
Nova-dependent splicing silencers in neogenin (Neog, Figure 4A) and syntaxin binding protein 2 (Stxbp2/Munc18–2, Figure 4B) pre-mRNAs illustrate the correlations we observed between the presence of YCAY clusters and alternative splicing. In both cases, the YCAY cluster in mouse pre-mRNA spans both sides of the 3′ splice site, so that half of the cluster is intronic, and the other half exonic. In neogenin pre-mRNA, the cluster and brain-specific splicing pattern of exon exclusion are conserved between zebrafish, chicken, and mouse (Figure 4A). However, Stxbp2 pre-mRNA contains a cluster of over three YCAY motifs in human and mouse, but not in opossum, chicken, and zebrafish (Figure 4B). This lack of YCAY clusters correlates with the absence of exon 3 exclusion in the brain and liver of chicken or zebrafish. This represents a case where the absence of YCAY cluster correlates with a lack of alternative splicing.
Nova-dependent splicing enhancers in protein tyrosine kinase, receptor type F (Ptprf, Figure 4C), and amyloid beta (A4) precursor-like protein 2 (Aplp2, Figure 4D) pre-mRNAs demonstrate that brain-specific splicing is conserved at least as much as the YCAY clusters. In Ptprf pre-mRNA, the YCAY cluster and the brain-specific splicing patterns are conserved in all species (Figure 4C). In Aplp2 pre-mRNA, the YCAY cluster is conserved in chicken but not in zebrafish, whereas the brain-specific splicing pattern is conserved in all species (Figure 4D). This represents one of the seven cases where brain-specific splicing pattern is conserved in spite of the absence of detectable YCAY clusters within the pre-mRNAs (Figure 3).
A Nova-dependent splicing enhancer in suppression of tumorigenicity (St7, Figure 4E) pre-mRNA demonstrates a correlation between the emergence of the YCAY cluster and the alternative exon 12a. Vertebrate alignment of the St7 genomic sequence shows an absence of conservation in the region containing exon 12a in opossum, chicken, and zebrafish (Figure 4F). We detected no evidence of St7 exon 12a exon inclusion in brain and liver of chicken or zebrafish, exemplifying another case where a lack of YCAY clusters correlates with a lack of alternative splicing. This observation suggests that the YCAY cluster in St7 might have served to create the exon 12a; in the future, this hypothesis can be further analyzed using the genomic sequences of species evolutionarily intermediate to opossum and mouse.
Since Nova-regulated exons encode proteins with synaptic functions [6], analysis of their splicing conservation can provide insight into the evolution of functionally coherent coregulated networks (Figure 5). We find that most RNAs with conserved YCAY clusters encode adhesion and cytoskeletal scaffold proteins, ion channels, and signaling proteins. Many of these proteins have been implicated in neuronal development, or function at the synaptic junction [26–28]. In comparison, RNAs with YCAY clusters that are conserved only in mammals encode receptors (such as glycine (GlyRα2) and kainaite (GluR6) neurotransmitter receptors) and signaling proteins (neurochondrin, Lrp12, Gpr45). It is possible that RNAs with conserved YCAY clusters more often encode proteins that are indispensable for synaptic development and function, while the RNAs with YCAY clusters present only in mammals more often encode proteins that modulate synaptic function. This hypothesis cannot yet be statistically analyzed since not enough is known about the detailed synaptic functions of each Nova-regulated gene.
Previous studies have shown that Nova regulates a module of alternative exons that encode a functionally coherent set of proteins. How did this relationship between an RNA-binding protein and its set of RNA targets evolve? We approached this question by analyzing evolutionary changes in Nova proteins, YCAY clusters, and splicing of corresponding alternative exons in mouse, chicken, and zebrafish. In this paper, we developed a hypothesis that evolution of Nova-regulated splicing proceeds mainly through changes in cis-acting elements, while Nova proteins themselves show little evolutionary change.
This is the first study to analyze the conservation of tissue-specific splicing between species other than mouse and human. While bioinformatic analysis shows alternative exon conservation between mouse and human to be roughly 75% [14], we find ∼80% (31/39) conservation of brain-specific splicing patterns between evolutionarily much more diverse mouse, chicken, and zebrafish species. This estimate might be a bit higher than the average conservation rate, since we were not able to define orthologous positions in chicken and zebrafish genomes to analyze splicing of a subset of Nova-regulated exons, due to the lower level of their genomic sequence conservation. However, the finding agrees with previous comparisons of splicing in mouse and human tissues showing that exons with tissue-specific splicing are more conserved than the rest of the alternative exons [16,17]. It will be interesting in the future to study the conservation of tissue-specific exons between such diverse species as fish and human in order to test more generally if exons with brain-specific pattern, in addition to those regulated by Nova, have particularly high conservation rates.
This study modified the original YCAY cluster algorithm by relaxing the requirement for the maximal distance between two YCAY motifs from three to nine nucleotides. The greater ability of the modified algorithm to detect conserved YCAY clusters is evident by analyzing the 31 cases where RT-PCR analysis in this study detected conservation of brain-specific splicing in chicken or zebrafish. Whereas the original more-stringent algorithm detects YCAY clusters in only 15 cases, the modified algorithm detects a conserved cluster in 24/31 cases (Table S1). A shortcoming of our current algorithm is that it analyzes the YCAY motifs as being separated by a linear RNA molecule, even though in reality YCAY motifs that appear dispersed on a linear RNA may be in relatively close proximity once the RNA assumes its in vivo structure. This may be one reason why the original study did not detect YCAY clusters in all the pre-mRNAs regulated by Nova in mouse brain, and why the current study didn't detect conserved YCAY clusters in all cases where brain-specific splicing pattern was conserved.
Of 15 pre-mRNAs in which we didn't detect a conserved YCAY cluster, seven had conserved brain-specific splicing patterns. In these instances, our algorithm may be failing to detect some types of conserved Nova-binding sites; in particular, those where RNA structure might bring multiple YCAY motifs in proximity. Alternatively, many alternative exons are regulated by multiple cis-acting elements [29]. For instance, splicing of alternative exon 19 in NMDA receptor1 (NR1) is regulated by Nova and several other factors including Napor, hnRNP A1, hnRNP H, and CaM kinase IV [30–32]. Thus, it is reasonable to believe that during a long evolutionary time, the importance of different cis-acting sites for the tissue-specific splicing pattern varies, allowing for other factors to compensate for the absence of YCAY clusters in these eight cases. We note that in our previous study [24], we demonstrated that YCAY clusters located within 200 nucleotides of splice sites predict Nova-dependent splicing regulation, but we do not know whether these clusters are the only ones required for Nova action. It is possible that in some cases Nova may be able to regulate its target exons via additional YCAY clusters located in introns further than 200 nucleotides from the splice sites, which would have been missed in this study.
We found seven cases of pre-mRNAs lacking brain-specific splicing pattern and also lacking YCAY clusters. Interestingly, in all of these cases, rather than just lacking tissue-specific splicing, we detected no evidence of alternative splicing at all in brain and liver. This observation contrasts with the currently prevailing model, which proposes that tissue-specific splicing generally evolves from the regulation of preexisting alternative exons, so that mutations in consensus 5′ and 3′ splice sites and enhancer sequences would precede subsequent mutations that lead to tissue-specific splicing [10,33]. We observed no case where an alternative exon would subsequently evolve a tissue-specific splicing pattern, although our sample size was limited. Our data suggest that tissue-specific splicing might have evolved directly, at least in some cases. This model of direct evolution of Nova-dependent alternative splicing is also supported by the observation that splice site consensus scores and exonic enhancer density in Nova-regulated exons are on average similar to constitutive exons [24].
One of the exons that obtained a Nova-dependent brain-specific splicing pattern in more recent evolution is Munc18–2 (Stxbp2) exon 3. The role of Munc18–2 in the brain is unexplored since an early study has termed it a non-neuronal Munc18–1 (Stxbp1/Sec1) homologue [34]. However, given that Nova regulates brain-specific splicing of its exon 3 in higher mammals, Munc18–2 might play a similar role in regulating neurotransmitter release as Munc18–1 [35], in a way that would contribute to brain function of higher mammals.
There is an ongoing debate regarding the functional significance of splice variants, particularly those encoded by minor exons [13,33,36–39]. Our study showed a high conservation of Nova-regulated alternative exons regardless of whether they are minor or major exons as assessed by bioinformatic analysis of EST and cDNA sequences (Tables S4 and S5). Furthermore, the functional significance of Nova-regulated exons is corroborated by conservation of reading frame; 70% (54/77) of internal cassette exons had a length multiple of three in all species in which they were present (Table S4), agreeing with previous reports that conserved alternative exons show significant increase in frame preservation [16,17,33,40].
The composite Nova binding site allows variation in the number and density of YCAY motifs to fine tune the affinity of the RNA for Nova [18,19,21]. Furthermore, the quantitative outcome of Nova binding might be refined by changing the position of YCAY clusters within the RNA [24]. However, we observed no case where YCAY cluster would change in position in a way that would qualitatively change the outcome of Nova binding (i.e., change from a splicing enhancer to silencer). We hypothesize that these features of Nova–RNA binding might allow selection in higher vertebrates to act gradually via changing the Nova binding site, rather than Nova itself, similarly to other cases where cis-acting elements were suggested as the main force for phenotypic evolution [11]. Furthermore, the dynamic nature of the cis-acting elements could help fine tune the functional coherence of the whole network of RNAs regulated by Nova (Figure 5), which were found in previous studies to primarily encode synaptic proteins [6,22,24]. In contrast, tight evolutionary fixation of Nova is essential to preserve regulation of an array of brain-specific splicing events. Interestingly, evolutionary fixation in the vertebrate lineage does not apply to all heterogeneous nuclear ribonucleoproteins, some of which were found to vary in domain structure and even in their presence/absence between mammals and fish [41,42].
The finding that a trans-acting factor such as Nova, which regulates multiple genes, is more conserved than the cis-acting sites that regulate individual genes might be anticipated, as it agrees with previous studies showing that transcriptional and micro-RNA regulators evolve more slowly than their target binding sites [11,43,44]. However, the conservation of the cis-acting sites observed in this study is surprisingly high, with 94% conservation of YCAY clusters between mouse and human. In contrast, biochemical analysis of cis-acting sites binding tissue-specific transcriptional regulators has found that 19%–59% of sites are conserved between mouse and human [45]. We have similarly observed that when using a biochemical approach (CLIP) [22], the majority of Nova-binding sites isolated from mouse brain are not conserved in the human genome ([22], unpublished observation). How can we reconcile the high conservation of the functional Nova binding sites analyzed in the current study with the much lower conservation of sites isolated via biochemical studies? We hypothesize that in addition to the highly conserved functional sites that were analyzed in this study, Nova might also bind less conserved sites, which could contribute to ability of the organism to evolve. To explore this question further in the future, biochemical analysis of Nova–RNA binding in brains of different species would need to be related to the alternative exons regulated by Nova in each species.
Taken together, the current work finds that brain-specific splicing patterns of Nova-regulated exons are highly conserved and are related to conservation of YCAY clusters in the pre-mRNAs. The presence of YCAY clusters is 100% predictive for the brain-specific alternative splicing of exons, whereas their absence significantly decreases the ratio of conserved splicing to 47%. The data also agree with the hypothesis that alternative splicing might in some cases evolve via addition of cis-acting sites that bind tissue-specific splicing factors such as Nova. Moreover, it is shown that the genes encoding adhesion and cytoskeletal proteins generally display deeper evolutionary fixation of YCAY clusters than the genes encoding receptors and signaling molecules, suggesting evolution of Nova function through evolution of target RNA sequences.
We searched for human (Homo sapiens), opossum (Monodelphis domestica), chicken (Gallus gallus), frog (Xenopus tropicalis), and zebrafish (Danio rerio) genomic orthologous sequences with the Liftover tool on University of California Santa Cruz (UCSC) Genome Browser [46] using genomic data from March 2006 for human, February 2006 for mouse, January 2006 for opossum, May 2006 for chicken, August 2006 for frog, and March 2006 for zebrafish. We used the UCSC Table Browser [47] (http://genome.ucsc.edu/cgi-bin/hgTables) to download sequence flanking the splice sites; we analyzed 80 nucleotides of exonic sequence upstream of each 5′ splice site, 200 nucleotides of intronic sequence downstream of each 5′ splice site, 200 nucleotides of intronic sequence upstream of each 3′ splice site, and 80 nucleotides of exonic sequence downstream of each 3′ splice site. Sequences were further verified for proper alignment using MacVector ClustalW and T-Coffe multiple alignment tool, and cases where multiple possible alignments were present in the genome due to multiple paralogous matches were discarded.
We adjusted our original YCAY cluster score algorithm [24] for the purpose of this evolutionary study. Whereas the original paper [24] required stringent filtering of false positives due to analysis made on genomic scale, the current algorithm allowed for larger toleration in distance between YCAY motifs [22]. The majority of YCAY motifs are located in introns, which display fast mutation rates between such distant species as fish and mouse, and therefore the current algorithm was adapted to tolerate such mutations in a way that still detects the core feature of Nova binding site, i.e., three proximal YCAY motifs. In addition, the definition of the boundaries between the areas where the clusters act as silencers or enhancers was the same as in the previous study [24]. The YCAY cluster score was calculated by searching for the first YCAY motif, and then giving it a score relative to the pattern of YCAY motifs that followed it:
if YCAY[N>23]YCAY then s = 0
if YCAY[19<N<24]YCAY[[N<4]Y]CAY then s = 2
if YCAY[9<N<20]YCAY[3<N<20]YCAY then s = 2
if YCAY[9<N<20]YCAY[[N<4]Y]CAY then s = 4
if YCAY[3<N<10]YCAY[9<N<20]YCAY then s = 2
if YCAY[3<N<10]YCAY[3<N<10]YCAY then s = 4
if YCAY[3<N<10]YCAY[[N<4]Y]CAY then s = 6
if YCAY[[N<4]Y]CAY[19<N<24]YCAY then s = 2
if YCAY[[N<4]Y]CAY[9<N<20]YCAY then s = 4
if YCAY[[N<4]Y]CAY[3<N<10]YCAY then s = 6
if YCAY[[N<4]Y]CAY[[N<4]Y]CAY then s = 8
After the score is given to the first YCAY, the analysis moves on to the next YCAY, and so on in an iterative way until the end of the 45-nucleotide sequence window is reached. At that point, the score is calculated:
S = log(s1 + s2 + … + sn), where n is the number of YCAY motifs in the 45-nucleotide sequence window.
In order to predict the direction of Nova-dependent splicing regulation based on YCAY cluster position, net YCAY cluster score was calculated as described previously [24] using the following formula:
Net conserved S = 1/2(MAX(NISE1, NISE2, NISE3, SUM(NISE2, NISE3)*2/3) − MAX(NISS1, NISS2, NESE))
where NISE1 is nova intronic splicing enhancer 1, NISE2 is nova intronic splicing enhancer 2, NISE3 is nova intronic splicing enhancer 3, NISS1 is nova intronic splicing silencer 1, NISS2 is nova intronic splicing silencer 2, NESE is nova exonic splicing enhancer.
Purified RNA from brain or liver of chicken (G. gallus) and zebrafish (D. rerio) was reverse transcribed using random hexamers, and cDNA products were amplified using Taq PCR Master Mix Kit (Qiagen) with 40 pmol of each primer and 0.5 pmol of one γ-32P-ATP-labeled primer at Tm = 55 °C. The primers used are listed in Table S6. PCR products were resolved on polyacrylamide gel electrophoresis and confirmed by size and sequencing.
Brain and liver from chicken (G. gallus) and zebrafish (D. rerio) were homogenized and protein concentration was determined with Bradford dye assay (BioRad). Proteins were separated by SDS/PAGE, transferred to nitrocellulose, and probed with rabbit polyclonal anti-Nova antibody that was made by immunization with full-length Nova1 [20] or rabbit polyclonal eIF3a (Santa Cruz). Blots were developed with horseradish peroxidase–linked secondary antibodies and enhanced chemiluminiescence (Amersham).
The National Center for Biotechnology Information (NCBI) Entrez Gene (http://www.ncbi.nlm.nih.gov) accession numbers of the mouse genes discussed in this paper are: Neo, 4756; Stxbp2/Munc18–2, 81804; Ptprf, 5792; Aplp2, 11804; St7, 64213; Nova1, 18134; Nova2, 384569; GlyRα2, 237213; GluR6, 54257; neurochondrin, 26562; Lrp12, 239393; Gpr45, 93690; NR1, 14810; Napor, 14007; hnRNP A1, 15382; hnRNP H, 59013; and CaM kinase IV, 12326. |
10.1371/journal.ppat.1002836 | Cedar Virus: A Novel Henipavirus Isolated from Australian Bats | The genus Henipavirus in the family Paramyxoviridae contains two viruses, Hendra virus (HeV) and Nipah virus (NiV) for which pteropid bats act as the main natural reservoir. Each virus also causes serious and commonly lethal infection of people as well as various species of domestic animals, however little is known about the associated mechanisms of pathogenesis. Here, we report the isolation and characterization of a new paramyxovirus from pteropid bats, Cedar virus (CedPV), which shares significant features with the known henipaviruses. The genome size (18,162 nt) and organization of CedPV is very similar to that of HeV and NiV; its nucleocapsid protein displays antigenic cross-reactivity with henipaviruses; and it uses the same receptor molecule (ephrin- B2) for entry during infection. Preliminary challenge studies with CedPV in ferrets and guinea pigs, both susceptible to infection and disease with known henipaviruses, confirmed virus replication and production of neutralizing antibodies although clinical disease was not observed. In this context, it is interesting to note that the major genetic difference between CedPV and HeV or NiV lies within the coding strategy of the P gene, which is known to play an important role in evading the host innate immune system. Unlike HeV, NiV, and almost all known paramyxoviruses, the CedPV P gene lacks both RNA editing and also the coding capacity for the highly conserved V protein. Preliminary study indicated that CedPV infection of human cells induces a more robust IFN-β response than HeV.
| Hendra and Nipah viruses are 2 highly pathogenic paramyxoviruses that have emerged from bats within the last two decades. Both are capable of causing fatal disease in both humans and many mammal species. Serological and molecular evidence for henipa-like viruses have been reported from numerous locations including Asia and Africa, however, until now no successful isolation of these viruses have been reported. This paper reports the isolation of a novel paramyxovirus, named Cedar virus, from fruit bats in Australia. Full genome sequencing of this virus suggests a close relationship with the henipaviruses. Antibodies to Cedar virus were shown to cross react with, but not cross neutralize Hendra or Nipah virus. Despite this close relationship, when Cedar virus was tested in experimental challenge models in ferrets and guinea pigs, we identified virus replication and generation of neutralizing antibodies, but no clinical disease was observed. As such, this virus provides a useful reference for future reverse genetics experiments to determine the molecular basis of the pathogenicity of the henipaviruses.
| Henipaviruses were first discovered in the 1990s following investigation of serious disease outbreaks in horses, pigs and humans in Australia and Malaysia [1], [2] and comprise the only known Biosafety Level 4 (BSL4) agents in the family Paramyxoviridae [3]. Depending upon the geographic locations of outbreaks, and the virus and animal species involved, case mortality is between 40% to 100% in both humans and animals [4], [5], making them one of the most deadly group of viruses known to infect humans. The genus Henipavirus in the subfamily Paramyxovirinae currently contains two members, Hendra virus (HeV) and Nipah virus (NiV) [6]. Fruit bats in the genus Pteropus, commonly known as flying foxes, have been identified as the main natural reservoir of both viruses although serological evidence suggests that henipaviruses also circulate in non-pteropid bats [7], [8], [9], [10].
The discovery of henipaviruses had a significant impact on our understanding of genetic diversity, virus evolution and host range of paramyxoviruses. Paramyxoviruses, such as measles virus and canine distemper virus, were traditionally considered to have a narrow host range and to be genetically stable with a close to uniform genome size shared by all members of Paramyxovirinae [3]. Henipaviruses shifted this paradigm on both counts having a much wider host range and a significantly larger genome [6]. Identification of bats as the natural reservoir of henipaviruses also played an important role in significantly increasing international scientific attention on bats as an important reservoir of zoonotic viruses, including Ebola, Marburg, SARS and Melaka viruses [11], [12], [13], [14].
Since the discovery of the first henipavirus in 1994, much progress has been made in henipavirus research, from identification of functional cellular receptors to the development of novel diagnostics, vaccine and therapeutics [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25]. By contrast, there is little understanding of the pathogenesis of these highly lethal viruses. This is due in part to the requirement of a high security BSL4 facility for any live infection studies and in part to the limited range of research tools and reagents for the current small animal models. Research into the mechanisms of henipavirus pathogenesis is also hampered by the lack of related, but non-pathogenic or less pathogenic viruses, thus preventing targeted comparative pathogenetic studies.
Early serological investigations in Australia and more recent studies in other regions (e.g., China) indicated the presence of cross-reactive, but not cross-neutralizing, antibodies to henipaviruses in bats of different species [8]. These findings were further supported by the detection of henipavirus-like genomic sequences in African bats [26]. Discovery and isolation of these related viruses will be highly important to our further understanding of henipavirus evolution, mechanism of cross-species transmission, and pathogenesis in different animal species.
Here we report the isolation and characterization of a new bat henipavirus which, based on preliminary infection studies, is non-pathogenic in two of the small animal infection models currently used in henipavirus research. We believe that this new virus will provide a powerful tool to facilitate our future study into different aspects of henipaviruses, especially in the less advanced area of pathogenesis.
As part of our on-going field studies on HeV genetic diversity and infection dynamics in the Queensland flying fox populations, urine samples were collected on a regular basis for PCR and virus isolation. Since the establishment of the Pteropus alecto primary cell lines in our group [27], we have intensified our effort to isolate live virus from these urine samples by routinely inoculating separate primary cell lines derived from kidney, spleen, brain, and placenta, as well as Vero cells. Syncytial CPE was observed in kidney cell (PaKi) monolayers 5 days post inoculation (dpi) with two different urine samples (Fig. S1) collected in September 2009 from a flying fox colony in Cedar Grove, South East Queensland (see Fig. S2 for map location). No CPE was observed in any of the four other cell lines. Supernatant harvested 6 dpi was used to inoculate fresh PaKi cell monolayers. After two passages in PaKi cells, the virus was able to infect and cause CPE in Vero cells. However, the CPE morphology of CedPV infection in Vero cells was different from that of HeV infection. Further analysis using HeV-specific PCR primers indicated that the new bat virus was not an isolate of HeV.
Considering the formation of syncytial CPE by this new virus and the previous success in isolating paramyxoviruses from bat urine [28], [29], [30], paramyxovirus family-specific and genus-specific primers were used to determine whether this new virus was a member of the family Paramyxoviridae. Positive PCR fragments of the expected sizes were obtained from the Paramyxovirinae and Respirovirus/Morbillivirus/Henipavirus primer sets developed by Tong et al [31]. Sequencing of the PCR products indicated that it was a new paramyxovirus most closely related to HeV and NiV. Based on these preliminary data, the virus was named Cedar virus (CedPV) after the location of the bat colony sampled.
Full length genome sequence was determined by a combination of three different approaches, random deep sequencing using 454 technology, sequencing of PCR products obtained using degenerate primers designed based on known henipaviruses, and RACE to determine the precise genome terminal sequences. As shown in Fig. 1, the genome of CedPV is 18,162 nt in length most similar to that of HeV in the family. The full genome sequence has been deposited to GenBank (Accession No. JQ001776). The genome size is a multiple of 6, hence abiding by the Rule-of-Six observed for all known members of the subfamily Paramyxovirinae [3]. It has a 3-nt intergenic sequence of CTT absolutely conserved at all seven positions and highly conserved gene start and stop signals similar to those present in HeV and NiV (Fig. S3). Also similar to the HeV genome is the presence of relatively large non-coding regions in the CedPV genome (Fig. 1 and Table 1). The overall protein-coding capacity of the CedPV genome is 87.41% which is significantly lower than the average of 92.00% for other family members but higher than HeV at 82.12%. As the genome size of CedPV and HeV is very similar, the increased coding capacity of CedPV is attributed to an increase in protein sizes for five of the six major proteins, with the L protein being 257-aa larger (Table 1). At 2,501 aa, the CedPV L protein is the largest, not only in the family Paramyxoviridae but also for all known viruses in the order Mononegavirale.
Phylogenetic analysis based on the full length genome sequence and the deduced amino acid sequences of each structural protein confirmed the initial observation that CedPV is most closely related to henipaviruses in the family. A phylogenetic tree based on the deduced sequences of the nucleocapsid protein (N) is presented in Fig. 2. Phylogenetic tree based on whole genome sequences gave similar results (Fig. S4). CedPV is more closely related to HeV and NiV than henipavirus-like sequences detected in African bats [26], [32] as shown in a phylogenetic tree based on the only sequences available of a 550-nt L gene fragment (Fig. S5).
First discovered for the parainfluenza virus 5 (PIV5, previously known as simian virus 5), almost all members of Paramyxovirinae have a P gene which produces multiple proteins through an RNA editing mechanism by addition of non-templated G residues leading to production of N-terminal co-linear proteins from different reading frames downstream from the editing site [3], [33]. These multiple gene products are known to play a key role in antagonizing the innate response of susceptible hosts [3]. A search of CedPV for open reading frames (ORF) in the P gene revealed a 737-aa P protein and a 177-aa C protein, but failed to find the highly conserved, cysteine-rich V ORF present in most other paramyxoviruses. The RNA editing site with the sequence of AAAAGGG, which is absolutely conserved in all known HeV and NiV isolates discovered to date, is also missing from the CedPV P gene sequence. To further verify that there are no multiple mRNAs produced from the CedPV P gene, direct sequencing of P gene transcripts was conducted from CedPV-infected Vero cells using multiple sets of primers generating overlapping fragments covering the entire coding region of the P gene. Each produced uniform trace files indicating a lack of RNA editing activities, which is very different from the mixed peaks generated by HeV and NiV immediately after the editing site (Fig. S7). To our knowledge, CedPV is the first member of Paramyxovirinae that lacks both RNA editing and any V-related coding sequence in its P gene. Further investigation is required to exclude the possibility that the P-gene editing in CedPV is cell- or tissue-specific and not present or present at an extremely low level in the current virus-cell system.
The striking similarity in genome size and organization and the presence of highly conserved protein domains among the N, M and L proteins between CedPV and henipaviruses prompted us to investigate the antigenic relatedness of these viruses. Staining of CedPV- infected Vero cells using rabbit anti-henipavirus antibodies indicated the presence of cross-reactivity. This cross-reactivity was further confirmed in reverse by staining of HeV-infected Vero cells using a rabbit serum raised against a recombinant CedPV N protein (Fig. 3). However, analysis by virus neutralization test using either polyclonal or monoclonal antibodies found that henipavirus-neutralizing antibodies were unable to neutralize CedPV. Similarly, CedPV-neutralizing antibodies obtained in our infection studies (see below) also failed to neutralize either HeV or NiV. It can therefore be concluded that CedPV and henipaviruses share cross-reactive antigenic regions, but not cross-neutralizing epitopes.
To further investigate the relationship between CedPV and recognized henipaviruses, we investigated the use of the henipavirus receptors, the ephrin-B2 and -B3 host cell proteins, as potential receptors for CedPV infection. Our previous studies have demonstrated that the ephrin-B2 and -B3 expression negative HeLa-USU cell line could support henipavirus infection and formation of syncytial CPE only when either the ephrin-B2 or -B3 gene was transiently expressed in the cells [22], [34]. For CedPV, similar observations were made with respect to the ephrin-B2 receptor. As shown in Fig. 4, CedPV failed to infect HeLa-USU, but was able to infect and cause syncytial CPE when the human ephrin-B2 gene was expressed. In contrast, when ephrin-B3 molecule was introduced, there was no evidence of infection.
Ferrets, guinea pigs, and mice exhibit differing responses to the previously described henipaviruses HeV and NiV, with ferrets and guinea pigs, but not mice developing severe disease characterized by systemic vasculitis [20], [35], [36], [37], [38]. In contrast, ferrets and guinea pigs exposed to CedPV by, respectively, oronasal and intraperitoneal routes remained clinically well although neutralizing antibody was detected in serum between 10 to 21 days pi (Table 2). Balb-C mice exposed to CedPV by the oronasal route remained clinically well and did not develop neutralizing antibody in serum by day 21 pi. In ferrets electively euthanized at earlier time-points, there was reactive hyperplasia of tonsillar lymphoid tissue, retropharyngeal and bronchial lymph nodes, accompanied by edema and erythrophagocytosis. CedPV antigen was detected in bronchial lymph node of one animal euthanized on day 6 pi, consistent with viral replication in that tissue; cross-reactive immunostaining against anti-NiV N protein antibodies was also noted (Fig. 5). No other significant histological lesions were identified. Viral RNA was detected in selected lymphoid tissues of 3 (of 4) ferrets sampled day 6 to 8 pi, including pharynx, spleen, and retropharyngeal and bronchial lymph nodes, as well as the submandibular lymph node of the ferret euthanized on day 20 pi. This pattern of lymphoid involvement suggests that there may be transient replication in the upper and lower respiratory tracts although CedPV genome was not recovered from nasal washes, oral swabs, pharynx or lung tissue of affected animals. Virus isolation was unsuccessful for all PCR positive tissues.
As a first step towards the understanding of the pathogenicity difference between CedPV and HeV, we examined the IFN responses in human HeLa cells upon virus infection. As shown in Fig. 6, while the induction of IFN-α was similar in cells infected with HeV or CedPV, there was a significant difference of IFN-β production upon infection by HeV or CedPV, with CedPV-infected cell producing a much higher level of IFN-β.
To investigate the CedPV exposure status of pteropid bats in Queensland and potential co-infection (either concurrent or consecutive) of CedPV with HeV, we tested 100 flying fox sera collected previously for other studies for antibody against the two viruses. Due to the cross-reactivity observed above, virus neutralization tests were conducted to obtain more accurate infection data for each virus. Overall, 23% of the sera were CedPV-positive and 37% HeV-positive (Table S1). Co-infection was reflected in 8% of the sera tested.
The emergence of bat-borne zoonotic viruses (including HeV, NiV, Ebola, Marburg, and SARS) has had a significant impact on public health and the global economy during the past few decades. With the rapidly expanding knowledge of virus diversity in bat populations around the world, it is predicted that more bat-borne zoonotic viruses are likely to emerge in the future. The discovery of a novel ebolavirus-like filovirus in Spanish microbats demonstrates that the potential for such spill over events is not limited to Africa or Asia [39]. It is therefore important to enhance our preparedness to counter future outbreaks by conducting active pre-emergence research into surveillance, triggers for cross-species transmission, and the science of identification of potential pathogens.
Henipaviruses represent one of the most important bat-borne pathogens to be discovered in recent history. Although CedPV displays some differences from existing members of the genus Henipavirus, we propose that CedPV be classified as a new henipavirus based on the following shared features with known henipaviruses: 1) it is antigenically related to current henipaviruses; 2) its genome size and organization is almost identical to those of HeV and NiV; 3) it has a similar prevalence in flying foxes; and 4) it uses ephrin-B2 as the cell entry receptor.
The lack of cross-neutralization between CedPV and HeV or NiV was not unexpected from the comparative sequence analysis of all the deduced proteins, especially the G protein (see Table 1). It is clear that the genetic relatedness of CedPV with HeV or NiV is much lower than between HeV and NiV. However, the percentage sequence identities of the major viral proteins between CedPV and HeV/NiV are on average at least 10% higher than that between HeV/NiV and any other known paramyxoviruses. Also, there was no antigenic cross-reactivity observed between CedPV and representative viruses of the other paramyxovirus genera in the subfamily Paramyxovirinae (Fig. S6).
Like other paramyxoviruses, the P gene of henipaviruses produces multiple proteins which play a key role in viral evasion of host innate immune responses [4], [40], [41]. One of these is the Cys-rich V protein: all members of the subfamily Paramyxovirinae produce the V protein with the exception of the human parainfluenza virus 1 (hPIV1). Although a putative RNA editing sequence (AAGAGGG) is present at the expected editing site of the P gene, the hPIV1 RNA polymerase does not produce an edited mRNA of the P gene [42]. There are remnants of the V ORF easily detectable in the hPIV1 P gene although the predicted 68-aa ORF region is interrupted by multiple in-frame stop codons. Of the 7 Cys residues conserved between bovine parainfluenza virus 3 and Sendai virus, four are still present in the non-functional V ORF of hPIV1[42]. In contrast, an extensive ORF and sequence homology search of the CedPV P gene only identified one aa coding region with minimal sequence identity to the V ORFs of HeV and NiV (see Fig. S8). In this region, out of the 9 Cys residues conserved between HeV and NiV V proteins, only 2 are present in the CedPV P gene. Furthermore, the sequence (AGATGAG) upstream from this putative ORF V coding region does not match the consensus RNA editing site. It can therefore be concluded that CedPV is the only member of Paramyxovirinae which lacks both the functional V mRNA/protein and the coding capacity for the RNA editing site and ORF V. The evolutionary significance of this finding needs further investigation.
Our in vitro study indicated that ephrin B2, but not ephrin B3, was able to restore CedPV infection in the ephrin B2-deficient HeLa cells. While this is highly suggestive that ephrin B2 is the functional entry receptor for CedPV, it should be emphasized that this was not a direct proof that ephrin B2 is the receptor. Further investigation is required to confirm this.
In our preliminary studies, it was shown that CedPV was able to replicate in guinea pigs and ferrets, but failed to cause significant clinical diseases, unlike that of the closely related HeV and NiV. These first infection experiments were conducted with a high dose if virus to establish whether the CedPV could replicate in these animals and determine the degree of any clinical disease. A second experiment was then carried out in ferrets to determine the site of replication and tissue tropism in sequentially sacrificed animals. A lower dose was used to gain better comparison with similar infection experiments using HeV and NiV [18], [35]. Although these initial experimental infection studies indicate that CedPV is less or non-pathogenic in these species, it is possible that CedPV may be pathogenic in other hosts, such as horses. We hypothesize that the lack of a V protein may impact on the pathogenicity. In this regard, it was encouraging to observe that infection of human cells by CedPV induced a much more robust IFN-β response than HeV. Further study is required to dissect the exact molecular mechanism of this observed difference.
Due to the close relationship between CedPV and HeV, it was important to investigate the possibility of co-infection by these two viruses in the Australian bat population. Based on the detection of neutralizing antibodies at 23% for CedPV, 37% for HeV and 8% for both, it can be concluded that the co-infection rate is very close to the theoretical rate of 8.5% (the product of the two independent infection rates). Based on this limited preliminary analysis, it appears that infection of bats by one henipavirus neither prevents nor enhances the likelihood of infection by the other.
In summary, the discovery of another henipavirus in Australian flying foxes highlights the importance of bats as a significant reservoir of potential zoonotic agents and the need to expand our understanding of virus-bat relationships in general. Our future research will be directed at determining whether spill-over of CedPV into other hosts has occurred in the past in Australia, whether CedPV is pathogenic in certain mammalian hosts, and whether CedPV exists in bat populations in geographically diverse regions.
All animal studies were approved by the CSIRO Australian Animal Health Laboratory's Animal Ethics Committee and conducted following the Australian National Health and Medical Research Council Code of Practice for the Care and Use of Animals for Scientific Purposes guidelines for housing and care of laboratory animals.
Cell lines used this study were Vero (ATCC), HeLa-USU [22], and the P. alecto primary cell lines derived from kidney (PaKi), brain (PaBr), (spleen) PaSp and placenta (PaPl) recently established in our group [27]. Cells were grown in Dulbecco's Modified Eagle's Medium Nutrient Mixture F-12 Ham supplemented with double strength antibiotic-antimycotic (Invitrogen), 10 µg/ml ciprofloxacin (MP Biomedicals) and 10% fetal calf serum at 37°C in the presence of 5% CO2.
Urine (approximately 0.5–1 ml) was collected off plastic sheets placed underneath a colony of flying foxes (predominantly Pteropus alecto with some P. Poliocephalus in the mixed population) in Cedar Grove, South East Queensland, Australia and pooled into 2-ml tubes containing 0.5 ml of viral transport medium (SPGA: a mix of sucrose, phosphate, glutamate and albumin plus penicillin, streptomycin and fungizone). The tubes were temporarily stored on ice after collection and transported to a laboratory in Queensland, frozen at −80°C, and then shipped on dry ice to the CSIRO Australian Animal Health Laboratory (AAHL) in Geelong, Victoria for virus isolation. The samples were thawed at 4°C and centrifuged at 16,000×g for 1 min to pellet debris. Urine in the supernatant (approximately 0.5–1 ml) was diluted 1∶10 in cell culture media. The diluted urine was then centrifuged at 1,200×g for 5 min and split evenly over Vero, PaKi, PaBr, PaSp and PaPl cell monolayers in 75-cm2 tissue culture flasks. The flasks were rocked for 2 h at 37°C, 14 ml of fresh cell culture media was added and then incubated for 7 d at 37°C. The flasks were observed daily for toxicity, contamination, or viral cytopathic effect (CPE).
Cells showing syncytial CPE were screened using published broadly reactive primers [31] for all known paramyxoviruses and a subset of paramyxoviruses. PCR products were gel extracted and cloned into pGEM T-Easy (Promega) to facilitate sequencing using M13 primers. Sequences were obtained and aligned with known paramyxovirus sequences allowing for initial classification.
Whole genome sequence was determined using a combination of 454 sequencing [43] and conventional Sanger sequencing. Virions from tissue culture supernatant were collected by centrifugation at 30,000×g for 60 min and resuspended in 140 µl of PBS and mixed with 560 µl of freshly made AVL for RNA extraction using QIAamp Viral RNA mini kit (Qiagen). Synthesis of cDNA and random amplification was conducted using a modification of a published procedure [44]. Briefly, cDNA synthesis was performed using a random octomer-linked to a 17-mer defined primer sequence: (5′-GTTTCCCAGTAGGTCTCNNN NNNNN-3′) and SuperScript III Reverse Transcriptase (Invitrogen). 8 µl of ds-cDNA was amplified in 200 µl PCR reactions with hot-start Taq polymerase enzyme (Promega) and 5′-A*G*C*A*C TGTAGGTTTCCCAGTAGGTCTC-3′ (where * denotes thiol modifications) as amplification primers for 40 cycles of 95°C/1 min, 48°C/1 min, 72°C/1 min after an initial denaturation step of 5 min at 95°C and followed by purification with the QIAquick PCR purification kit (Qiagen). Sample preparation for Roche 454 sequencing (454 Life Sciences Branford, CT, USA) was according to their Titanium series manuals, Rapid Library Preparation and emPCR Lib-L SV.
To obtain an accurate CedPV genome sequence, 454 generated data (after removing low quality, ambiguous and adapter sequences) was analysed by both de novo assembly and read mapping of raw reads onto the CedPV draft genome sequence derived from Sanger sequencing. For 454 read mapping, SNPs and DIPs generated with the CLC software were manually assessed for accuracy by visualising the mapped raw reads (random PCR errors are obvious compared to real SNPs and DIPs especially when read coverage is deep). Consensus sequences for both 454 de novo and read mapping assembly methods were then compared to the Sanger sequence with the latter used to resolve conflicts within the low coverage regions as well as to resolve 454 homopolymer errors.
Sequences of genome termini were determined by 3′- and 5′-RACE using a protocol previously published by our group [45]. Briefly, approximately 100 ng of RNA was ligated with adaptor DT88 (see reference for sequence information) using T4 RNA ligase (Promega) followed by cDNA synthesis using the SuperScript III RT kit (Invitrogen) and an adaptor-specific primer, DT89. PCR amplification was then carried out using DT89 and one or more genome-specific primers. PCR products were sequenced directly using either DT89 or genome specific primers by an in-house service group on the ABI Sequencer 3100.
The CLC Genomics Workbench v4.5.1 (CLC Inc, Aarhus, Denmark) was used to trim 454 adapter and cDNA/PCR primer sequences, to remove low quality, ambiguous and small reads <15 bp and to perform de novo and read mapping assemblies all with default parameters. Clone Manager Professional ver 9.11 (Scientific and Educational Software, Cary, NC, USA) was used to join overlapping contigs generated by de novo assembly. Phylogenetic trees were constructed by using the neighbor-joining algorithm with bootstrap values determined by 1,000 replicates in the MEGA4 software package [46].
Quantitative PCR assays (qPCR) were established based on CedPV-specific sequences obtained from the high throughput sequencing. A TaqMan assay on the P gene was developed and used for all subsequent studies. The sequences of the primer/probe are as follows: forward primer, 5′-TGCAT TGAGC GAACC CATAT AC; reverse primer, 5′-GCACG CTTCT TGACA GAGTT GT; probe, 5′-TCCCG AGAAA CCCTC TGTGT TTGA-MGB.
The coding region for the CedPV N protein was amplified by PCR with a pair of primers flanked by AscI (5′ end) and NotI (3′ end) sites for cloning into our previously described GST-fusion expression vector [47]. The expression and purification by gel elution was conducted as previously described [48]. For antibody production, purified protein was injected subcutaneously into 4 different sites of 2 adult (at a dose of 100 µg per animal) New Zealand white female rabbits at days 0 and 27. The CSIRO's triple adjuvant [49] was used for the immunization. Animals were checked for specific antibodies after days 5 and 42 and euthanized at day 69 for the final blood collection.
For immunofluorescence antibody test, Vero cell monolayers were prepared in 8-well chamber slides by seeding at a concentration of 30,000 cells/well in 300 µl of cell media and incubating over night at 37°C. The cell monolayers were infected with an MOI of 0.01 of CedPV, HeV or NiV and fixed with 100% ice-cold methanol at 24 h post-infection. The chamber slides were blocked with 100 µl/well of 1%BSA in PBS for 30 min at 37°C before adding 50 µl/well of rabbit sera against CedPV N or NiV N diluted 1∶1000. After incubation at 37°C for 30 min, the slides were washed three times in PBS-T and incubated with 50 µl/well of anti-rabbit 488 Alexafluore conjugate (Invitrogen) diluted 1∶1000 at 37°C for 30 min. The slides were then washed three times in PBS-T and mounted in 50% glycerol/PBS for observation under a fluorescence microscope.
For virus neutralization test, serial two-fold dilutions of sera were prepared in duplicate in a 96-well tissue culture plate in 50 µl cell media (Minimal Essential Medium containing Earle's salts and supplemented with 2 mM glutamine, antibiotic-antimycotic and 10% fetal calf serum). An equal volume containing 200 TCID50 of target virus was added and the virus-sera mix incubated for 30 min at 37°C in a humidified 5% CO2 incubator. 100 µl of Vero cell suspension containing 2×105 cells/ml was added and the plate incubated at 37°C in a humidified 5% CO2 incubator. After 4 days, the plate was examined for viral CPE. The highest serum dilution generating complete inhibition of CPE is defined as the final neutralizing titer.
Human ephrin B2 and B3 genes were cloned into pQCXIH (Clontech) and the resulting plasmids packaged into retrovirus particles in the GP2–293 packaging cell line (Clontech) and pseudotyped with vesicular stomatitis virus G glycoprotein (VSV-G) following the manufacturer's instructions. HeLa-USU cell line [22] was infected with the VSV-G pseudotyped retrovirus particles in the presence of 1 µg/ml polybrene (Sigma). 8 h post infection, the medium was changed and the cells were allowed to recover for 24 h, allowing time for the retroviral insert to be incorporated into the cell genome and for expression of the hygromycin resistance gene. 24 h post infection, cells transformed by the retrovirus were selected for by the addition of 200 µg/ml hygromycin in the media. Stocks of cells that were resistant to hygromycin were prepared and frozen. HeLa-USU and ephrin-expressing HeLa-USU cells were seeded in 6-well tissue culture plates at a density of 250,000 cells/well overnight. The viruses (HeV and CedPV) were diluted to give an MOI of 0.01 and inoculated into the wells. The cell monolayers were examined daily for syncytial CPE.
Animal studies were carried out in the BSL4 animal facility at AAHL. Ferrets, guinea pigs and mice were used on the basis of their known and varying responses to exposure to other henipaviruses.
Firstly, 2×106 TCID50/ml CedPV passaged twice in bat PaKi cells was administered to 2 male ferrets (1 ml oronasally); 4 female guinea pigs (1 ml intraperitoneally); and 5 female Balb-C mice (50 µl oronasally). Guinea pigs and mice were implanted with temperature sensing microchips (LifeChip Bio-thermo, Destron Fearing) and weighed daily. Ferret rectal temperature and weight was recorded at sampling times. Animals were observed daily for clinical signs of illness and were euthanized at 21 d post-inoculation. Sera were collected on days 10, 15 and 21 to test for neutralizing antibody against CedPV.
Secondly, on the basis of asymptomatic seroconversion to CedPV noted in ferrets in the first study, 7 further female ferrets were exposed by the oronasal route to a lower dose of 3×103 TCID50. Two animals were euthanized on each of days 6, 8 and 10 post-inoculation and one on day 20. Nasal washes, oral swabs, and rectal swabs were collected on days 2, 4, 6, 8 and 10 and urine was sampled on the day of euthanazia; each specimen was assessed for CedPV genome. A wide range of tissue samples were collected at post mortem examination and assessed by routine histology, immunohistochemistry (using rabbit antibodies raised against recombinant CedPV and NiV N proteins, respectively), qPCR (see above) and virus isolation using reagents and procedures previously established in our group [16].
HeLa cells were infected with Hendra and Cedar viruses at an MOI 0.5 for 24 hours, at which time total cellular RNA was extracted and IFN-α and IFN-β mRNA levels were quantified by real-time PCR using Power SYBR Green RNA-to-CT 1-Step Kit (Applied Biosystems). Primers were as previously described [50].
Sera from 100 flying foxes collected during 2003–2005 from Queensland, Australia were screened for neutralizing antibodies to CedPV. Virus neutralization test was conducted as described above (antibody tests). All serum samples were tested at a dilution of 1∶20.
|
10.1371/journal.pbio.1001881 | Hypoxia-Inducible Factor-2α Is an Essential Catabolic Regulator of Inflammatory Rheumatoid Arthritis | Rheumatoid arthritis (RA) is a systemic autoimmune disorder that manifests as chronic inflammation and joint tissue destruction. However, the etiology and pathogenesis of RA have not been fully elucidated. Here, we explored the role of the hypoxia-inducible factors (HIFs), HIF-1α (encoded by HIF1A) and HIF-2α (encoded by EPAS1). HIF-2α was markedly up-regulated in the intimal lining of RA synovium, whereas HIF-1α was detected in a few cells in the sublining and deep layer of RA synovium. Overexpression of HIF-2α in joint tissues caused an RA-like phenotype, whereas HIF-1α did not affect joint architecture. Moreover, a HIF-2α deficiency in mice blunted the development of experimental RA. HIF-2α was expressed mainly in fibroblast-like synoviocytes (FLS) of RA synovium and regulated their proliferation, expression of RANKL (receptor activator of nuclear factor–κB ligand) and various catabolic factors, and osteoclastogenic potential. Moreover, HIF-2α–dependent up-regulation of interleukin (IL)-6 in FLS stimulated differentiation of TH17 cells—crucial effectors of RA pathogenesis. Additionally, in the absence of IL-6 (Il6−/− mice), overexpression of HIF-2α in joint tissues did not cause an RA phenotype. Thus, our results collectively suggest that HIF-2α plays a pivotal role in the pathogenesis of RA by regulating FLS functions, independent of HIF-1α.
| Rheumatoid arthritis (RA) is a systemic autoimmune disorder characterized by chronic inflammation in joint tissues leading to destruction of cartilage and bone. Despite some therapeutic advances, the etiology of RA pathogenesis is not yet clear, and effective treatment of RA remains a significant, unmet medical need. Hypoxia is a prominent feature of inflamed tissue within RA-affected joints, and earlier work has implicated limited involvement of hypoxia-inducible factor (HIF)-1 α. We explored the role of a second HIF family member, HIF-2α, in RA pathogenesis. We showed that HIF-2α is markedly increased in the tissue lining the RA-affected joints. Notably and in contrast to HIF-1α, when overexpressed in normal mouse joint tissues, HIF-2α is sufficient to cause RA-like symptoms. Conversely, an HIF-2α deficiency blocks the development of experimental arthritis in mice. We discovered further that HIF-2α regulates RA pathogenesis by modulating various RA-associated functions of joint-specific fibroblast-like cells, including proliferation, expression of cytokines, chemokines, and matrix-degrading enzymes, and bone-remodeling potential. HIF-2α also increases the ability of these cells to promote interleukin-6–dependent differentiation of TH17 cells, a known effector of RA pathogenesis. We thus show that HIF-1α and HIF-2α have distinct roles and act via different mechanisms in RA pathogenesis.
| Rheumatoid arthritis (RA) is a chronic inflammatory autoimmune disease that mainly targets the synovial membrane, resulting in destruction of the joint architecture. The pathophysiology of RA involves numerous cell types, including T cells, B cells, macrophages, synoviocytes, chondrocytes, and osteoclasts, all of which contribute to the process of RA pathogenesis [1]. T-cell–mediated autoimmune responses play an important role in RA pathogenesis, in which interleukin (IL)-17–producing T-helper cells (TH17) act as crucial effectors [1],[2]. RA is characterized by synovial hyperplasia and synovitis with infiltration of immune cells. Synovial tissues express numerous cytokines that have been directly implicated in many immune processes of RA pathogenesis [1],[3]. Additionally, an aggressive front of hyperplastic synovium, called the pannus, invades and destroys mineralized cartilage and bone through the action of osteoclasts [1],[3]. Synovial hyperplasia results from a marked increase in macrophage-like and fibroblast-like synoviocytes (FLS). Accumulating evidence indicates that activated FLS are among the key players in RA joint destruction [4]. FLS actively contribute to the initiation, propagation, and maintenance of synovial inflammation through secretion of factors and direct cell–cell interactions. For instance, cytokines and chemokines produced by FLS attract T cells to RA synovium, and the interaction of FLS with T cells results in activation of both cell types. FLS in the inflamed synovium also contribute to RA pathogenesis by producing matrix-degrading enzymes involved in cartilage destruction; RANKL (receptor activator of nuclear factor–κB ligand), which regulates osteoclast differentiation, leading to bone erosion; and angiogenic factors associated with blood vessel formation [4]. Despite therapeutic advances, the etiology of RA pathogenesis has not yet been entirely elucidated, and effective treatment of RA remains a significant unmet medical need.
A prominent feature of the inflamed RA synovium is hypoxia [5]–[7], suggesting a possible role for hypoxia-inducible factors (HIFs) in RA pathogenesis. HIFs are members of a transcription factor family that act as “master regulators” of the adaptive response to hypoxia [8],[9]. Of the three isoforms, HIF-1α (encoded by HIF1A) and HIF-2α (encoded by EPAS1) are the most extensively studied. HIF-1α is up-regulated in RA synovium [10]–[12], where it appears to be associated with angiogenesis [5]–[7]. HIF-1α is also expressed in TH17 cells, where it serves to regulate TH17/Treg balance; a lack of HIF-1α in TH17 cells impairs their differentiation [13],[14]. Additionally, loss of HIF-1α in myeloid cells reduces the RA pathogenesis caused by K/BxN serum transfer [15]. Although these results suggest that HIF-1α is an important mediator of RA pathogenesis, whether HIF-1α is sufficient to cause RA pathogenesis in vivo has not been previously demonstrated. Most strikingly, HIF-2α, which is closely related to HIF-1α, has not yet been investigated for a role in RA pathogenesis. Indeed, despite many similarities between HIF-1α and HIF-2α, these two isoforms show different sensitivity to oxygen tension and display distinct, and sometimes opposing, cellular activities [8],[9]. Here, we present an extensive study of the function of HIF-2α in experimental inflammatory arthritis in mice. We also investigated whether the role of HIF-2α is independent of, complementary to, or redundant with that of HIF-1α in the development and pathogenesis of experimental RA. We report here that HIF-2α is an essential catabolic regulator of RA pathogenesis, independent of the action of HIF-1α.
To explore possible functions of HIFs in RA pathogenesis, we first examined the expression pattern of HIFs by immunostaining human RA joint sections. HIF-2α was highly expressed in the intimal lining of human RA synovium, where other markers of inflamed RA synovium were expressed, including IL-6, matrix metalloproteinase (MMP)3, and MMP13 (Figure 1A). Indeed, double immunostaining for HIF-2α and these markers revealed their co-localization in human RA synovium (Figure 1B). HIF-2α was also up-regulated in tartrate-resistant acid phosphatase (TRAP)-positive osteoclasts in bone tissue and chondrocytes in damaged cartilage, but not in the intact, undamaged part of human RA cartilage (Figure S1A). In contrast, HIF-1α was detected only in a few cells in the sublining and deep layer of human RA synovium (Figure 1A). However, neither HIF-1α nor HIF-2α was detected in human osteoarthritis (n = 10), psoriatic arthritis (n = 2), or gouty arthritis (n = 2) synovium (Figure S1B). These results indicate RA-specific differential up-regulation of HIF-1α and HIF-2α in synovial tissues.
We extended these results using the collagen-induced arthritis (CIA) model of RA in DBA/1J mice. This is a commonly used experimental model of inflammatory joint arthritis caused by a T-cell–dependent, antibody-mediated autoimmune response directed against cartilage type II collagen [16]. Compared with nonimmunized (NI) control joints, joints in CIA mice exhibited destruction typical of RA (Figure S1C–E). HIF-2α was highly up-regulated in the region lining the CIA synovium (Figure 1C), where it was co-localized with the RA-synovium markers, IL6, MMP3, and MMP13 (Figure 1B). Unlike HIF-2α expression, HIF-1α was rarely detected in the intimal lining, but was detected in cells of the sublining and deep layer (Figure 1C). Similar to human RA joint tissues, HIF-2α was also detected in pannus and damaged cartilage (Figure S1F). Quantitation of relative HIF expression levels further confirmed the marked up-regulation of HIF-2α compared with HIF-1α in human RA and mouse CIA synovia (Figure 1D). HIF-2α–positive cells were much more abundant in synovial lining cells (fibroblast-like and macrophage-like synoviocytes) compared with sublining macrophages and endothelial cells in blood vessels of RA synovium (Figure 1D).
The expression patterns of HIF-1α and HIF-2α in RA synovium suggested differential roles of HIF isoforms. To explore the possible in vivo functions of HIFs, we overexpressed HIF-1α or HIF-2α in the knee joint tissues of DBA/1J mice via intra-articular (IA) injection of Ad-Hif1a or Ad-Epas1 adenoviruses (1×109 plaque-forming units [PFUs]), respectively. Immunostaining of joint tissue sections 3 wk after IA injection revealed that the respective adenoviruses caused marked overexpression of HIF-1α and HIF-2α in the synovium, cartilage, and meniscus of joint tissues (Figure 2A and B). HIF-2α expression in joint tissues caused typical RA-like phenotypic manifestations, including synovial hyperplasia and severe synovitis, determined by hematoxylin and eosin (H&E) staining and scoring of inflammation (Figure 2C and D); marked cartilage destruction, determined by safranin-O staining and scored by Mankin's method (Figure 2E); pannus formation and invasion into calcified cartilage and bone, determined by hematoxylin/safranin-O staining and scoring (Figure 2E); and angiogenesis in the synovium, determined by immunostaining for CD31 and counting blood vessels in synovia of knee and ankle joints (Figure 2E). Overexpressed HIF-2α in the synovium of Ad-Epas1–injected mice was co-localized with the RA-synovium marker IL6, as determined by double-immunofluorescence microscopy (Figure S1G). In contrast to HIF-2α, HIF-1α overexpression did not cause any changes in joint architecture, including hallmarks of RA such as synovitis, pannus formation, angiogenesis, and cartilage destruction (Figure 2C–E). Collectively, these results indicate that ectopic expression of HIF-2α, but not HIF-1α, causes typical RA-like joint destruction in mice, suggesting distinct functions of HIF-1α and HIF-2α in RA pathogenesis.
We confirmed the role of HIF-2α using HIF-2α–knockout mice or local deletion of HIF-2α in joint tissues. We first examined HIF-2α functions using mice with reduced expression of the Epas1 gene encoding HIF-2α. Because homozygous deletion of Epas1 (Epas1−/−) is embryonic lethal [17], we used heterozygous Epas1+/− mice. We have previously shown that deletion of one allele of Epas1 is sufficient to inhibit OA cartilage destruction [18]. Whereas Epas1+/− DBA/1J mice showed reduced expression levels of HIF-2α mRNA, HIF-1α mRNA levels were not altered in these mice (unpublished data). Compared with wild-type (WT) littermates, Epas1+/− DBA/1J mice showed a significantly reduced incidence (89.4%±7.1% versus 33.2%±6.5%, p = 0.0004) and severity (2.85%±0.26% versus 1.10%±0.10%, p = 0.004) of CIA on day 60 after the first injection of type II collagen (Figure 3A). Epas1+/− DBA/1J mice under CIA conditions also showed a significant reduction in all the examined hallmarks of RA. These include paw swelling and increased ankle thickness (Figure 3B), elevated serum levels of autoantibodies against type II collagen (Figure 3C), synovitis (Figure 3D), cartilage destruction (Figure 3E and F), pannus formation and invasion (Figure 3E and F), and angiogenesis in inflamed synovium (Figure 3E and F).
We further validated HIF-2α functions in CIA by locally deleting Epas1 in joint tissues via IA injection of Ad-Cre (1×109 PFU) in Epas1fl/fl mice. Immunostaining of joint sections revealed that Ad-Cre injection effectively reduced the elevated levels of HIF-2α induced by CIA in joint tissues, including synovium, cartilage, and pannus (Figure 4A). Moreover, local deletion of Epas1 in joint tissues by Ad-Cre injection significantly inhibited RA pathogenesis by blocking synovitis and synovial hyperplasia, pannus formation and invasion into calcified cartilage and bone, angiogenesis in inflamed synovium, and cartilage destruction (Figure 4B and C). These results collectively indicate that Epas1 knockdown (Epas1+/−) or local deletion (Ad-Cre) inhibits experimental RA in mice.
Next, we investigated the inhibitory mechanisms of RA pathogenesis in Epas1+/− DBA/1J mice by examining immune responses. Epas1+/− mice showed normal populations of CD4+ and CD8+ T cells in lymph nodes, as determined by flow cytometry (Figure 5A). Flow cytometry also revealed no differences in immune cell populations between WT and Epas1+/− DBA/1J mice, including CD4+ and CD8+ T cells in spleen and thymus; Foxp3-expressing regulatory T cells (Treg) in lymph node, spleen, and thymus; naïve (CD44lowCD62Lhigh) and memory (CD44highCD62Llow) CD4+ T cells in lymph node and spleen; and B220+ B cells and CD11c+ dendritic cells in lymph node and spleen (Figure S2A–D). Proliferation of CD4+ T cells and B220+ B cells isolated from lymph nodes and spleens was similar between WT and Epas1+/− DBA/1J mice (Figure S2E and F). Additionally, CD4+ T cells isolated from lymph nodes and spleens of Epas1+/− mice showed a normal capacity to differentiate into TH1, TH2, and TH17 cells, as determined by the detection of specific cytokines (Figure 5B and Figure S2G and H).
Although immune system development was not affected in Epas1+/− mice, HIF-2α knockdown in these mice modulated immune responses under CIA conditions. The population of IL17A–producing cells in lymph nodes and spleens as well as the levels of secreted IL17A, which plays a key role in TH17 cell differentiation and autoimmune responses, were significantly down-regulated under CIA conditions in Epas1+/− mice (Figure 5C). We further validated the effects of Epas1 knockdown on pathogenic cytokine expression in synovial cells using a total mixed-cell population isolated from synovial tissues of WT and Epas1+/− mice. mRNA levels of the pathogenic cytokines IL1β, IL6, IL12, IL17A, IL17F, TNFα, and interferon (IFN)-γ under CIA conditions were significantly down-regulated in the total synovial cell population isolated from Epas1+/− mice compared with WT littermates (Figure 5D). Conversely, IA injection of Ad-Epas1 (1×109 PFU) significantly increased mRNA levels of IL6, IL17A, and IL17F in the total synovial cell population compared with those in Ad-C–injected mice (Figure 5D). Collectively, our results indicate that Epas1 knockdown in Epas1+/− DBA/1J mice does not alter the development pattern of the immune system, but does significantly reduce the production of pathogenic cytokines under CIA conditions.
HIF-2α is up-regulated mainly in the intimal lining of RA synovium, which primarily consists of FLS and macrophage-like synoviocytes [4]. We therefore examined which cell types overexpress HIF-2α in inflamed RA synovium. Double-immunofluorescence microscopy of human RA (Figure 6A) and mouse CIA (Figure 6B) synovia revealed co-localization of HIF-2α with FLS markers (vimentin or CD55), whereas only a subset of CD68-positive macrophages expressed HIF-2α. We further examined HIF-2α expression in primary cultures of the total synovial cell population isolated from CIA mice; these cells consist of FLS, macrophages, and dendritic cells, among others (Figure 6C). Most FLS (∼92%) were positive for HIF-2α staining, whereas only ∼32% of macrophages were positive for HIF-2α staining (Figure 6C). To elucidate the role of HIF-2α expression in macrophages, we stimulated Raw264.7 cells (a murine macrophage cell line) with TNFα or lipopolysaccharide (LPS). Both stimuli caused up-regulation of the inflammatory mediators COX2 (cyclooxygenase 2) and iNOS (inducible nitric oxide synthase), without affecting HIF-2α expression (Figure 6D). These results collectively suggest that HIF-2α is mainly up-regulated in FLS of RA synovium, where it may play a major role in RA pathogenesis.
Next, we investigated the mechanisms regulating HIF-2α expression using primary cultures of mouse FLS. The pro-inflammatory cytokines IL1β and TNFα induced up-regulation of HIF-2α in FLS, whereas IL6 and IL17 did not affect HIF-2α expression (Figure 6E). A pharmacological analysis using inhibitors of nuclear factor–kappaB (NF-κB) and mitogen-activated protein (MAP) kinase subtypes indicated that IL1β- and TNFα-induced HIF-2α expression in FLS is mediated by the NF-κB pathway, but not by the MAP kinase pathway (Figure 6F). Because hypoxia is a prominent feature of the inflamed RA synovium [5]–[7], we additionally examined the role of hypoxia in HIF-2α expression in FLS. Mouse CIA synovium was hypoxic, as determined by pimonidazole staining (Figure 6G). However, unlike HIF-1α protein levels, which were markedly elevated in FLS under hypoxic conditions, HIF-2α protein showed only minimal accumulation under the same conditions; however, Ad-Epas1 infection under normoxic conditions caused marked expression of HIF-2α protein (Figure 6G). Collectively, these findings suggest that pro-inflammatory cytokines, rather than hypoxia, are the leading cause of HIF-2α expression in FLS under CIA conditions.
FLS play a crucial role in RA pathogenesis by producing various regulatory factors [4]. We therefore explored whether up-regulated HIF-2α in FLS modulates FLS functions and thereby RA pathogenesis. Because increased survival and/or proliferation of FLS contribute to synovial hyperplasia [4], we first examined HIF-2α regulation of apoptosis and proliferation in these cells. Ad-Epas1–mediated HIF-2α overexpression in primary cultured FLS did not cause apoptosis or modulate apoptosis induced by an anti-Fas antibody (unpublished data) known to cause FLS apoptosis [4]. However, HIF-2α overexpression significantly increased proliferation of FLS, and IL1β-induced proliferation was inhibited in Epas1+/− FLS (Figure 7A). Moreover, staining for the cell proliferation marker Ki67 revealed the presence of proliferating cells in the intimal lining of both CIA and Ad-Epas1–infected synovia; notably, this staining was markedly reduced in Epas1+/− DBA/1J mice (Figure 7B). Double immunostaining for HIF-2α and Ki67 indicated that 16% and 24% of HIF-2α–positive cells were proliferative in inflamed synovia caused by CIA and Ad-Epas1 injection, respectively (Figure 7C).
Pannus formation and invasion into adjacent cartilage and bone are important regulatory steps in cartilage and bone erosion, which is mediated by the actions of osteoclasts [1],[3],[4]. Osteoclastogenesis is regulated by RANKL, which is produced by FLS and T cells, and requires physical contact of precursor cells with RANKL-expressing FLS or T cells in RA synovium [3],[19]. We therefore examined a possible role for HIF-2α in FLS regulation of RANKL expression, osteoclastogenesis, and pannus formation. HIF-2α overexpression or IL1β treatment of FLS caused significant up-regulation of RANKL mRNA levels (Figure 7D). Additionally, immunostaining indicated markedly increased levels of RANKL protein in CIA synovium of WT mice, an effect that was reduced in Epas1+/− mice (Figure 7E). HIF-2α and RANKL were co-localized in CIA synovium, as determined by double immunostaining (Figure 7E). Consistent with this, TRAP staining revealed an increase in the number of multinucleated osteoclasts in the pannus of the bone–cartilage interface of CIA and Ad-Epas1–infected joints of WT mice; this too was also significantly reduced in Epas1+/− mice (Figure 7F). The role of HIF-2α in osteoclastogenesis was further determined using Epas1+/− precursor cells, with and without HIF-2α overexpression. Osteoclastogenesis of Epas1+/− precursor cells was significantly reduced compared with that of WT cells (Figure 7G). Moreover, overexpression of HIF-2α in precursor cells by Ad-Epas1 infection enhanced osteoclastogenesis (Figure 7G). These results collectively support the idea that HIF-2α–mediated production of RANKL in FLS and osteoclastogenesis of precursor cells contribute to cartilage and bone erosion during HIF-2α–induced RA pathogenesis.
FLS regulate RA pathogenesis by producing various cytokines, chemokines, and matrix-degrading enzymes involved in inflammation, chemotaxis, cartilage destruction, and bone erosion [4]. This led us to explore a possible role for HIF-2α in the expression of these regulatory factors. Ad-Epas1–infected FLS exhibited significantly increased mRNA levels of matrix-degrading enzymes (MMP3, MMP9, MMP12, MMP13, and ADAMTS4), chemokines (CCL2, CCL5, CCL7, CXCL1, CXCL2, CXCL4, CXCL5, and CXCL10), and inflammatory mediators (COX2 and iNOS) (Figure 8A). Among the cytokines examined (IL1β, IL6, IL11, IL12, IL17, IL21, LIF, and TNFα), both mRNA and protein levels of IL6 and TNFα were increased in response to HIF-2α overexpression (Figure 8B and C). Moreover, IL1β-induced up-regulation of catabolic factors (matrix-degrading enzymes, cytokines, and chemokines) was abolished by the knockdown of Epas1 with two independent small interfering RNAs (siRNAs) (Figure S3). In contrast to the effects of HIF-2α, overexpression of HIF-1α in FLS caused up-regulation of MMP9, IL6, COX2, and VEGF, but not that of other factors regulated by HIF-2α (Figure 8C).
TH17 cells are crucial effectors of RA pathogenesis [2],[3], and HIF-1α has been previously shown to regulate TH17 cell differentiation [13],[14]. For instance, enhanced HIF-1α expression in TH17 cells positively regulates TH17 differentiation by up-regulating RORγt, an isoform of RAR-related orphan receptor gamma [13]. We therefore examined possible functions of HIF-2α in TH17 cell differentiation, and thereby RA pathogenesis. We first examined mRNA levels of HIF-1α and HIF-2α during TH17 cell differentiation. Compared with precursor CD4+ T cells, differentiated TH17 cells exhibited significant down-regulation of HIF-2α expression and significant up-regulation of HIF-1α expression (Figure 9A). Unlike the case with HIF-1α, which enhances TH17 cell differentiation [13], overexpression of HIF-2α in precursor CD+ T cells did not affect TH17 cell differentiation (Figure 9B), suggesting that HIF-2α in CD4+ T cells does not directly modulate TH17 cell differentiation.
It is well established that IL6 plays a key role in TH17 cell differentiation [2],[3]. Consistent with this, in vitro TH17 cell differentiation was dependent on the addition of exogenous IL6 protein (Figure 9C). Given the marked overexpression of IL6 in FLS induced by HIF-2α, we explored possible functions of HIF-2α–regulated, FLS-derived IL6 in TH17 cell differentiation by treating CD4+ precursor T cells with conditioned medium (CM) prepared from Ad-C (control)- or Ad-Epas1–infected FLS. TH17 cell differentiation was evaluated by monitoring IL17A production using an enzyme-linked immunosorbent assay (ELISA). As shown in Figure 9C, addition of CM from Ad-Epas1–infected FLS from WT mice induced TH17 cell differentiation, even in the absence of exogenous IL6 protein. The specific role of IL6 in CM was confirmed by preparing CM from FLS of Il6−/− mice or by adding IL6 neutralizing antibody to the CM. Compared with CM from WT FLS, CM of Ad-Epas1–infected FLS from Il6−/− mice did not affect in vitro TH17 cell differentiation (Figure 9C). Furthermore, addition of a neutralizing antibody against IL6, but not TNFα, blocked stimulation of TH17 cell differentiation by the CM of Ad-Epas1–infected FLS (Figure 9D). We additionally confirmed TH17 cell differentiation by monitoring mRNA levels of IL17A and IL17F using quantitative reverse transcription–polymerase chain reaction (qRT-PCR) analysis (Figure S4). Immunostaining of synovial sections also revealed the presence of IL17A-producing cells in mouse synovium infected with Ad-Epas1 or under CIA conditions, whereas no positive immunostaining was observed in Ad-C–infected or NI synovium (Figure S5A–C). Indeed, IL17A-positive cells were located in close proximity to HIF-2α–positive cells in human RA and mouse CIA synovia, as determined by double immunostaining (Figure S5D).
The above results suggest that FLS-derived IL6 plays an important role in HIF-2α regulation of experimental RA by regulating TH17 cell differentiation. To confirm this, we investigated IL6 functions in HIF-2α–induced experimental RA using Il6−/− mice. Consistent with the inhibition of CIA by Il6 knockout [20],[21], we also observed significantly greater inhibition of synovitis, pannus formation and invasion, cartilage destruction, and angiogenesis in inflamed synovium under CIA conditions in Il6−/− DBA/1J mice compared with WT littermates (Figure 9E and F). More importantly, the development of RA-like phenotypic manifestations, including synovitis, pannus formation and invasion, cartilage destruction, and angiogenesis, induced in inflamed synovium by IA injection of Ad-Epas1 was markedly diminished in Il6−/− DBA/IJ mice compared with WT mice (Figure 9E and F). Our results collectively suggest that FLS-derived IL6 plays an important role in TH17 cell differentiation and thereby contributes to HIF-2α regulation of experimental RA.
Our current findings provide two novel insights into the regulation of RA pathogenesis by HIF pathways: the catabolic role of HIF-2α in RA pathogenesis and the differential actions of HIF-1α and HIF-2α in this disease.
In the first case, we demonstrate an essential role for HIF-2α in the pathogenesis of RA. Despite circumstantial evidence for the hypoxic status of RA synovium [5]–[7] and increased expression of HIF-2α in the synovial lining of human RA patients [9], little is currently known about the role of HIF-2α in RA pathogenesis. The results of our loss-of-function studies utilizing Epas1 knockdown in mice (Epas1+/−) or local deletion in Epas1fl/fl mice by Ad-Cre injection strongly support our conclusion that HIF-2α is necessary for RA pathogenesis. This conclusion is reinforced by the marked up-regulation of HIF-2α observed in RA synovia of humans and mouse models of RA as well as the RA-like phenotype revealed in gain-of-function studies involving IA injection of Ad-Epas1. In RA joint tissues, HIF-2α is up-regulated in various tissues, including synovium, pannus, cartilage, meniscus, and TRAP-positive osteoclasts. IA injection of Ad-Epas1 also caused up-regulation of HIF-2α in these tissues. Because numerous cell types in joint tissues contribute to the process of RA pathogenesis [1], up-regulated HIF-2α in any of these tissues could contribute to RA pathogenesis. However, because HIF-2α levels were most markedly increased in synovial cells, which are also the primary targets of adenovirus infection, we characterized HIF-2α functions in synovial tissue in the regulation of RA development. In RA synovial tissue, HIF-2α was up-regulated in most FLS in the synovium lining compartment, although some other cell types, such as macrophages, also exhibited HIF-2α up-regulation. Although we cannot rule out a contribution of these other cell types, we were able to demonstrate that HIF-2α regulates RA-associated FLS functions in experimental RA pathogenesis. These include proliferation; expression of cytokines, chemokines, and matrix-degrading enzymes; RANKL expression and osteoclastogenesis; IL6 production; and IL6-dependent TH17 cell differentiation. Among these, IL6-dependent TH17 cell differentiation is a crucial effector of RA pathogenesis. In this context, we demonstrated that IL6 present in CM prepared from FLS caused TH17 cell differentiation. Moreover, IL17A-positive cells were located in close proximity to HIF-2α–positive cells, suggesting that IL6 production mediated by HIF-2α in the inflamed RA synovium affects differentiation of neighboring TH17 cells. Additional support for this relationship is provided by our demonstration that global deletion of Il6 abolished HIF-2α–induced RA pathogenesis. Although it remains possible that production of IL6 by cell types in synovial tissue besides FLS could also contribute to the regulation of TH17 cell differentiation, establishing this definitively would likely require a conditional FLS-specific Il6-knockout model, which, to our knowledge, has not yet been developed.
The second novel finding of this study is that HIF-1α and HIF-2α have distinct roles and act via different mechanisms in RA pathogenesis. HIF-1α is up-regulated in RA synovium [10]–[12], where it is associated with angiogenesis [5]–[7]. It has previously been demonstrated that HIF-1α regulates RA pathogenesis by directly modulating TH17 cell functions [13],[14]. In the current study, HIF-1α expression, in contrast to that of HIF-2α, was detected in a small number of cells in the sublining and deep layer of RA synovium in both humans and experimental mouse models of RA, a result consistent with other reports [10],[11]. We did not extensively explore the underlying mechanisms of this differential expression of HIF isoforms in the current study. However, HIF-1α and HIF-2α show different sensitivity to oxygen tension and display distinct, and sometimes opposing, cellular activities [8],[9]. Indeed, we found in this study that sensitivities to hypoxia and to pro-inflammatory cytokines differed between HIF-1α and HIF-2α in FLS. These differences may reflect the differential expression pattern of HIF-2α and HIF-2α in RA synovium. Nevertheless, ectopic expression of HIF-1α in joint tissues by IA injection of Ad-Hif1a did not cause an RA-like phenotype, suggesting that HIF-1α overexpression is not sufficient to induce RA pathogenesis. In striking contrast, HIF-2α overexpression was sufficient to activate RA pathogenesis and did so by regulating FLS functions. Collectively, our results suggest that HIF-2α regulates RA pathogenesis by acting globally to modulate the RA pathogenesis program, including angiogenesis and FLS functions, whereas HIF-1α contributes to RA pathogenesis by modulating the effector functions of myeloid and T cells. Moreover, the observation that HIF-2α deficiency, which does not affect HIF-1α expression, is sufficient to inhibit experimental RA underscores the specific roles played by HIF-2α.
RA and OA are the most common types of joint arthritis. We have previously shown that HIF-2α is a catabolic regulator of OA cartilage destruction [18],[22]–[24], demonstrating that HIF-2α causes OA pathogenesis by up-regulating catabolic enzymes such as MMP3 and MMP13 in chondrocytes, and further showing that chondrocyte-specific Col2a1-Epas1 TG mice exhibit spontaneous cartilage destruction with no evidence of synovitis [18]. Although RA and OA phenotypes share certain features, such as cartilage destruction, their etiology and pathogenesis are completely different. RA and OA also differ with respect to outcomes, cell types associated with the pathogenesis, and therapeutic approaches. For instance, OA is a degenerative joint disease (“wear and tear” arthritis) that begins with the destruction of surface articular cartilage, subchondral bone sclerosis, and osteophyte formation in a single joint. In this type of arthritis, mechanical stresses, including joint instability and injury, and factors that predispose toward OA, such as aging, are important causes of pathogenesis [25],[26]. In contrast to OA, RA is a systemic autoimmune disorder, which manifests as chronic inflammation that results in destruction of cartilage and bone tissues [2]–[4]. The inflammatory process initially affects a single joint, but the disease usually progresses to affect nearly all joints [27]. Thus, our results indicate that, despite their different etiologies and pathogenesis, both RA and OA are regulated by HIF-2α via completely different mechanisms: HIF-2α regulates OA pathogenesis by up-regulating matrix-degrading catabolic enzymes in articular chondrocytes, whereas it appears to regulate RA pathogenesis by regulating angiogenesis, various functions of FLS, and IL6-dependent TH17 cell differentiation.
In summary, our current studies suggest that HIF-2α is an essential catabolic regulator of RA pathogenesis that acts by modulating various RA-associated FLS functions. Because the etiology of RA pathogenesis has not yet been entirely elucidated and effective treatment of RA remains a significant unmet medical need, HIF-2α may serve as an effective therapeutic target in RA treatment. In this context, an important question that remains to be evaluated is whether recently developed small-molecule inhibitors of HIF-2α [28],[29] inhibit RA pathogenesis in vitro and in vivo. Additionally, because HIF-1α and HIF-2α appear to regulate RA pathogenesis through different mechanisms, both HIF isoforms could be alternative therapeutic targets in the treatment of RA disease.
The use of human materials was approved by the Institutional Review Board of Chonnam National University Hospital and Wonkwang University Hospital, and written informed consent was obtained from all individuals before the operative procedure. Mice were housed in specific pathogen-free barrier facilities and were used in accordance with protocols approved by the Animal Care and Ethics Committees of the Gwangju Institute of Science and Technology.
Human RA, psoriatic arthritis, gouty arthritis, and OA joint tissues were collected from patients undergoing knee arthroplasty (Tables S1, S2, S3) and then embedded in paraffin. All RA patients had a median disease duration of ∼6 y, high disease activity (i.e., median DAS of 5.61), and received medications, including a variety of disease-modifying antirheumatic drugs (Table S1). Because joint tissues were obtained from patients undergoing knee arthroplasty, our samples represent relatively late-stage RA.
Male DBA/1J, C57BL/6, Epas1+/−, Epas1fl/fl, and Il6−/− mice were used for experimental RA studies. The C57BL/6 strains of Epas1+/−, Epas1fl/fl, and Il6−/− mice were described previously [18],[22]. Epas1+/−, Epas1fl/fl, and Il6−/− (C57BL/6) mice were backcrossed against the DBA/1J strain for eight generations to generate Epas1+/− DBA/1J, Epas1fl/fl DBA/1J, and Il6−/− DBA/1J mice, respectively. CIA was produced in WT and Epas1+/− DBA/1J mice using a standard protocol [16]. Briefly, mice were intradermally injected at the base of the tail with incomplete Freund's adjuvant alone (control) or Freund's adjuvant containing 100 µg of collagen type II; a booster injection was given 21 d later. Epas1fl/fl DBA/1J mice were IA-injected with Ad-C or Ad-Cre (1×109 PFU) on days 0, 3, and 6, followed by a booster injection with collagen type II. Mice were maintained for an additional 2 wk. The incidence and severity of RA were evaluated on the indicated days after the first immunization. Severity was evaluated using a clinical score (grade 0–4) of paw swelling based on the level of inflammation in each of the four paws [16]. Joint tissues from mice were fixed, decalcified with 0.5 M EDTA (pH 8.0), embedded in paraffin, and sectioned at 5-µm thickness. Synovitis was evaluated by H&E staining of joint sections, and synovial inflammation (grade 0–4) was scored as described by Tang et al. [30]. The pannus in joint tissues adjacent to cartilage and bone was visualized by H&E staining with or without safranin-O staining of cartilage, and pannus formation was scored (grade 0–4) as described by Tang et al. [30]. Cartilage destruction was examined by safranin-O staining and scored using Mankin's method, as previously described [18],.
Human and mice joint tissues were sectioned at 5-µm thickness for immunohistochemical staining. Antigen retrieval was performed by incubating sections with 0.1% trypsin for 40 min at 37°C or with citrate buffer for 20 min at 95°C. The following primary antibodies were used for immunohistochemistry: rabbit anti–HIF-2α and rabbit anti-RANKL (Santa Cruz), rabbit anti-MMP3 and anti-MMP13 (Abcam), goat anti-IL6 (R&D Systems), rabbit anti-IL17A (Abcam), mouse anti-HIF-1α (Sigma), and rat anti-CD31 (Dianova). For double-immunofluorescence labeling of human and mouse joint tissues, the following primary antibodies were used: rabbit anti–HIF-2α (Novus Biologicals for human tissues and Santa Cruz for mouse tissues), rabbit anti-MMP3 (Abcam), mouse anti-MMP13 (Calbiochem, for human synovia), rabbit anti-MMP13 (Abcam, for mouse synovia), goat anti-IL6 (R&D Systems), mouse anti-vimentin (BD Biosciences), rabbit anti-CD55 (Santa Cruz), rabbit anti-CD68 (Abcam, for human synovia), rat anti-CD68 (Abcam, for mouse synovia), rabbit anti-TRAP (Santa Cruz), rabbit anti-VEGF (Santa Cruz), rabbit anti-IL17A (Santa Cruz), rabbit anti-Ki67 (Abcam), and rabbit anti-RANKL (Santa Cruz). Expression levels of HIF-2α in RA synovium were quantified using Image J software. The percentage of cells expressing HIF-2α was analyzed in synovial lining cells (fibroblast-like and macrophage-like synoviocytes), sublining macrophages, and endothelial cells [32]. Cell types were distinguished according to their characteristic morphology and confirmed by immunoreactivity with anti-vimentin (FLS), anti-CD68 (macrophages), and anti-CD31 (endothelial cells) antibodies. For immunofluorescence staining of total synovial cells or FLS cultured on coverslips, the following antibodies were used: rabbit anti–HIF-2α (Novus Biologicals), goat anti-IL6 (R&D Systems), rabbit anti-CD68, and mouse anti-vimentin (BD Biosciences).
Total synovial cells were isolated from knee joint synovium of CIA mice. Synovial tissues were minced and digested in collagenase for 4 h at 37°C. The cells were plated on coverslips in RPMI-1640 medium and incubated for 4 d. FLS were isolated from NI and CIA joint tissues of WT and Il6−/− mice [33]. FLS between passage 4 and 8 were used for further analysis. Pure FLS (>90% CD90+/<1% CD14+) were identified by flow cytometry using antibodies against the fibroblast marker CD90 and the macrophage marker CD14 (Abcam). For the preparation of CM, FLS were infected with Ad-C or Ad-Epas1 at a multiplicity of infection (MOI) of 800 for 2 h and incubated on 35-mm culture dishes containing 1 ml of RPMI-1640 medium. CM was used to treat CD4+ precursor T cells during differentiation into TH17 cells. FLS proliferation in culture was quantified by measuring BrdU incorporation during DNA synthesis. Proliferating cells in synovial sections were identified by detecting Ki67 using an antibody obtained from Novus Biologicals.
CD4+ T cells from WT and Epas1+/− mice were purified from lymph nodes and spleens. TH cell differentiation was induced by plating cells (2×106 cells/ml) on culture dishes coated with anti-CD3 antibody (1 µg/ml) in the presence of soluble anti-CD28 antibody (2 µg/ml) under the following TH cell-skewing conditions: TH1 cells, IL12 (10 ng/ml) and anti-IL4 antibody (10 µg/ml); TH2 cells, IL4 (20 ng/ml) and 10 µg/ml of antibodies against IFNγ and IL12; TH17 cells, tumor growth factor (TGF)-β (3 ng/ml), IL6 (30 ng/ml), and 10 µg/ml of antibodies against IL4, IFNγ, and IL12. IL2 (100 U/ml) was added after 24 h, and cells were cultured for 6 d. Antibodies and cytokines were purchased from BD Biosciences or PeproTech. The cells were stimulated with PMA (50 ng/ml), ionomycin (1 µM), and brefeldin A (1 mg/ml; eBioscience). TH cell differentiation was evaluated by flow cytometry after staining for intracellular cytokines and by detecting cytokines by ELISA and qRT-PCR. Where indicated, activated CD4+ T cells were transfected by electroporation with empty vector or vector carrying Epas1. The cells were cultured under neutralizing conditions (10 µg/ml of antibodies against IL4 and IFNγ) or TH17 cell-skewing conditions for 4 d. The effects of HIF-2α overexpression on skewed CD4+ T cells were evaluated by detecting IL17A production by ELISA. For cell proliferation assays, CD4+ T and B220+ B cells were isolated from lymph nodes and spleens of WT and Epas1+/− DBA/1J mice. T-cell proliferation was induced by stimulating cells with anti-CD3 antibody (10 µg/ml), and B-cell proliferation was induced by stimulating cells with LPS (10 µg/ml), LPS plus IL4 (5 ng/ml), or antibodies against IgM (20 µg/ml; Jackson ImmunoResearch) and CD40 (10 µg/ml; BioLegend). Proliferation was assessed by measuring [3H]thymidine (0.5 µCi/well) incorporation during the last 18 h of a 72-h culture period.
For detection of the hypoxic status of mouse CIA synovium, mice immunized with type II collagen were intraperitoneally injected with hypoxyprobe-1 (pimonidazole HCl; Hypoxyprobe Inc.) at a dosage of 60 mg/kg body weight and sacrificed 6 h after injection. Paraffin-embedded joint tissues were sectioned at 5-µm thickness, and pimonidazole was detected by immunofluorescence microscopy, according to the manufacturer's instructions. For hypoxic culture of mouse FLS, cells were exposed to hypoxia for 12, 18, or 24 h in a GasPak anaerobic chamber (BBL GasPak Pouch; Becton Dickinson) at 37°C, as described previously [18]. The proportion of oxygen in each chamber was ≤1%.
Leukocytes were prepared from lymph nodes draining the inflamed joint, spleen, and thymus of NI and CIA mice. Synovial cells were harvested by digesting synovial tissues with collagenase. The cells were incubated with primary antibodies for 15 min at 4°C. Antibodies against CD4, CD8, CD44, Foxp3, B220, and CD11c were purchased from eBioscience; the anti-CD62L antibody was from BD Pharmingen. Nonspecific staining was ascertained using isotype-matched control antibodies. TH cells were fixed in fixation/permeabilization buffer for 30 min; resuspended in 100 µl of permeabilization buffer; incubated for 30 min at 4°C with Alexa 488- or phycoerythrin (PE)-conjugated anti-IFNγ (eBioscience), fluorescein isothiocyanate (FITC)-conjugated anti-IL4, PE-conjugated anti-IL17A or isotype control antibodies (eBioscience); and analyzed by flow cytometry using EPICS XL and EXPO32 software (Beckman Coulter).
Representative cytokines involved in RA pathogenesis (IFNγ, TNFα, IL4, and IL17A) and produced by TH subsets were detected using ELISA kits (eBioscience), according to the manufacturer's protocol. IL6 and TNFα secreted into serum-free culture media by FLS infected with Ad-C or Ad-Epas1 were quantified by ELISA. Collagen type II–specific antibodies were measured by ELISA. Sera from NI and CIA mice were added into 96-well plates coated with type II collagen (5 µg/ml), incubated overnight at 4°C, washed, and incubated for 1 h with alkaline phosphatase-labeled monoclonal antibodies against mouse IgG1, IgG2a, or IgG2b (Immunology Consultants Lab). Wells were developed using p-nitrophenyl phosphate as a substrate, and the resulting color reaction was quantified using an ELISA plate reader.
Bone marrow cell culture, osteoclastogenesis, and TRAP staining were performed as described previously [34]. Briefly, bone-marrow–derived macrophages were isolated from WT or Epas1+/− mice, seeded in 48-well plates (4×104 cells/well), and cultured for 4 d (Ad-C and Ad-Epas1 infection in WT cells) or 5 d (WT and Epas1+/− precursor cells) with M-CSF (macrophage colony stimulating factor; 30 ng/ml) and RANKL (100 ng/ml) to induce osteoclastogenesis. The surface area of TRAP-stained multinuclear osteoclasts containing three or more nuclei was measured using an Osteomeasure system (Osteometrics). TRAP activity was also determined in paraffin sections of joint tissues from NI and CIA mice or mice IA-injected with 1×109 PFU of Ad-C or Ad-Epas1. The numbers of TRAP-positive osteoclasts and their precursor cells were counted in a blinded fashion in all regions of pannus-formed bone–cartilage interface and synovium for each knee joint.
PCR primers and experimental conditions are summarized in Table S4. Two different siRNA sequences that silenced Epas1 effectively were used in this study (Table S5). Nonsilencing, scrambled siRNA was used as a negative control. Cells were transfected for 6 h with siRNA using Lipofectamine 2000 (Invitrogen), and then infected with Ad-Epas1 or treated with IL1β. In qRT-PCR, the relative levels of target mRNA were normalized to those of glyceraldehyde 3-phosphate dehydrogenase. The following antibodies were used for Western blotting: rabbit anti-MMP2, -3, -9, -12, -13, and -14 (Epitomics); rabbit anti-ADAMTS4 (Abcam); goat anti-IL6 (R&D Systems); rabbit anti-ADAMTS5 (Thermo Scientific); rabbit anti-iNOS, rabbit anti-VEGF, goat anti-RANKL, and goat anti-TNFα (Santa Cruz); and mouse anti-COX2 (Cayman Chemical).
The nonparametric Mann–Whitney U test was used for the analysis of data based on an ordinal grading system, such as synovitis, pannus, and Mankin scores. For results obtained in qRT-PCR assays, ELISAs, and analyses of blood vessel numbers, joint thickness, TRAP-positive cells, BrdU incorporation, thymidine incorporation, and apoptotic cell numbers, data were first tested for conformation to a normal distribution using the Shapiro–Wilk test and then were analyzed by Student's t test (pair-wise comparisons) or analysis of variance (ANOVA) with post hoc tests (multicomparison), as appropriate. Significance was accepted at the 0.05 level of probability (p<0.05).
|
10.1371/journal.pntd.0006175 | Rift Valley Fever: A survey of knowledge, attitudes, and practice of slaughterhouse workers and community members in Kabale District, Uganda | Rift Valley Fever virus (RVF) is a zoonotic virus in the Phenuiviridae family. RVF outbreaks can cause significant morbidity and mortality in humans and animals. Following the diagnosis of two RVF cases in March 2016 in southern Kabale district, Uganda, we conducted a knowledge, attitudes and practice (KAP) survey to identify knowledge gaps and at-risk behaviors related to RVF.
A multidisciplinary team interviewed 657 community members, including abattoir workers, in and around Kabale District, Uganda. Most participants (90%) had knowledge of RVF and most (77%) cited radio as their primary information source. Greater proportions of farmers (68%), herdsmen (79%) and butchers (88%) thought they were at risk of contracting RVF compared to persons in other occupations (60%, p<0.01). Participants most frequently identified bleeding as a symptom of RVF. Less than half of all participants reported fever, vomiting, and diarrhea as common RVF symptoms in either humans or animals. The level of knowledge about human RVF symptoms did not vary by occupation; however more farmers and butchers (36% and 51%, respectively) had knowledge of RVF symptoms in animals compared to those in other occupations (30%, p<0.01). The use of personal protective equipment (PPE) when handling animals varied by occupation, with 77% of butchers using some PPE and 12% of farmers using PPE. Although most butchers said that they used PPE, most used gumboots (73%) and aprons (60%) and less than 20% of butchers used gloves or eye protection when slaughtering.
Overall, knowledge, attitudes and practice regarding RVF in Kabale District Uganda could be improved through educational efforts targeting specific populations.
| Rift Valley Fever (RVF) virus is transmitted to humans from contact with infected livestock and through mosquito bites. Several human cases of RVF were diagnosed in Kabale District, Uganda in March 2016, over 40 years after the last RVF case was identified in Uganda. We administered a knowledge, attitudes, and practice survey to people living in Kabale District, near where the cases occurred. Survey results demonstrated that knowledge, attitudes and practice surrounding RVF could be improved within the community.
| Rift Valley Fever (RVF) virus is a single-stranded negative sense RNA virus of the genus Phlebovirus in the Phenuiviridae family [1,2]. RVF virus outbreaks can cause significant morbidity and mortality in animals and humans [3]. In sheep, goats, and cattle, infection with RVF can lead to increased abortions and stillbirths. RVF symptoms in humans range from asymptomatic or a mild flu-like illness to a more severe illness including hepatitis, retinitis, or encephalitis; approximately 1% of human cases develop hemorrhagic disease [4]. Case fatality is estimated at 1–2% [5], however during an outbreak in Saudi Arabia, case fatality was estimated to be as high as 17% [6]. RVF virus is maintained in infected mosquitoes, and transmitted to ruminants including cows, sheep, and goats via mosquito bites [3]. Mosquito vectors that have been implicated in RVF transmission include species from six genera: Aedes, Culex, Anopheles, Eretmapodites, Mansonia, and Coquilletidia [3,7]. These vectors have been identified in regions spanning the African continent [7]. Experimental models have demonstrated that North American mosquito species are competent vectors [8,9]. Humans primarily become infected with RVF from contact with blood or body fluids of infected animals through slaughtering, caring for sick animals, or assisting with animal birth. As a result, herdsmen and butchers are at increased risk of RVF infection due to exposure to blood and body fluids of infected animals [10–12]. Mosquito bites are another source of RVF transmission to humans. Consumption of raw meat and dairy is not a confirmed mode of RVFV transmission but is a risk factor for other zoonotic diseases such as brucellosis and listeria.
In agricultural communities, RVF outbreaks can cause significant economic losses [3]. The 2007 RVF outbreak in Kenya impacted agricultural production and employment, resulting in an estimated loss of $32 million USD, due to both direct effects from livestock death and indirect effects including reduced income from the closure of livestock markets and reduced sales of animal derived products [13]. Outbreaks typically occur after periods of heavy rainfall and flooding lead to increased mosquito populations; therefore it is possible to use geospatial analysis to predict areas with increased risk of RVF outbreaks [14–16]. Early identification of cases in livestock and humans is an important tool in outbreak detection and may aid in the rapid deployment of disease control measures, such as animal vaccines or mosquito control. There are no approved RVF vaccines for humans. Therefore, the main methods of disease prevention in people are early detection, awareness, and behavior modification such as the use of personal protective equipment and increased hand hygiene. Therefore it is important to ensure that at-risk populations in endemic regions are aware of RVF.
In March 2016, two human RVF cases were confirmed in Kabale District, Uganda [17]. These were the first RVF cases identified in Uganda since 1968 [18]. Following the identification of these two cases, an investigation was coordinated by the Uganda Ministry of Health, Uganda Ministry of Agriculture Animal Industry and Fisheries (MAAIF), the Uganda Virus Research Institute (UVRI), and United States Centers for Disease Control and Prevention (CDC) to respond locally to the outbreak and to assess the knowledge, attitudes, and practices (KAP) of community members living in the area. The main objective of a KAP study is to identify knowledge gaps in order to create programs and materials to address those needs. Although RVF KAP studies have been performed previously in Eastern Africa [19–23], we could not find published RVF KAP studies performed in Uganda.
Kabale District is located in the southwest corner of Uganda. According to the 2014 Uganda Census, Kabale had an estimated population of 534,160 people, with the majority living in a rural setting (457,592/534,160; 86%) [24]. The altitude of Kabale ranges from 1,219 m to 2,347 m above sea level. Agriculture is an important source of revenue in this region. Corn, beans, potatoes, bananas, millet, coffee, apples, and tea are all grown in the region. Livestock also provide a source of income for families through the sale of meat and dairy. Most families own goats, however sheep, cattle, and pigs are also common. In addition to agricultural lands, Kabale District also has areas with high altitude forest and savannahs.
A multidisciplinary team, consisting of individuals from the Uganda Ministry of Health, Uganda MAAIF, UVRI, and CDC, visited 34 sites from April 1–12 2016 in Kabale District, Uganda to recruit participants to complete a RVF KAP cross-sectional survey (Fig 1). Recruitment sites included the main abattoir in the town/city of Kabale, villages where a human RVF case had been identified, and areas where no human cases had been identified previously. The team arrived at a site and local health workers recruited participants from the village. Convenience sampling was used; all participants who arrived at the site were eligible for the study. Children older than seven years were allowed to participate if a parent provided written consent.
The survey was written in English but was read aloud in a private setting to participants by trained interviewers in the local language. The survey questionnaire asked participants about sociodemographic variables as well as epidemiological risk factors and exposures, including contact with animals and exposure to raw milk, raw meat, and mosquitoes (S1 Fig).
Occupation was self-reported and individuals described their primary occupation as a farmer, herdsman, butcher, or another occupation. Interviewers also asked about the use of personal protective equipment (e.g. masks, gloves, gumboots) when in contact with animals. Participants were asked questions about knowledge of RVF symptoms, transmission, and prevention. Participants answered questions about their attitudes towards the existence of RVF, welcoming RVF survivors into their community, and health seeking practices. Most questions were closed-ended. However, participants were asked open-ended questions about why they did not use a mosquito net, about fears and risks of contracting disease, methods to protect themselves from disease, and ways to prevent RVF transmission.
Data were analyzed using STATA 13.0 (StataCorp. 2013. Stata Statistical Software: Release 13. College Station, TX: StataCorp LP.). Differences in frequency distribution between groups were compared using Pearson’s chi-square for categorical variables. Results were considered statistically significant if the p-value was <0.05.
Approval for the study was obtained from UVRI Research Ethics Committee and the CDC Human Subjects Advisor group. The CDC National Center for Emerging and Zoonotic Infectious Diseases (NCEZID) Human Subjects Advisor group determined that this was part of an outbreak investigation and was classified as non-research. The NCEZID Human Subjects Advisory group was aware of the study protocol and consent process. Participants completed a consent form for the questionnaire that was translated from English into local languages by team members. Parents of minors consented to participation prior to interviewing. All participation was voluntary and participants did not receive any compensation.
A total of 657 participants were interviewed, most (40%) were males aged 20–49 years (Table 1). The mean age of participants was 40 years, with a range of 7 years to 90 years. Most participants were from sites where no RVF case had been identified (293/647; 45%). In addition, 238 (37%) of participants were recruited from sites where a previous RVF case had been identified, and 117 (18%) of participants were butchers who worked at the main abattoir in Kabale. Most participants owned domestic animals and had contact with goats (353/504; 70%), cattle (299/505; 59%), pigs (129/505; 26%), sheep (98/506; 19%), and poultry (91/505; 18%). Participants who owned livestock, had on average three cattle, four goats and four sheep. Few participants had contact with ducks (3/506; 0.6%) or rabbits (9/506; 2%).
Participants most frequently had contact with live animals during grazing (Table 2).
A third (33%) of participants also reported assisting with animal birth. Contact with dead animals was mainly through handling raw meat (77%). However, nearly half of participants said that they were involved in slaughtering or butchering (Table 2). Significantly more men than women were engaged in milking, birth assistance and slaughtering/butchering (Table 2). However, significantly more women assisted with animal grazing (Table 2). Few participants drank raw milk or ate raw meat.
Most (90%) of participants said that they had heard of RVF previously; a greater proportion of butchers (95%), herdsmen (94%), and farmers (92%), had heard of RVF compared to persons in other occupations (85%) (X2 = 10.4; p = 0.016). Most respondents had heard of RVF from the radio (77%).
Interviewers asked participants if they had knowledge of RVF symptoms in humans and animals, participants responded either “yes, no, or do not know” (S1 Fig). Most participants, regardless of their occupation, said that they could identify RVF symptoms in humans (Table 3 and Table 4). Clinical symptoms of RVF are listed in Table 3. The most common human symptom identified by participants was bleeding. Less than half of participants cited fever, vomiting, or diarrhea as RVF symptoms in humans. Of all participants, 241 (37%) had knowledge of RVF symptoms in animals (Table 3). Knowledge of RVF symptoms in animals was significantly higher among herdsmen and butchers. Bleeding in animals was the most common RVF symptom reported by participants, particularly among butchers, farmers, and individuals in other occupations. Herdsmen were significantly more likely to identify fever as an RVF symptom in animals. In contrast, butchers were more likely to identify nasal discharge as an RVF symptom in animals.
In general, knowledge of RVF symptoms did not vary significantly by gender (Table 4). However, more women than men identified bleeding as a sign of RVF in humans. Significantly more men stated that they were familiar with RVF symptoms in animals and significantly more men identified nasal discharge as a symptom in animals.
Of all participants, 53% (345/655) said that they had any knowledge about how RVF is transmitted to humans or animals. Participants most frequently identified animal contact as a mode of transmission (269/348; 77%). We asked participants if RVF could be transmitted by human to human bodily contact (e.g. handshaking) and through contact with human bodily fluids (e.g. blood). Of all participants who responded, 83 of 348 (24%) said RVF could be transmitted through mosquitoes and through contact with body fluids of an infected person (84/348). Of farmers, 76 of 115 (66%) identified animals as a source of transmission, compared to 80–93% of butchers, herdsmen, and persons in other occupations (p = 0.003). Some participants thought that RVF could be transmitted by human-to-human bodily contact (107/348; 31%) and through air (25/348; 7%). The belief that RVF could be transmitted by bodily contact was most common among farmers (40%), compared to butchers (20%), herdsmen (20%), and others (29%) (p = 0.021). Most participants identified goats (267/288; 93%), cattle (260/290; 90%), and sheep (203/289; 70%) as sources of RVF transmission. Although our survey did not specifically ask whether or not eating raw meat or drinking raw milk were risk factors, 38 participants mentioned these as other possible sources of RVF transmission.
Our questionnaire also asked about ways to prevent RVF transmission; most participants cited avoiding animals (257/612; 42%) and sick people (122/612; 20%) as modes of prevention. A small proportion of respondents (92/612, 15%) cited sleeping under a mosquito net as a method of prevention. Although our questionnaire did not specifically ask about boiling milk and cooking meat, participants frequently mentioned them as possible methods to prevent RVF transmission.
Mosquito nets were used by most participants (82%; 539/655). Of those 116 individuals who did not use a mosquito net, the most frequently cited reason was that nets were uncomfortable, which was cited by 17% of persons. Of all participants, 29% (149/510) said that they used personal protective equipment (PPE; e.g. gloves, mask, gumboots) when handling animals. Participant responses regarding the use of PPE varied by occupation and gender (Tables 5 and 6). Butchers used PPE significantly more often than persons in other professions. However, they mostly used gumboots and aprons. Few butchers used gloves.
Most participants (88%; 501/567) believed that RVF exists; this finding did not vary by gender but did vary significantly by occupation. Compared to herdsmen (90%) and farmers (95%), fewer butchers (83%) and persons in other occupations (85%) (p<0.05) believed in the existence of RVF. The most common reason for not believing RVF exists is that people were not aware of an RVF case near Kabale (23%; 36/156); 28% (177/631) of all individuals were aware of an RVF case. Significantly more males (30%) than females (22%) had heard of an RVF case (p<0.05). Most people believed that they are at risk of getting RVF (69%; 448/651). Significantly more butchers identified themselves as being at risk of RVF (88%) compared to persons in other occupations (60%), farmers (68%), and herdsmen (79%) (p<0.01). Similarly a greater proportion of males (74%) identified themselves as being at risk of RVF compared to females (59%) (p<0.05). When asked about the treatment of RVF, most individuals said that they would seek care from modern medicine (91%; 579/636). Most participants said that they would interact with an RVF survivor (75%; 461/613) and welcome them back into their community (73%; 478/652).
Our study found that in the agricultural community of Kabale District, Uganda, most survey participants had heard of RVF, although few human cases had ever been identified in Uganda, and none had been identified since the 1960’s. Most of these participants received their information regarding RVF from the radio. This may have been in part due to local efforts to increase RVF awareness recently through radio announcements following the detection of the two human cases.
We found that although 90% of participants had heard of RVF, many did not recognize the most common signs and symptoms of RVF in humans and animals. The most common symptom recognized in both humans and animals was bleeding; however, this is a rare symptom of RVF. Survey participants may be confusing RVF symptoms with symptoms of Marburg virus, since a Marburg virus cluster was identified in Kabale district in 2012 and because Marburg can present with more hemorrhagic symptoms [25]. Our findings are similar to previous findings in other studies done in Kenya and Tanzania, despite differences in our study populations [19,22]. Compared to similar studies in Kenya and Tanzania, our study population had higher education; nearly 55% of our participants had secondary educations compared to 88% of participants having no formal education in the Kenya study by Abdi et al [19]. Further, Kenya has more documented RVF outbreaks than Uganda, with a particularly large outbreak in 2006–2007 that significantly impacted the Kenyan pastoralist economy [13]. In the Abdi study, most individuals (92%) recognized hemorrhage as an RVF symptom and less than 40% recognized headache, myalgias and visual changes as RVF symptoms in humans. Our study and the Abdi study suggest that even in settings where RVF is endemic or where formal education is prevalent, misconceptions about RVF exist.
Recognition of symptoms and signs in animals and humans is imperative to prevent the spread of disease, especially in a community such as Kabale, where 60% of survey participants own animals. During the 2007 RVF outbreak in Sudan, Hassan et al. noted that an RVF outbreak was only recognized after human cases were identified, indicating that improved awareness of the disease in animals should be emphasized [26]. When at-risk communities have a heightened awareness of RVF symptoms in animals they may be able to improve surveillance and identify an outbreak early. Jost et al. conducted focus groups with pastoralists in Kenya and Tanzania during the 2006–2007 RVF outbreak and found that pastoralists not only recognized changes in weather patterns and mosquito swarms but also recognized RVF signs among animals and humans [21].
Many participants (47%) said that they did not have an understanding of RVF transmission. Nearly a third (31%) of all of our participants thought that RVF could be transmitted by bodily contact and 66% of farmers thought that RVF could be transmitted by animal contact. Additionally, mosquito bites were an under-recognized mode of RVF transmission (this was also noted in the Abdi study) [19]. Interestingly, in another study in Kenya by Owange et al, pastoralists felt that mosquitoes were a very important risk factor for RVF transmission in cattle [20]. Although our study demonstrated that participants did not consider mosquitoes to be an important risk factor for RVF, participants said that they frequently slept under mosquito nets, likely due to concern for malaria. Improving education regarding other mosquito-borne diseases, including yellow fever and RVF, can further stress the importance of prevention against mosquito bites.
Studies of Ebola and Marburg survivors in Uganda and elsewhere have demonstrated that many survivors experience stigma within their community [27]. RVF, unlike Ebola and Marburg, cannot be transmitted from human to human. Our study demonstrated that stigma is not commonly associated with RVF and that most community members would interact with survivors and welcome them into their community.
The objective of our study was to conduct a rapid, timely KAP study in order to inform an educational campaign, and as a result our study had several limitations. Because our sampling strategy was convenience based, our survey participants may not accurately represent the true population in Kabale district. However, given that we interviewed over 650 participants in the region, focusing in part on high risk populations, we believe that our study is still very useful. Many of our questions were yes/no or multiple choice questions and we did not conduct any focus groups, therefore our results are mainly descriptive. Some of the questions regarding RVF symptoms were not specific to RVF and could overlap with other diseases including Marburg.
Since there is no currently approved RVF vaccine for use in humans [28] and no available veterinary RVF vaccine in Uganda, prevention of RVF is primarily through education. Educational programs should encourage the use of PPE during care of sick animals and slaughtering, discuss ways to decrease mosquito bites, and promote safe consumption of meats and dairy. In addition, heightened awareness of symptoms in humans and animals may lead to more rapid identification of outbreaks and may result in decreased transmission. Our study showed that RVF knowledge could be improved in Kabale district, Uganda.
Following the analysis of the survey, the MAAIF, Uganda Ministry of Health and CDC worked together to develop RVF educational materials targeting abattoir workers, farmers, herdsmen, and other community members (S2 Fig).
The posters emphasized that signs and symptoms of RVF more commonly include diarrhea, vomiting, and fever. These posters were created to target a low literacy audience using culturally appropriate messaging highlighting key findings from the KAP study, including disease transmission, signs and symptoms of RVF in humans and animals, and safe cooking practices [29]. For example, a poster targeting farmers and herdsmen contains images of symptomatic animals and humans and advises that farmers and herdsmen contact veterinary staff if their animal is ill. These materials have been translated to French for use in Niger after the identification of RVF cases in 2016 [30]. The CDC Viral Special Pathogens website also serves as a resource for information about previous outbreaks and RVF symptoms and diagnosis [31].
|
10.1371/journal.pbio.2002214 | The in vivo structure of biological membranes and evidence for lipid domains | Examining the fundamental structure and processes of living cells at the nanoscale poses a unique analytical challenge, as cells are dynamic, chemically diverse, and fragile. A case in point is the cell membrane, which is too small to be seen directly with optical microscopy and provides little observational contrast for other methods. As a consequence, nanoscale characterization of the membrane has been performed ex vivo or in the presence of exogenous labels used to enhance contrast and impart specificity. Here, we introduce an isotopic labeling strategy in the gram-positive bacterium Bacillus subtilis to investigate the nanoscale structure and organization of its plasma membrane in vivo. Through genetic and chemical manipulation of the organism, we labeled the cell and its membrane independently with specific amounts of hydrogen (H) and deuterium (D). These isotopes have different neutron scattering properties without altering the chemical composition of the cells. From neutron scattering spectra, we confirmed that the B. subtilis cell membrane is lamellar and determined that its average hydrophobic thickness is 24.3 ± 0.9 Ångstroms (Å). Furthermore, by creating neutron contrast within the plane of the membrane using a mixture of H- and D-fatty acids, we detected lateral features smaller than 40 nm that are consistent with the notion of lipid rafts. These experiments—performed under biologically relevant conditions—answer long-standing questions in membrane biology and illustrate a fundamentally new approach for systematic in vivo investigations of cell membrane structure.
| The structure and organization of the cell membrane are central to many biological functions, and although they have been extensively studied, there is still much that we don’t understand. A wealth of detailed information has been obtained from studies of model lipid membranes. However, these systems are often highly simplified in composition and always lack the active processes found in cells. The major obstacle to studying membranes in vivo in living cells has been that high-resolution physical methods used to investigate membrane structure are not compatible with living organisms. To overcome this obstacle, we employed a new in vivo approach that allows membrane structure to be observed directly in the gram-positive bacterium B. subtilis. This approach relies on tuning the isotopic content of hydrogen within the membrane, and other parts of the bacterium, to generate neutron scattering spectra exclusively from the membrane. Using this approach, we confirmed that the structure of the B. subtilis membrane is lamellar and has an average hydrophobic thickness of 23.9 ± 0.9 Ångstroms (Å). We were also able to observe nanoscopic lateral membrane structures, consistent with the notion of lipid rafts. This experimental approach will allow for a wide range of future structural studies of the cell membrane (and possibly other classes of biomolecules) without the need for extrinsic probes or labels. In this way, it fundamentally changes the scope of nanoscale structural questions that can be addressed in vivo.
| Because of the inherent complexity of living cells, high-resolution analytical methods are generally used ex vivo, and many depend on labels to differentiate the species of interest. In the case of cellular membranes, their nanoscopic structure (ultrastructure) has been elucidated primarily through electron microscopy (EM) with heavy-atom labels[1], supplemented with in vitro biophysical studies of model membranes.[2–12] In vivo membrane studies at lower resolution, notably with fluorescence microscopy, have provided additional details of lateral structure as well as important information on diffusion and other dynamic processes.[13] Through these techniques, a vivid picture of the membrane has emerged.[14–19] Nonetheless, they have significant limitations, and fundamental questions about the ultrastructure of the membrane in vivo remain unresolved.[20]
For example, membrane hydrophobic thickness is a basic structural parameter with implications for membrane transport, as well as membrane protein folding and function. Yet the hydrophobic thickness of membranes has never been determined in vivo. Another outstanding question concerns the existence of nanoscopic domains, or lipid rafts. The lipid raft hypothesis[21] invokes lateral organization of membrane lipids and proteins into distinct domains in the plane of the membrane to facilitate the assembly and regulation of multi-molecular complexes. This hypothesis provides a compelling rationale for numerous observations relating to membrane trafficking, endocytosis, signal transduction, and other processes.[22–25] However, evidence for lipid domains has been largely inferential, and it is now widely believed that these domains are nanoscopic as well as transient[26,27], making them difficult to detect. New methods to interrogate lateral structure in vivo—ideally without the use of extrinsic probes—would help us to understand the role of lateral heterogeneity in the many cellular processes where it has been implicated.
In this context, small-angle neutron scattering (SANS) has emerged as a uniquely powerful tool for the study of lipid bilayer structure in vitro. Neutrons have wavelengths on the order of Ångstroms (Å) and are thus inherently nanoscopic probes, well-matched to the dimensions of membrane structure. Using appropriately designed experiments, both transverse (normal to the membrane plane)[7] and lateral (within the membrane plane)[28,29] structure can be accurately determined. Importantly, neutron-scattering techniques do not require extrinsic molecules or heavy atoms as labels and rely instead on hydrogen (H)/deuterium (D) isotopic substitution. Cold and thermal neutrons are also nondestructive[30–32], making them ideal for the study of living systems. However, applications of powerful neutron-based ultrastructural methods have been restricted to in vitro membrane models due to the complex scattering signal arising from cells as a result of their diverse biomolecular composition.
Scattering and imaging experiments require that the feature(s) of interest have an observable signal, contrasted from the background. Where contrast is insufficient in the native system, it must be imparted through the use of labels. In the case of fluorescence imaging, contrast arises from the distinctive excitation and emission properties of native, or more commonly, introduced fluorophores. With X-rays and electrons, scattering contrast arises from differences in electron density, which scale with atomic number. Neutrons, on the other hand, are scattered by atomic nuclei, and the key contrast parameter—analogous to electron density—is the scattering length density (ρ). Neutron scattering lengths (b) are unrelated to atomic number and, in fact, vary among isotopes of the same element (S1 Fig). Most importantly, the scattering lengths of hydrogen (bH = −3.74 fm) and deuterium (bD = 6.67 fm) are substantially different.[31] Thus, neutron contrast is unique in that it can be varied in hydrogen-rich biological systems with H/D isotopic labels, as opposed to the heavy-atom labels used in X-ray and electron scattering, or fluorescent labels used in microscopy.
Different classes of biomolecules (i.e., proteins, lipids, carbohydrates, and nucleic acids) have different elemental compositions, which gives each class a different value of ρ (Fig 1a and 1b; Table A in S1 Text). Multicomponent systems can be designed such that 2 or more components have the same ρ, for example, protein in approximately 42% D2O/H2O. In this case, the protein and solvent are said to be contrast-matched, and the protein does not generate a distinct scattering signal—i.e., it is effectively invisible to neutrons.[30] If a third component is introduced having a different ρ (e.g., DNA), it then becomes the only contributor to the net scattering. Through judicious H/D-labeling of the sample and the aqueous medium, ρ for the different biomolecules can be tuned to enhance or attenuate their scattering. Thus, contrast can be varied without changing the chemical composition of the system.
In this report, we describe a chemical–biological approach to controlling neutron contrast variation in vivo and its application to the study of cell membrane ultrastructure. As a platform, we chose the gram-positive bacterium B. subtilis. This organism has a number of attractive features, such as being genetically tractable, having a well-characterized lipid metabolism, and growing readily in deuterated media. Importantly, it has a single membrane and uses only saturated fatty acids (FAs) that can be prepared readily in deuterated form to allow tuning of membrane contrast. After suppressing cellular contrast through global D-labeling (i.e., replacing most of the H with D throughout the cell), we selectively reintroduced contrast to the membrane by supplying H-FAs, which the cell incorporated. In this way, we were able to isolate the SANS spectrum of the membrane, which showed it to be a lamellar structure with a hydrophobic thickness of 24.3± 0.9 Å. A modification of the contrast-labeling scheme led to the observation of nanoscopic lipid structures of <40 nm within the plane of the cell membrane, providing experimental evidence for the existence of nanoscopic lipid domains (lipid rafts) in an active biological membrane.
The neutron scattering signal from a cell reflects the sum of contributions from all cellular components, here, taken to be water, protein, RNA, DNA, carbohydrate, and lipid (Fig 1a). Each of these has a characteristic ρ (Fig 1b), and from their relative abundances in B. subtilis,[33] we estimated the total scattering from the organism as a function of D2O concentration in the aqueous environment. Water accounts for approximately 85% of the cellular volume and exchanges rapidly across the membrane. Therefore, immersion of cells in deuterated buffer changes the neutron contrast, ergo the net scattering, with the relative contributions of the different molecular species varying as a function of D2O concentration (Fig 1c; S2 Fig).
We then measured the total scattering from B. subtilis cultured in standard M9 minimal medium (H-M9) and resuspended in phosphate buffered saline (PBS) at different D2O concentrations. The observed scattering corresponded closely to predictions of the simple compositional model, as shown in Fig 1c. Note that the cells scatter strongly over the entire range of percent D2O, and the total signal at any concentration of D2O reflects contributions from the multiple classes of biomolecules present in the cell. Structural analysis based on the overlapping signals is impractical, if not impossible.
To isolate a meaningful scattering signal from the membrane, it was first necessary to suppress neutron contrast from the entire cell by manipulating its H/D composition. This objective was achieved by culturing B. subtilis in deuterium-enriched M9 minimal medium (D-M9, prepared in 90% D2O with H-glucose as the carbon source). Due to metabolic H/D exchange, deuterium from the growth medium becomes permanently incorporated into the carbon skeletons of biosynthetic molecules. Overall, skeletal deuteration was approximately 70% (S4 Fig, Tables B and C in S1 Text), which was predicted to create a near-contrast matched condition for all biomolecules when the cells were immersed in approximately 85% D2O buffer (Fig 2a). This expectation was borne out in the scattering experiment, where a strong reduction in the total scattering was observed at approximately 85% D2O (Fig 2b, open circles; S2 Fig).
With contrast suppressed, the next step was to reintroduce contrast specifically into the membrane by providing exogenous FAs for incorporation into the membrane phospholipids. Membrane FAs, after conversion to FA methyl esters (FAMEs), are readily analyzed for structure and isotope substitution by gas chromatography/mass spectrometry (GC/MS). As shown in Fig 2c (top), B. subtilis uses a mixture of 7 main FAs, all of which are saturated (have no double bonds), and all but one of which are branched.[34,35] Initial supplementation experiments with wild-type B. subtilis showed that exogenous FAs were not incorporated into the membrane lipids intact, so the native pathways for both catabolism and anabolism of FAs had to be blocked (Fig 2d).
Catabolism was blocked genetically by the deletion of yusL, the gene encoding a critical enzyme in β-oxidation, enoyl-CoA hydratase.[36] Anabolism was blocked chemically using cerulenin, an irreversible inhibitor of β-ketoacyl-ACP synthase that suppresses de novo FA biosynthesis.[37] This combination resulted in conditional FA-dependent growth (demonstrated in S5 Fig). Cerulenin-induced growth inhibition can be rescued with a mixture of just 2 of the native complements of FAs—palmitic acid (normal-hexadecanoic acid [n16:0]) and 12-methyltetradecanoic acid (anteiso-pentadecanoic acid [a15:0]).[38] These 2 constitute a minimal set of 1 high-melting (n16:0) and 1 low-melting (a15:0) FA, with which the cell can regulate the fluidity and structure of its membrane. Unlabeled (H) and perdeuterated (D) forms of these 2 FAs were then used in various mixtures to tune neutron contrast in the membrane.
Growth of B. subtilis ΔyusL in D-M9 medium did not alter the FA composition of the membrane (Fig 2c, upper and middle panels; see S4 Fig for peak assignments and Table B in S1 Text for peak areas and deuteration analysis). However, when cerulenin-treated ΔyusL cells were grown in the same D-M9 medium, supplemented with H-n16:0 and H-a15:0, their membrane FAs were found to consist exclusively of these 2 FAs, with no deuterium incorporated from the medium (Fig 2c, bottom panel). The total neutron scattering from these cells showed a large increase in scattering at 85% D2O (marked by the blue arrow in Fig 2b), which can be attributed entirely to the H-FAs incorporated into the membrane phospholipids.
Having imparted contrast to the membrane (Fig 3a), we used SANS to determine the transverse membrane structure by recording the scattered intensity, I(q), as a function of the scattering wavevector, q. Cells for this experiment were cultured as described above in D-M9 medium, supplemented with cerulenin and H-n16:0 and H-a15:0 to label the membrane, then transferred to 85% D2O buffer. Because these are more time-consuming experiments, the cells were resuspended in an 85% D2O buffer supplemented with glucose, Mg2+, and cerulenin. These additives prevent autolysis and preserve the membrane potential,[39,40] such that cells in suspension remained >90% viable over a period of 4 h at 25°C, as determined by direct cell counts, optical density measurements, and live/dead staining (S6 and S7 Figs). The residual background was recorded using cerulenin-treated ΔyusL cells, which were fed a mixture of FAs contrast-matched to 85% D2O (a15:0 and n16:0, each 30% H and 70% D), and which do not contribute substantially to the net neutron scattering signal.
In SANS, the shape of the I(q) versus q data describes the structure of the sample on the order of Ångstroms to approximately 100 nm. Formally this data is modeled using a lamellar form factor representing the acyl core of the bilayer, with the 2 identical head group regions on either side, contrast-matched to the scattering density of the surrounding cellular environment. The lamellar form factor is characterized by a q−2 dependence at low-q, arising from the 2D membrane surface of the entire bacterium, transitioning to a q−4 dependence typical of 3D objects with smooth surfaces and sharp interfaces—more detail is available in the literature[41–44] and the materials and methods.
Subtraction of the background from the sample scattering revealed a pure membrane spectrum, which displayed a lamellar form factor characteristic of a lipid bilayer (Fig 3b), with the expected q−2 dependence at low-q (dashed line). A fit of the data (solid red line) revealed the average membrane hydrophobic thickness (2DC) to be 24.3 ± 0.9 Å at 25°C (S8 Fig). The hydrophobic thickness of the living B. subtilis membrane is thus comparable to that of synthetic phosphatidylcholine (PC) bilayers, such as dimyristoyl PC (2DC = 25.7 Å at 30°C) or 1-palmitoyl-2-oleoyl PC (2DC = 28.8 Å at 30°C).[46] It is also comparable to that of purified basolateral plasma membranes from rat hepatocytes (approximately 26 Å), as measured by small-angle X-ray scattering,[6] highlighting the conservation of membrane structure across animal and eubacterial kingdoms.
Finally, we sought to examine the lateral structure within the membrane to determine whether the lipids in the B. subtilis membrane are uniformly mixed or display nanoscopic organization, as predicted by the lipid raft hypothesis.[21] Canonical mammalian lipid domains are believed to be enriched in the high-melting lipids, cholesterol and sphingomyelin. Bacteria generally lack these lipids, but are nonetheless believed to have lipid domains[47–53] formed from lipid species playing analogous roles. [54,55] The expected hallmark of lipid domains, in any system, is the lateral separation of higher- and lower-melting lipids with associated proteins into distinct phases.
We recently introduced a contrast-variation strategy to detect lateral lipid organization in synthetic vesicles using SANS[28,29]. This strategy relies on differential H/D labeling of the lipid phases to control contrast in the plane of the membrane. With a suitable labeling strategy, the separated phases can be either contrasted or matched with respect to each other and the buffer. A particularly important case is where a mixture of lipids has an average contrast matching that of the buffer, but partitions into H- and D-enriched phases that contrast each other and the buffer. Experimentally, uniform mixing creates a null-scattering condition, whereas phase separation in the plane of the membrane (domain formation) induces neutron contrast and an observable signal in the SANS spectrum. In S9 Fig and Table F in S1 Text we present a series of similar SANS experiments which show how neutrons are sensitive to contrast in the plane of the bilayer and how neutron contrast can be controlled by thermally induced mixing of the 2 phases or by selecting specific isotopic mixtures which result in contrast-matched phases.
The in vivo adaptation of this strategy compares scattering from cells with 2 different H- and D-FA mixtures that are, on average, contrast-matched to the medium (Fig 4a). In the control mixture, there can be no contrast regardless of whether or not the membrane lipids are uniformly mixed, because all species are present at the same H/D ratio. However, in the experimental mixture, de-mixing among lipids creates contrast as described above, producing a measurable increase in the scattered intensity as a result of local inhomogeneities in the H/D distribution.
We implemented this strategy in cerulenin-treated B. subtilis ΔyusL cells using an experimental mixture of 100% D n16:0 (high-melting) and 40/60 H/D a15:0 (low-melting). When corrected for the relative abundances of these FAs in the cell (S10 Fig), this mixture creates an average membrane contrast matching the 30/70 H/D FA ratio (a15:0 and n16:0, each 30% H and 70% D) used for the control mixture (inset to Fig 4b) as well as the rest of the cellular components, including the 85% D2O buffer. SANS spectra from B. subtilis fed the experimental mixture reproducibly displayed an excess scattering over the q range 0.015–0.2 Å−1 compared to cells fed the control mixture (results from a repeat experiment are shown in S11 Fig). As the only source of contrast in the experiment was the membrane H/D FA pool, this result supports the notion that there is lateral de-mixing of lipids containing the high- and low-melting FAs on a length scale ℓ of 3–40 nm (ℓ = 2π/q) (Fig 4c), consistent with the lipid raft hypothesis.
Understanding how the nanoscale structure of biological systems relates to function is a challenging, ongoing pursuit. One difficulty is that only a few probes are capable of directly interrogating structure at this scale: electrons, X-rays, and neutrons. Electrons have enjoyed the widest application in cell biology, and EM remains the single most powerful tool for studying cellular ultrastructure. With regard to the B. subtilis membrane, a cryo-EM study of sectioned, freeze-substituted cells provided a striking picture of the cellular architecture, including the cell wall.[56] However, the structure of the plasma membrane was not well determined, and its thickness was estimated at 66 ± 8 Å, a surprisingly large value. Our results complement this EM picture of the cell envelope by providing a high-resolution hydrophobic thickness determination obtained under physiological conditions.
X-ray and neutron scattering have been widely used for studying the structure of model membranes composed of defined lipid mixtures[3–5,28,29,46] or natural lipid extracts.[6–12] X-ray scattering has also been used recently for the ex vivo study of cellular membranes,[6] but its application to intact cells is confounded by the issue of background scattering from water and biomolecules. Neutron scattering uniquely provides a solution to the background problem in the form of isotopic contrast variation. We have shown here that in vivo contrast variation through metabolic labeling can effectively suppress scattering from the complex cellular milieu, while highlighting specific features of interest, even when they arise from minor components such as lipids (approximately 1% of cellular wet mass). Furthermore, the cold neutrons used for scattering experiments are well suited for studies on living cells because of their low kinetic energy (<0.025 eV) and their nonionizing character—in contrast to high-energy X-ray and electron beams (>5,000 eV).
Prior applications of neutron scattering in vivo have relied upon external solvent contrast, only, which in some cases has been sufficient to observe Bragg scattering from repeat structures in thylakoid membranes[57,58] and mitochondria.[59] These studies have not revealed membrane structure per se, but have provided information on the arrangement of closely packed membranes. By creating internal contrast, i.e., by differentially labeling specific cellular components, we have, for the first time enabled high-resolution ultrastructural measurements on a single membrane. In this work, we demonstrated the power of a chemical-biology–based approach to create selective internal contrast, thereby enabling high-resolution measurements of the in vivo membrane thickness.
Our in vivo contrast variation approach also provides a new tool to study lateral membrane structure in living cells. Neutron-based structural methods offer distinct advantages in that they report nanoscopic lipid structure directly, without the need for models or extrinsic probes. Indications of nonuniform mixing[60,61] within the plasma membrane emerged contemporaneously with the landmark fluid–mosaic model proposed in 1972,[16] the concept of membrane domains became well established by the mid-1970s,[62] and the lipid raft hypothesis was formalized in 1997.[21] Nonetheless, the existence of lipid domains has remained controversial,[63] and because they are believed to be both transient and smaller than the diffraction limit of light (200 nm), they eluded observation by conventional microscopic techniques. Recently, however, ultra-resolution fluorescence microscopy was used to identify diffusionally restricted islands on the scale of 20 nm in the plasma membrane of rat kidney epithelial cells.[6] Our observation of lipid segregation on a comparable scale in the plasma membrane of B. subtilis is consistent with the existence of analogous lipid domains in bacteria and supports the notion that nanoscopic lipid assemblies are an integral feature of biological membranes.
The critical barrier that has prevented application of high-resolution neutron scattering techniques in vivo was lack of a means to create internal contrast. In this work, we overcame the barrier and showed that B. subtilis is an ideal in vivo model system for the application of neutron contrast variation strategies. Through specific growth conditions and select genotypes, we were able to attenuate cellular contrast globally and precisely reintroduce contrast into the membrane. With the ability to control both the chemical and isotopic properties of the membrane lipids, we were able to interrogate both transverse and lateral membrane structure. The same general approach to selective contrast can potentially be extended to other biomolecules and model organisms for applications outside the membrane arena. More immediately, the in vivo experimental platform can be used to investigate the response of the plasma membrane to a diverse range of physical, chemical, genetic, and environmental stimuli. We anticipate that this capability will therefore prove valuable in many areas, such as antibiotic development, biofuel production, membrane protein function, and understanding the interplay between the membrane, cytoskeleton and cell wall in creating a protective, adaptable, multifunctional interface.
Deuterium oxide (99.9% D) and algal amino acids (unlabeled and uniformly D-labeled, with an isotopic purity of 98%) were obtained from Cambridge Isotope Laboratories. Palmitic acid (H-n16:0) and palmitic acid-d31 (D-n16:0) were obtained from Sigma–Aldrich. 12-Methyltetradecanoic acid (H-a15:0) was purchased from Sigma–Aldrich or prepared from s-butyl magnesium chloride (Sigma–Aldrich) and 11-bromoundecanoic acid (Sigma–Aldrich) according to the method of Baer and Carney[64] and purified by vacuum distillation. Perdeuterated D-a15:0 was prepared from H-a15:0 through 3 cycles of H/D exchange with D2O catalyzed by 10% Pt/C at 220°C as described by Yepuri et al.,[65] followed by chromatography on silica gel and vacuum distillation. The final product was chemically homogeneous and had an isotopic purity of 99%, as determined by analysis of its derived methyl ester by GC/MS. Cerulenin was obtained from Alfa Aesar (Tewskbury, MA) and stored in the dark at –80°C as a solid. FA-free bovine serum albumin (BSA, catalog number A8806) was obtained from Sigma–Aldrich (St. Louis, MO). All other materials were obtained from commercial suppliers and used as received.
B. subtilis 168 (parent strain) and a ΔyusL mutant (strain BKE32840) were obtained from the Bacillus Genetic Stock Center (The Ohio State University, Columbus, OH). The ΔyusL strain lacks enoyl-CoA hydratase/3-hydroxyacyl-CoA dehydrogenase activity and is severely deficient in its ability to degrade exogenous long-chain FAs;[36] therefore, this strain was used for all experiments where exogenous FAs were supplied in the cultivation medium. General growth and maintenance of B. subtilis was performed in either Luria–Bertani (LB) rich medium or M9 minimal medium with 2% glucose supplemented with 5 mM l-tryptophan. Solid media were prepared by the addition of 1.5% Bacto or Noble Agar (Difco) to LB or M9 medium, respectively. Erythromycin was added to 0.5 μg/mL for routine maintenance of BKE32840. Cultures were incubated at 37°C, and liquid cultures were aerated by shaking at 250 rpm. Where applicable, supplemental FAs were added to a final concentration of 8 mg/L each of a15:0 and n16:0 from 25 mg/mL stock solutions in ethanol, along with 10 g/L of FA-free BSA as a carrier to aid solubility. Cerulenin (10 mg/mL in ethanol stock solution) was prepared fresh and added immediately prior to inoculation. The required final concentration was determined empirically and varied by supplier and batch. For the work described here, cerulenin from Alfa Aesar was used at a final concentration of 50 μg/mL, which was sufficient to fully suppress endogenous FA synthesis, as judged by inhibition of growth, and rescued by exogenous FA (S5 Fig).
Partially deuterated cells were grown in M9, prepared using 90% v/v D2O and H-glucose. This medium produced approximately 60%–70% deuteration of the carbon skeletons in biosynthetic molecules (analysis described below). B. subtilis 168 and BKE32840 were adapted to growth in M9-Gluc + 5 mM l-tryptophan prepared with 90% (v/v) D2O by serial passage (1:100 inoculum) in media prepared with increasing concentrations of D2O (H2O:D2O 100:0, 50:50, and 10:90). Cerulenin-treated, FA-supplemented cells were grown starting from a culture of untreated, unsupplemented cells. The starter culture (OD600 0.8–1.0) was diluted 1:20 in cerulenin/FA medium to an OD600 of ca. 0.05, incubated for growth to an OD600 of 0.8–1.0, then passaged by dilution (1:20) into fresh medium. FA composition was monitored at each passage by GC/MS; 5–8 passages were required to clear the native FAs and adapt the cells to the FA-dependent condition. Control cultures without supplied FA (no-growth controls) were monitored in parallel for each experiment to ensure that the cerulenin remained active during the incubation period. Optical densities were determined at 600 nm using a Synergy Mx plate reader (BioTek, Winooski, VT, USA), using a 96-well microtiter plate and 300-μL well volumes.
For contrast experiments (Figs 1 and 2, and S2 Fig), B. subtilis 168 cells were resuspended in PBS (10 mM Na2HPO4, 1.8 mM KH2PO4, 137 mM NaCl, and 2.7 mM KCl), prepared either with H2O or D2O (H-PBS or D-PBS, respectively), then mixed in appropriate proportions. For all H cells (Fig 1c), 250 mL of fresh culture (OD600 = 1.5) in H-M9 medium was split into 2 parts, which were processed in parallel. Cells were harvested by centrifugation at 6000 × g for 20 min at 4°C, then washed by centrifugation/resuspension with 3 × 10 mL of H- or D-PBS, allowing 5 min for equilibration at each step. The washed cell pellets were then resuspended in 11 mL of the same buffer to provide a cell concentration approximately 50 mg/mL wet weight, equivalent to approximately 10 mg/mL dry weight. For deuterated cells (Fig 2), the same procedure was followed except that a 125-mL culture grown in D-M9 prepared with 90% D2O was used, providing a final cell concentration of approximately 5 mg/mL dry weight. Cell suspensions were loaded into quartz “banjo” cells (diameter 22 mm, path length 1 mm) for study by SANS. Measurements were made at 37°C, using a single detector position, and the total scattering was analyzed as described below.
For membrane structural analyses (Figs 3 and 4, S7 and S11 Figs), FA-fed B. subtilis 168 ΔyusL cells were grown in M9-Gluc prepared with 90% D2O. Cells from a 35-mL culture were harvested at an OD600 of 0.5, washed with 3 × 3 mL of PBS, prepared with 85% D2O, and resuspended in 1 mL of the same buffer, providing a final cell concentration of approximately 5 mg/mL dry weight. Because of the long data collection times required (up to 4 h), glucose (0.1% w/v), MgSO4 (10 mM), and cerulenin (50 μg/mL) were added to the final resuspension buffer, the pH was reduced to 6.8, and the measurements were made at 25°C to prolong cell viability and minimize autolysis.[39,40] As discussed below under cellular viability and integrity and shown in S6 and S7 Figs, cells remained >90% viable and displayed consistent SANS spectra over a period of 4 h under these conditions.
Lipids were extracted from cells and characterized for FA content by GC/MS of derived FAMEs (schematic description provided in S4 Fig). Total lipid extracts were prepared using a modification[66] of the method of Bligh and Dyer.[67] In brief, cells were pelleted by centrifugation at 6000 × g for 15 min, followed by 3 washes in 1% (w/v) NaCl. Samples were lyophilized in 10 mL glass test tubes with Teflon-faced screw caps, to each of which was sequentially added 0.5 mL of chloroform, 1 mL of methanol, and 0.4 mL of water, with vigorous agitation at each stage. This mixture forms a single phase and was left to stand for 18 h at room temperature with occasional agitation. After 18 h, phase separation was induced by the addition of 0.5 mL of chloroform and 0.5 mL water. The lipids were recovered from the lower chloroform phase by evaporation of the solvent in a new 10-mL glass test tube under an argon stream. FAMEs were prepared by acidic methanolysis of dried lipid extracts or intact cells.[68] Solvents, if present, were evaporated under a stream of argon prior to the addition of 1 mL of concentrated HCl/methanol (10% v/v). The test tube was then purged with argon, sealed, and heated to 85°C for 2 h. After cooling, 1 mL of water and 1 mL of hexane were added, and the contents were vortex-mixed. After phase separation, a portion (approximately 700 μL) of the upper phase was taken out for GC/MS analysis.
Cellular amino acids were analyzed by GC/MS, as described by Dauner and Sauer.[69] Cells from 10 mL of culture at OD600 ≈ 0.7 were harvested by centrifugation, washed 3 times with water by resuspension/centrifugation, and stored frozen. For analysis, cells were resuspended in 1 mL of water in a microcentrifuge tube and lysed by sonication. After centrifugation at 14,000 × g for 15 min, a 500-μL portion of the supernatant was transferred to a glass vial and mixed with 1.5 mL of 6 M HCl. Standard solutions were prepared from H- or D-algal amino acid mixtures and processed in the same manner. The vials were sealed with Teflon-lined caps and heated at 110°C for 24 h to hydrolyze protein, after which the volatiles were removed by rotary evaporation. The hydrolysates were resuspended in tetrahydrofuran (100 μL) and N-tert-butyldimethylsilyl-N-methyltrfluoroacetamide (100 μL) and then heated to 60°C for 1 h to produce tert-butyldimethylsilyl amino acid derivatives suitable for GC/MS analysis. Samples were diluted with hexane 1:10 (v/v) prior to analysis.
GC/MS analysis was performed using an Agilent 5890A gas chromatograph with a 5975C mass-sensitive detector operating in electron-impact mode (Agilent Technologies, Santa Clara, CA). The instrument was equipped with an HP-5ms capillary column (30 m long, 0.25 mm outside diameter, and 0.25 μm coating thickness) using helium at 1 mL/min as the carrier gas. Samples of 1 μL were introduced using splitless injection at an inlet temperature of 270°C. FAMEs were eluted using a temperature program of 2 min at 60°C; 20°C/min to 170°C; 5°C/min to 240°C; and 30°C/min to 300°C for 2 min. Derivatized amino acids were eluted with a temperature program of 2 min at 100°C; 10°C/min to 280°C for 2 min.; and 25°C/min to 325°C for 2 min. Peak assignment, integration, and mass spectral analysis were performed using the instrument's ChemStation Enhanced Data Analysis software and the NIST mass spectral database. Peaks for deuterated compounds were identified on the basis of retention times and spectral comparison with nondeuterated compounds. The extent of deuteration was assessed by determining the gain in molecular mass for parent ions of FAMEs (Table B in S1 Text) or of selected fragment ions for amino acids (Table C in S1 Text).
SANS data were collected on Bio-SANS[70] and the Extended Q-range Small Angle Neutron Spectrometer (EQ-SANS)[71] at the at the High Flux Isotope Reactor and the Spallation Neutron Source, respectively, both located at Oak Ridge National Laboratory. Two-dimensional scattering data from both instruments were reduced using the Mantid[72] software and normalized to a porous silica standard to establish an absolute scale, and corrected for pixel sensitivity, dark current, and sample transmission. Background scattering was subtracted from the 1D intensity versus q, which is defined as:
q= 4πsin(θ)/λ
(1)
where λ is the neutron wavelength and 2θ is the scattering angle relative to the incident beam. At EQ-SANS, the data were collected in 60 Hz mode with 2 instrumental configurations: 1.3-m sample-to-detector distance with 4–7 Å neutrons (q = 0.050–0.4 Å−1) and 4.0 m sample-to-detector distance with 10–13.4 Å neutrons (q = 0.009–0.07 Å−1), yielding a total q-range from approximately 0.009 to 0.4 Å−1. At BioSANS, 6-Å neutrons were used at 2 sample-to-detector distances, 2.5 m (q = 0.050–0.30 Å−1) and 15.3 m (q = 0.005–0.060 Å−1), yielding a total q-range from 0.005 to 0.30 Å−1. Collection times did not exceed 4 h, at which point cells were determined to be better than 90% viable, as shown below in cell viability and sample integrity (S6 and S7 Figs). The scattering data from the cells demonstrate no statistically significant change in scattered intensity after 4 h (S7 Fig).
Model lipid mixtures (S9 Fig) were produced using synthetic lipids from Avanti Polar Lipids (Alabaster, AL) and prepared as follows: lipids were dissolved in chloroform, dispersed as a film by evaporation in a 20-mL scintillation vial, and dried overnight under vacuum (>6 h). The lipid films were then rehydrated in the isotopically appropriate solvent (H2O, D2O, or a combination thereof) and subjected to 5 freeze-thaw cycles prior to extrusion through 100-nm pore-diameter track-etched polycarbonate membrane filters. The final concentration of vesicles for scattering measurements was 10 mg/mL.
Data used for the contrast experiments (Figs 1 and 2) were collected on EQ-SANS for 0.009 Å−1 < q < 0.06 Å−1. The data were evaluated as the Porod invariant:
Q*=∫q2I(q)dq=2πI(0)Vp
(2)
The proportionality of Q* to the total scattering intensity I(0) makes it a useful metric for comparison to structure-independent estimates of I(0) based only on composition and scattering length density, ρ. Vp is the Porod volume. To analyze the acyl thickness of the bilayer, the data were modeled using the expression:[73]
I(q)=NVVs2(ρm−ρs)2〈|F(q)|2〉
(3)
for an arbitrary number of bilayers, N, of volume, Vs, with scattering length density ρ s, in a solvent of ρm, where F(q) is the form factor describing the lamellar shape, and (ρm − ρs), is the contrast term. The scattering law used to model the data was the lamellar form factor representing the hydrocarbon core of the bilayer, with 2 identical head group regions on either side, with scattering densities matching that of the surrounding cellular environment. Fitting of the SANS data was performed in SASview.[74]
Estimates of total scattering from B. subtilis (Figs 1 and 2) were made as follows. A neutron scattering signal arises when there is a difference in neutron scattering length density, ρ, between a species, s, and the medium in which it resides in, m. The difference, (ρm−ρs), is called contrast, and the scattered intensity is proportional to its square. For any species,
ρ=(Σinbi)/V
(4)
where n is the number of atoms, b is the coherent neutron scattering length for each atom, and V is the volume of the species. Because each class of biomolecule (i.e., protein, lipid, carbohydrate, RNA, or DNA) has a nearly constant chemical composition and density, ρ is well approximated as a single value for each class (Table A in S1 Text).[30]
Deuterium labeling increases the ρ of an unlabeled molecule (ρH) because the scattering length b is greater for D than H (6.67 versus –3.74 fm).[31] In biomolecules, the hydrogen atoms can be assigned to 2 categories, those bound to carbon (CH), which do not exchange with water, and those bound to heteroatoms (XH, X = N, O, or S), which do exchange with water. The fractions in the 2 categories are consistent within each class of biomolecule, which allows for the definition of the terms, ΔρCD and ΔρXD, that represent the increase in ρ resulting from the complete deuteration of each category. For a deuterated biomolecule:
ρ=ρH+fXDΔρXD+fCDΔρCD
(5)
where ρH is ρ for the unlabeled (all-H) molecule, and fXD and fCD are the fractions of deuterium substitution in the XH and CH categories, respectively. Estimates (Figs 1b and 2a) were generated using Eq 5 and the values for ρH, ΔρXD, and ΔρCD in Table A in S1 Text. The differing slopes for each class of biomolecule reflect differences in ΔXD.
The total scattering from a cell, Icell(0), is the sum of the scattering contributions from all of the cell’s biomolecular species, given by the relation
Icell(0)∝Σsχs(ρm−ρs)2
(6)
where χs is the volume fraction of each species s, ρs is its neutron scattering length density, and ρm is the average neutron scattering length density of the medium (all species s, including intracellular water). When ρs = ρm, the species is contrast-matched and thus effectively invisible to neutrons, provided the medium is uniform. The scattering attributable to an individual species j is given by:
Ij(0)∝0.5χjΣsχs(ρj−ρs)2
(7)
In an undeuterated cell, the χj for biomolecules is sufficiently high and the ρj is sufficiently different that the surrounding medium is not truly uniform. As a result, the total scattering for a given species j reflects interspecies contrast and is not completely nulled at the nominal contrast matchpoint defined by the average of the medium (χj = χm). However, judicious deuteration can be used to converge ρ for water and most biomolecules (cf. Figs 1b and 2a), effectively suppressing the interspecies contributions.
Knowledge of the cell’s composition (Table A in S1 Text) allows one to estimate the total cellular scattering—broken down by biomolecular species as a function of deuteration in the solvent and in the CH skeletons of the various biomolecules (Figs 1c and 2d). Approximately 80% of the dry mass of a B. subtilis cell is made up of protein (53%), RNA (18%), DNA (2.6%), lipid (5.2%), and carbohydrate (2.8%).[33] The remaining mass—which was neglected in our estimates—consists of small organic molecules, such as amino acids, cofactors, and nucleotides plus inorganic material. These data were used with Eqs 6 and 7 to calculate the predicted total scattering in Fig 1c.
Estimated scattering for cells grown in D-M9 (Fig 2b) relied on analyses of lipids and proteins extracted from B. subtilis under relevant conditions to adjust ρ according to Eq 5. The net deuteration of the FA pool in the absence of cerulenin and supplemental FAs was 68.5%, with deuteration of individual FAs ranging from 66% to 71% (Table B in S1 Text). The deuteration of the amino acid pool was more variable and had a lower average deuteration of 60% (Table C in S1 Text), which is similar to what would be expected in Escherichia coli.[75] A value of 70% was assumed for other biomolecules, and the extent of deuteration at water-exchangeable positions was assumed to match that of the medium.
Cell viability was evaluated using optical density measurements, manual cell counts with a hemocytometer, and fluorescent live/dead staining with the BacLight Bacterial Viability Kit (Catalog number L7012, Molecular Probes, Eugene, OR). The live/dead assay uses a mixture of 2 fluorescent nucleic acid stains (SYTO 9 and propidium iodide), which stain live cells green and dead cells with compromised membranes, red. Fluorescence micrographs of stained cells were acquired with a Zeiss Axioskop 2 Plus and analyzed with ImageJ[76] to count cells using green and red fluorescence channels. Nonirradiated suspensions of the ΔyusL strain prepared identically to the samples used for neutron beam experiments were used as controls. Cell suspensions were incubated in a sealed cuvette at 25°C, and optical density (OD600) was recorded at 1-min intervals over a period of 24 h (S6a Fig). Direct cell counts and live/dead staining of the nonirradiated samples were performed immediately after processing and at 4 h and 24 h (S6b Fig). These experiments showed close correspondence among the 3 measures of viability and that cells remained 90% viable over 4 h and 50%–60% viable over the course of 24 h under the conditions of the experiment. Irradiated cell suspensions were taken from the neutron beam after data collection and allowed to decay for approximately 30 min, at which time they were subjected to a radiological survey. Radiological safety protocols do not permit timely removal of cell suspensions for microscopic examination. Instead, optical density measurements of the cells at 4 h and 24 h were carried out using a Shimadzu UV-2700 UV–Vis spectrophotometer (S6a Fig). FA stability was monitored over the course of the 4 h it took to collect a complete SANS data set using GC/MS as described above for nonirradiated samples (S7 Fig). Finally, cell stability in the neutron beam was assessed by repeating the SANS measurement 3.5 h after the cells were put in the beam showing no change in the scattered intensity (S7 Fig).
|
10.1371/journal.pbio.2000638 | Approach-Induced Biases in Human Information Sampling | Information sampling is often biased towards seeking evidence that confirms one’s prior beliefs. Despite such biases being a pervasive feature of human behavior, their underlying causes remain unclear. Many accounts of these biases appeal to limitations of human hypothesis testing and cognition, de facto evoking notions of bounded rationality, but neglect more basic aspects of behavioral control. Here, we investigated a potential role for Pavlovian approach in biasing which information humans will choose to sample. We collected a large novel dataset from 32,445 human subjects, making over 3 million decisions, who played a gambling task designed to measure the latent causes and extent of information-sampling biases. We identified three novel approach-related biases, formalized by comparing subject behavior to a dynamic programming model of optimal information gathering. These biases reflected the amount of information sampled (“positive evidence approach”), the selection of which information to sample (“sampling the favorite”), and the interaction between information sampling and subsequent choices (“rejecting unsampled options”). The prevalence of all three biases was related to a Pavlovian approach-avoid parameter quantified within an entirely independent economic decision task. Our large dataset also revealed that individual differences in the amount of information gathered are a stable trait across multiple gameplays and can be related to demographic measures, including age and educational attainment. As well as revealing limitations in cognitive processing, our findings suggest information sampling biases reflect the expression of primitive, yet potentially ecologically adaptive, behavioral repertoires. One such behavior is sampling from options that will eventually be chosen, even when other sources of information are more pertinent for guiding future action.
| Human decision-making often appears irrational. A major challenge is to explain why apparently irrational behavior occurs and what potential benefits it might have conferred for our evolutionary ancestors. A well-studied behavior in experimental psychology is “confirmation bias,” where we sample information that simply confirms what we already believe. In this study, we show that one factor giving rise to such information sampling biases is Pavlovian approach: our natural tendency to approach items that are associated with reward. We demonstrate three novel information sampling biases in a large-scale smartphone experiment with >30,000 human subjects. We examine how these three biases are related to Pavlovian approach, as quantified via an entirely independent economic choice task. We also show that, within our population, information sampling is a stable trait of an individual that is related to demographic variables such as age and education. Although irrational in the context of our task, we postulate that approach-induced biases in information sampling may have been adaptive over evolutionary history. They would drive organisms towards gathering information about locations that they will eventually engage with to obtain reward.
| Many spheres of human behavior depend upon gathering and understanding evidence appropriately to inform decision-making. Yet the best way to sample information is a nontrivial problem, necessitating deciding where to sample information [1,2], when to cease information gathering [3,4], and weighing up how such evidence should guide behavior [5,6]. Normative approaches can help address these questions [7], but their computational complexity renders them unlikely candidates for controlling behavior. Instead, these approaches can be better used as a basis for understanding limitations in cognitive processes and why biases emerge in human behavior [8,9].
A particularly well-studied bias is that of confirming one’s prior beliefs [10]. Inspired by classic rule discovery and falsification studies of Wason [11,12], explanations of confirmation bias frequently appeal to limits in hypothesis testing as their latent cause. Several alternative accounts have been proposed. The “positive test account” [13] posits that humans form beliefs about a particular hypothesis and subsequently selectively seek and interpret evidence in support of this rule rather than against it. Yet it has been pointed out that this strategy may be normative in situations where possible competing hypotheses to explain the data are sparse [14]. Other accounts suggest that humans are simply limited in the number of hypotheses they can consider at any given time [15].
It is widely acknowledged that humans are also subject to more primitive influences on behavioral control. Whilst these have been overlooked as a potential source of confirmation bias, they are known to impact upon information seeking in other domains. For instance, a primitive behavior present in several species is the observing response [16,17]. Here, animals select actions to yield information (reduce uncertainty) about the probability of receiving future reward, even when these actions have no bearing upon reward receipt. This can also be related to human preferences for revealing advance information about rewards when that information is immaterial to the task at hand [18]. Critical here is the notion that, in nature, advance information is typically valuable in guiding future action (unlike in the experimental tasks used to demonstrate these behaviors). Preferences for early temporal resolution of uncertainty [19] is thus conserved across humans and other species and persists in influencing behavior even when rendered instrumentally irrelevant.
These considerations led us to consider how other primitive behaviors might bias information sampling. A notable characteristic of reward-guided behavior in many species is that of Pavlovian approach. Animals show greater efficacy in learning approach, as opposed to avoidance, actions that will lead to the delivery of reward [20,21]. Humans are also subject to similar approach biases [22]. Pavlovian approach effects also spill over into the domain of attentional control, as stimuli previously ascribed a high value capture attention even when they are contextually irrelevant [23]. As the locus of attention is intimately linked to information sampling during choice [24], this raises the possibility that Pavlovian approach may similarly influence information search.
To test this idea, we examined gameplay data from a large-scale smartphone app [25] in which we manipulated several factors of interest whilst probing subjects’ information sampling behavior. In brief, subjects played a card game in which they paid to sample information from different locations prior to deciding which option was most likely to yield reward. A framing manipulation meant that in half of all gameplays, approaching (choosing) the “biggest” option would be rewarded, but in the other half, approaching the “smallest” option would be rewarded. Crucially, the information structure of the task was identical across these matched conditions, such that any effects on information sampling could be ascribed to our manipulation as to the option subjects were instructed to approach.
We compared observed behavior to predictions derived from a normative dynamic programming model that computes the expected value associated with a perfect model of the task, treated as a Markov decision process (see Materials and Methods and [26]). This enabled us to isolate three distinct biases in subjects’ information search that respectively influenced where information was sought, when information collection terminated, and how information was used to guide eventual choices. Each of these three biases can be considered as a form of “approach” behavior towards locations that are more likely to yield reward. Also relevant here is our recent parameterization of human Pavlovian approach behavior in an approach-avoidance decision model on a separate economic decision task [27]. We demonstrate that the prevalence of all three biases is related to the key parameter from this model.
Subjects played a binary choice game that involved paying escalating costs for information (by turning over playing cards) while gambling on which option was best based upon card values that were revealed (Fig 1A). There were six possible conditions that subjects might play (Fig 1B). Across three of these conditions, subjects’ objective was to identify the pair (row) of cards with the largest product (“MULTIPLY BIGGEST”), largest sum (“ADD BIGGEST”), or largest single card (“FIND THE BIGGEST”). Across the remaining three conditions, the objective was inverted, such that they now sought the row with the smallest product, sum, or single card.
At the beginning of each trial, all cards start face down. Subjects then touch the first card (randomly located) to turn it over at no cost. This enters Task Stage 1 (Fig 1A). One of the three remaining cards is made available to be sampled at a cost of 10 points, but subjects can alternatively make a guess (gamble on which option will be rewarded) at no cost. If they choose to sample, the value of the second card is revealed and they enter Task Stage 2. Either of the two remaining cards can then be sampled at a cost of 15 points, or subjects can again choose to make a guess at no cost. If they choose to sample again, they enter Task Stage 3. The last remaining card can be sampled at a cost of 20 points, or they may again guess at no cost. At any Task Stage, making a guess means that subjects enter the Choice Stage. Here, subjects choose which row they think will be rewarded, and all remaining cards are then turned face up. The subject wins 60 points if the gamble is correct and loses 50 points if incorrect, minus the points paid for information sampling. Card values ranged, with a uniform distribution (sampled with replacement), from 1 to 10, with “picture cards” removed from the deck.
On each gameplay, subjects were randomly assigned to play two short blocks (11 trials each) of two from the six possible conditions. The symmetry between the “approach big” and “approach small” conditions is crucial to our experimental design. Revealing a card of a particular value yields the same information content in both versions of the task (with the exception of the FIND THE BIGGEST and FIND THE SMALLEST conditions). This means that subjects’ information gathering behavior should, normatively, be matched across these conditions. The only behavior that should change is the final gamble made by the subject, which should reverse. By comparing across ADD BIG and ADD SMALL conditions, and across MULTIPLY BIG and MULTIPLY SMALL conditions, we could probe the influence of the approach direction (i.e., big/small) on information sampling behavior and vice versa.
The first question we asked pertained to Task Stage 1 (Fig 1A). Here, subjects decided whether to sample or guess based upon two variables: the information seen, i.e., the card value, and also the location where information was made available for sampling. We label the first row sampled as “row A.” In some trials, subjects were constrained to sample the next card from row A (“AA trials”), whilst in other trials they were constrained to sample from row B (“AB trials”).
As can be seen from the optimal dynamic programming model (Fig 2A), the card value and (to a lesser extent) the trial type influences the relative expected value of choosing to guess versus choosing to sample. The U-shaped function of the graph reflects an intuition that high- or low-valued cards are informative about the correct option to approach, making it more valuable to guess early. Mid-valued cards, by contrast, provide less information and make it more valuable to sample more information. The differential influence of AA versus AB trials is because the potential reduction in uncertainty depends upon the information that has already been revealed. Intuitively, on a MULTIPLY TRIAL where a 1 has been revealed, then sampling from row A again yields little information relative to row B, as it is already known that row A will have a low value (between 1 and 10). On a MULTIPLY TRIAL where a 10 has been revealed, then sampling from row A yields more information than row B as it reduces the range of possible row A values from between 10 and 100 to an exact value.
As expected, the dynamic programming model predicts identical behavior irrespective of the subject’s approach goal. As an example, consider revealing a 2 in the MULTIPLY BIG condition on an AA trial. This yields a probability of 0.764 that row B will be rewarded, and the expected value of guessing is therefore 0.764 * 60 + (1–0.764) * (-50) = 34. The expected value of sampling again from the A row, calculated using dynamic programming, is 28.8, and so the relative expected value of guessing is 5.2 (Fig 2A). Now consider seeing a 2 in the MULTIPLY SMALL condition. This now yields the exact same probability of 0.764 that row A will be rewarded. Hence the expected value of guessing remains 34. The expected value of sampling further information remains 28.8, and so the relative expected value of guessing remains 5.2.
In contrast with these model predictions, subjects’ actual behavior showed a systematic difference between MULTIPLY BIG and MULTIPLY SMALL conditions (compare circles and plus signs in Fig 2B, see also S1 Fig). Subjects became more likely to guess if they had seen evidence that supported them approaching row A rather than avoiding it. In MULTIPLY BIG, a high-valued card (6 or above) carries evidence for choosing row A. Subjects become more likely to guess than when seeing the same card in MULTIPLY SMALL. However, a low-valued card in MULTIPLY BIG (5 or below) carries evidence for avoiding row A. Subjects now become more likely to sample than when seeing the same card in MULTIPLY SMALL. This framing effect is seen most clearly when subtracting behavior in MULTIPLY SMALL from MULTIPLY BIG (Fig 2C).
The observed bias is one of approaching an option if positive evidence has been provided in support of that option; consequently, we term this “positive evidence approach.” We observed that positive evidence approach was more pronounced in AB trials than AA trials (Fig 2C). This is again consistent with our hypothesis, as subjects are less inclined to sample further information if it is available on a row they wish to avoid than if it is on a row they wish to approach.
To quantify positive evidence approach across our population, we used the following summary statistic:
∑i=610(pGuessbig,i−pGuesssmall,i)+∑i=15(pGuesssmall,i−pGuessbig,i)
where pGuessbig,i and pGuesssmall,i denote the average probability of guessing in MULTIPLY BIG and MULTIPLY SMALL, respectively, having revealed card value i. As there should be no difference in the probability of guessing across the two conditions, the expected value of this statistic from the normative model is 0. By contrast, the value of this statistic across our population was 0.42 in AA trials and 0.70 in AB trials. To estimate our confidence in this summary statistic, we recomputed it on 1,000 bootstrapped samples of 10,000 gameplays from our population. This yielded 95% confidence intervals of [0.30, 0.54] in AA trials and [0.62, 0.79] in AB trials. (Throughout the paper, we focus on the reporting of effect sizes and 95% confidence intervals rather than p-values, as our large sample size renders p-values less informative [28].)
Similar results are seen by comparing the ADD BIG and ADD SMALL conditions (S2 Fig; AA trials: mean 0.52, 95% CIs [0.40, 0.64]; AB trials: mean 0.79, 95% CIs [0.70, 0.89]). See also S3 Fig for FIND THE BIGGEST/FIND THE SMALLEST, which are not directly matched for information content.
It is also notable that, overall, subjects’ behavioral choices to sample information were similar to predictions arising from the optimal model (Fig 2B), although not identical (S1 Fig). This alone does not imply that subjects are implementing the optimal model. Instead, it may simply reflect the fact that relatively simple behavioral strategies will often recapitulate many features of more sophisticated strategies [29]. For example, one straightforward strategy would be to compare the value of the presented card to the average value, estimate the current degree of uncertainty in making a choice, and then use these values with a softmax transformation [30] to calculate a probability for selecting row A, selecting row B, or sampling further information. We consider this question further in a latter section of the paper and show that this can approximate the average behavior of subjects in the task without recourse to an optimal model.
We next asked how decisions to sample or reject information might influence subsequent choices. If subjects elected to guess at the first stage, they entered the Choice Stage, in which they gambled on which option would be rewarded (Fig 1A). In ADD BIG, the relative expected value of choosing row A over row B increases with the value of the first card (Fig 3A, blue line), while in ADD SMALL, it decreases with the value of the first card (Fig 3A, purple line). This was reflected in subjects’ choices in both sets of trials (Fig 3B; see S4 Fig and S5 Fig for other conditions). However, this decision arises on two different types of trial. The subject will either have just declined the opportunity of sampling information from the A row (on AA trials), or the B row (on AB trials). Our hypothesis was that information sampling depends upon the underlying approach value of an item. A corollary is that declining to sample an item reflects an underlying preference for not approaching it.
When we compared choice preferences on AA and AB trials for identical card values on the same condition, we observed that, across all six conditions, subjects showed a systematic shift towards being less likely to choose the option that had just been left unsampled. Hence, subjects presented with the same card value on an AA trial were more likely to choose option B than on an equivalent AB trial (Fig 3B). This effect was most pronounced near the point of subjective equivalence in subjects’ choices and is revealed most clearly by subtracting subjects’ choice behavior in AB from AA trials (Fig 3C).
We term this a “rejecting unsampled options” bias. To quantify rejecting unsampled options across our population, we used the following summary statistic:
∑i=110p(Choice=A)i,AB−p(Choice=A)i,AA
where p(Choice = A)i,AB denotes the probability of choosing row A having observed card i on an AB trial, and p(Choice = A)i,AA denotes the same probability on an equivalent AA trial. There is no difference in the choice that is presented to the subject between AB and AA trials, and the expected value of this statistic is therefore 0. The mean value of this statistic across the population was 0.21 in ADD BIG (95% CIs [0.10, 0.32]) and 0.30 in ADD SMALL (95% CIs [0.18, 0.41]).
We also found the rejecting unsampled options bias to be present across all the other conditions: MULTIPLY BIG (mean 0.24, 95% CIs [0.12, 0.35]; S4 Fig), MULTIPLY SMALL (mean 0.27, 95% CIs [0.15, 0.39]; S4 Fig), FIND THE BIGGEST (mean 0.22, 95% CIs [0.11, 0.35]; S5 Fig), and FIND THE SMALLEST (mean 0.21, 95% CIs [0.09, 0.34]; S5 Fig).
Our design enabled us to also investigate where subjects chose to sample information. At Task Stage 2 on AB trials, we could determine subjects’ relative preference for sampling from row A versus row B (Fig 1A). Here, different conditions have different predictions for which row is more advantageous to sample. For example, in both the ADD BIG and ADD SMALL conditions, sampling from either row yields exactly the same amount of information about which row might be rewarded. The optimal dynamic programming model predicts no relative advantage for sampling from row A versus row B (S6 Fig).
In both MULTIPLY BIG and MULTIPLY SMALL conditions, however, dynamic programming predicts that sampling from the row that currently has the higher-valued card will be more informative. The intuition behind this is that the range of possible outcomes on the row with the higher-valued card is greater, and so sampling further information on this row leads to a greater reduction in uncertainty than sampling the row with the lower-valued card. This is borne out in a heat map of model predictions, showing the difference in relative value from sampling from row A versus row B (Fig 4A). Importantly, these predictions are identical for both MULTIPLY BIG and MULTIPLY SMALL conditions. Somewhat counterintuitively, it is therefore more advantageous to sample from the row with the largest card even in MULTIPLY SMALL. (Note that this is different from the relative expected value of guessing versus sampling, which is shown in S8 Fig)
In contrast to model predictions, we found that subjects preferred to sample from the option that currently had the higher value in the MULTIPLY BIG condition alone (Fig 4B, left). In the MULTIPLY SMALL condition, they preferred to sample from the option that currently had the lower value (Fig 4B, right). The influence of this bias in subjects’ information sampling is revealed more clearly by subtracting behavior in MULTIPLY SMALL from that of MULTIPLY BIG (Fig 4C). Whereas the optimal model shows no difference between these two conditions (i.e., the entire heat map should equal 0), subjects reliably sampled information from the row that they currently sought to approach rather than avoid.
We term this bias “sampling the favorite.” To quantify sampling the favorite, we derived two statistics for trials in which subjects decided to sample a third piece of information. We calculated one “strong evidence” statistic for trials in which the “favorite” (the item that would eventually be approached) was clear. We defined this as trials where the difference in card values was 4 or greater in magnitude:
∑i=510∑j=1i−4(p(Sample=A)i,j,big−p(Sample=A)i,j,small)−
∑j=510∑i=1j−4(p(Sample=A)i,j,big−p(Sample=A)i,j,small)
where P(Sample = A)i,j,big refers to the relative probability of choosing to sample from row A over row B, when card i is presented on row A and card j presented on row B, on MULTIPLY BIG trials. The top row of the equation denotes trials where row A has a higher-valued card than row B, favoring approaching A in MULTIPLY BIG but approaching B in MULTIPLY SMALL. The converse is true for the bottom row.
We calculated a second “weak evidence” statistic for trials in which the “favorite” was less clear. We defined this as trials in which the difference in card values was between 1 and 3 in magnitude:
∑i=210∑j=max(1,i−3)i−1(p(Sample=A)i,j,big−p(Sample=A)i,j,small)−
∑j=210∑i=max(1,j−3)j−1(p(Sample=A)i,j,big−p(Sample=A)i,j,small)
Crucially, because the optimal model predicts identical values for sampling from row A versus row B on MULTIPLY BIG and MULTIPLY SMALL, the expected value for both statistics is always 0. In contrast, the value of the “strong evidence” statistic across our population was 12.35 (95% CIs [10.55, 14.01]), whilst the value of the “weak evidence” statistic was 7.21 (95% CIs [5.85, 8.48]).
Note this bias was also observed in the ADD BIG/ADD SMALL condition (S6 Fig), where the value of the “strong evidence” statistic was 11.93 (95% CIs [10.39, 13.51]), whilst the value of the “weak evidence” statistic was 6.86 (95% CIs [5.61, 8.14]). See S7 Fig for FIND THE BIGGEST/FIND THE SMALLEST, which are not directly matched for information content.
We consider that subjects are unlikely to be implementing dynamic programming when they perform the task, yet their overall behavior shows a surprising resemblance to model predictions (e.g., Fig 2B, Fig 3B). We therefore constructed a simpler model that describes subjects’ performance without recourse to dynamic programming.
In this model, subjects first compute the value of choosing each option by comparing the presented value to the average value of all possible cards. In ADD BIG trials, at stage 1, for example, this would be the following:
V(A)=β1(1stCardValue-<1stCardValue>)
V(B)=β1(<1stCardValue>-1stCardValue)
where <1stCardValue> is the expected value of the first card (5.5) and β1 is a free parameter. In ADD SMALL trials, we simply inverted the value of the each card, such that 10 became 1, 9 became 2, and so on.
We also considered an AB trial (e.g., Fig 4B), where after turning the second card, the values of option A and B become the following:
V(A)=β1(1stCardValue-2ndCardValue)
V(B)=β1(2ndCardValue-1stCardValue)
At both stages, we compute the degree of uncertainty, ω, in choosing either option:
ω=−(11+eV(B)−V(A)−0.5)2
This is then used to derive the value of sampling information from option A or option B:
V(sampleA)=β2+β3ω
V(sampleB)=β2+β3ω
The probability of each choice C being option o is finally generated using a softmax choice rule:
p(C=o)=eV(o)τ∑ieV(i)τ
where i indexes the entire set of available options at a given stage of the task. Hence, at stage 1 of an AA trial, it indexes the set {choose A, choose B, sample A}; at stage 1 of an AB trial, it indexes the set {choose A, choose B, sample B}; at stage 2 of an AB trial, it indexes the set {choose A, choose B, sample A, sample B}. Note that the model therefore assumes a three- or four-way decision at each stage of the game. Although in the structure of the task this was made as two sequential binary decisions (guess or sample, then select a row), it reflects the intuition that when subjects decide to guess, they already have internally committed to choosing a particular row.
We fit parameters β1, β2, β3, and τ, using maximum likelihood estimation (fminsearch in MATLAB) separately at stage 1 and stage 2. Fitting was repeated at 50 random seed locations to avoid local minima, and the model was fit to the average behavior of all subjects across each condition using a sum of squared errors cost function. We performed fitting separately for both ADD and MULTIPLY conditions; we do not consider FIND THE BIGGEST/SMALLEST conditions here, as these are not matched for information content. Note that this model does not explicitly feature terms for the costs associated with sampling information; instead, these are implicitly factored into the constant term β2.
This model captures the main features of the behavioral data (S9 Fig). At stage 1, it displays a U-shaped effect of card value on information sampling (as in Fig 2B) caused by the effects of choice uncertainty on the value of sampling information. It also displays a softmax choice curve (as in Fig 3B) between options A and B, matching subjects’ real choice probabilities between these two options. At stage 2, it displays choice probabilities between options A and B that again closely match subjects’ behavior.
However, because this model makes symmetric predictions for BIG and SMALL trials, it fails to capture the three biases described above (S9 Fig). We therefore adapted the model with three additional parameters to capture these biases (Fig 5). At stage 1, before entering values into the softmax choice equation, we captured the “rejecting unsampled options” bias (Fig 5D–5F) by adding an “approach bonus” (β4) to the value of the item that could not be sampled. This makes subjects more likely to choose this option if they do not sample information.
V(A)=V(A)+β41stCardValue[onABtrialsonly];
V(B)=V(B)+β4<1stCardValue>[onAAtrialsonly].
A natural consequence of this value modulation by β4 is that it also induces a change in the form of the uncertainty ω, which determines the value of sampling information (see above). Whereas previously this would have been symmetric around the average value of the first card, it now becomes asymmetric, but in opposite directions on “add big” versus on “add small” trials. On AB trials, this feature of the model is thus sufficient to also capture the “positive evidence approach” bias (Fig 5B and 5C).
On AA trials, however, the “approach bonus” alone predicts the opposite pattern of “positive evidence approach” to that observed in the data. We instead found that we could capture the “positive evidence approach” bias (Fig 5A–5C) by modulating the value of sampling option A:
V(sampleA)=β2+β3ω–β5(1stCardValue)[onAAtrialsonly].
Notably, we found increasing first card value had a negative influence on the value of sampling A, reflected by the negative sign in front of parameter β5. We infer that on AA trials (where option A is available to be sampled), subjects are more inclined to choose option A when it is of high value than to sample again from it. This parameter is not needed on AB trials, where “positive evidence approach” is already captured by the β4 parameter alone. It should also be noted that it captures another general feature of the data, which is that subjects are more likely overall to guess on AA trials than on AB trials (see S1 Fig). This is because β5 reduces the value of sampling on AA trials selectively.
Finally, at stage 2, we found that we could capture the “sampling the favorite” bias (Fig 5G–5I) by introducing a parameter that affected subjects’ propensity to sample from higher-valued cards:
V(sampleA)=β2+β3ω+β6(1stCardValue-2ndCardValue)
V(sampleB)=β2+β3ω+β6(2ndCardValue-1stCardValue)
Parameter fits for stage 1 and stage 2 for both ADD and MULTIPLY trials are given in S1 Table. To confirm that the additional parameters provided additional complexity to model fits without overfitting, we used 10-fold nested cross-validation. Parameters were fit using 90% of the data (training set), and then the cost function was calculated for the remaining 10% of the data (test set, not used to train the model). This process was iterated ten times using different portions of the data as the test set each time. At both stage 1 and stage 2, for both ADD and MULTIPLY conditions, the model with additional parameters provided consistently better fits to the test set than the reduced model (S2 Table).
The close fit between model predictions and subject behavior reveals that a far simpler framework (comparing a card value to the average expected value) can approximate an optimal dynamic programming model. Moreover, subjects’ approach-induced biases in information sampling can be readily parameterized within this framework. We anticipate that further, more refined models will be subsequently tested by downloading the raw behavioral data from Dryad [31].
One potential drawback of the proposed model is that each of the three biases is captured by a separate parameter rather than a single unifying mechanism driven by Pavlovian approach. We therefore tested whether these parameters are related to each other across the entire population of subjects—that is, whether they might show positive covariance with each other. To this end, we adopted an alternative model fitting strategy using a mixed effects analysis to describe population behavior. A mixed effects analysis contains population-level hyperparameters to constrain individual subject model fits (see reference [32] for details). An added benefit is that one can examine the covariance structure of these hyperparameters to explore how β4, β5, and β6 are related to each other. Importantly, we found that, after model fitting, the covariance between all three parameters was positive. We normalized by the variance of each parameter to yield correlation coefficients between β4, β5, and β6; this yielded a positive correlation between β4 and β5 (r = 0.21), between β4 and β6 (r = 0.06), and between β5 and β6 (r = 0.12; p < 0.0001 for all comparisons). This analysis provides evidence that subjects who showed stronger expression of one of the biases also tended to show greater expression of the remaining two biases.
An advantage of large-scale data collection via a smartphone app is that it allows data to be gathered on a range of cognitive tasks across a large cohort of subjects. Recently, we reported learning and choice behavior on another gambling task contained within the same app platform [27,33]. In this simpler gambling task, subjects make binary choices between safe and risky options in three types of trials: gain trials (a certain gain versus a larger gain/zero gain gamble), mixed trials (certain zero gain versus a mixed gain/loss gamble), and loss trials (a certain loss versus a larger loss/zero loss gamble). Notably, subject behavior in this task was best characterized within a Pavlovian approach-avoidance decision model when compared to a range of models that also included a standard Prospect Theory model [27]. This decision model captures the influence on risk-taking behavior of both economic preferences and Pavlovian influences. It describes subjects’ value-independent propensity to approach or avoid gain gambles with a single parameter, βgain, and their value-independent propensity to approach or avoid loss gambles with a second parameter, βloss. Full details of modeling are provided in [27] and Materials and Methods.
For each subject who played both games within the app (n = 21,866 users), we estimated βgain and βloss and computed the difference between these two parameters. We performed a median split on these values to derive two subpopulations of subjects, one exhibiting a larger bias for approach potential rewards over avoid potential losses and one exhibiting a weaker bias. Next, we calculated the average behavior in our task of the subjects within these two subpopulations. We then fit the model described in the previous section to subjects’ aggregate behavior and compared the fits of β4 (rejecting unsampled options), β5 (positive evidence approach), and β6 (sampling the favorite) statistics across the different subpopulations. To estimate our confidence in these statistics, we performed 100 bootstraps using 10,000 samples drawn from each subpopulation.
All three of our information sampling biases were differentially present in the high approach-avoid versus low approach-avoid groups. Positive evidence approach was greater in the high approach-avoid group in both add and multiply trials (Fig 6A). Rejecting the unsampled option was also greater in the high approach-avoid group in the add condition, although this difference was slightly reversed in the multiply condition (Fig 6B). Sampling the favorite showed a subtler pattern of expression was greater in the high approach-avoid group in both add and multiply trials (Fig 6C). All of the different observed biases are linked by the tendency to sample information from locations that will eventually be approached. The present results show that this is also reflected in the expression of these biases in groups exhibiting differential levels of Pavlovian approach influence on their behavior.
An additional advantage of acquiring data via smartphone is that it enables examination of variation in information sampling across a much wider range of subjects than is typically examined in laboratory studies. In an initial exploration of this, we examined variation in a simple measure of information seeking, namely the average number of cards turned relative to the optimal model.
Subjects reliably sampled less information than predicted from the optimal model, but there was substantial variation across the population (Fig 7A). It is important to remember, however, that the model is only “optimal” in the sense of maximizing expected points per gameplay. It does not, for example, include additional factors such as the subjective cost of sampling information. Indeed, we found that adding a “subjective sampling cost” of 5 points per turn to the optimal model shifted the distribution in Fig 7A so that it was now centered around zero (S10 Fig). Nonetheless, variability in the extent to which individual subjects sampled information was highly reproducible across repeated gameplays (Fig 7B and S10 Fig), and we also found it to be stable irrespective of which set or ordering of conditions subjects played (S11 Fig). This suggests that it provides a measure that might be related to performance on other cognitive tasks or demographic information about participants. An example of the latter is our finding that the number of cards gathered was positively related to both the highest level of attained education and age group of our participants (Fig 7C, top panels). Importantly, this measure was decoupled from general performance on the task, which was positively related to educational attainment but negatively related to age (Fig 7C, bottom panels). There was a very slight tendency for subjects within the “high approach-avoid group” to gather more evidence versus subjects in the “low approach-avoid group,” but this difference was negligible relative to the overall variance in information sampling across the population (mean of 0.0066 more cards sampled in high approach-avoid group, 95%CIs [0.0037, 0.0094]).
Information seeking comprises interlinked decisions that include how much to sample, where to sample from, and, finally, which option to choose based upon sampled information. Whilst the complexity of our task allowed these different features to be indexed simultaneously within a single scenario, the task was sufficiently constrained such that it can be treated as a Markov decision process. As such, an optimal model of the task can be derived using dynamic programming [26]. Dynamic programming has rarely been considered as a normative basis for analysis of information search strategies in human information search [7]. Although computationally expensive, a distinct advantage for our purposes is that it straightforwardly derives a common currency for the expected value of sampling in different locations against the value of choosing a particular option.
Subjects rapidly learnt the task, with their performance in terms of points gained becoming relatively stable within ~4 trials; moreover, basic features of subject behavior (e.g., Figs 2B and 3B) matched with the overall pattern of predictions from the normative model. This confirms our previous observations concerning the validity of behavioral data acquired via smartphone [25]. We make the large behavioral dataset freely available for download [31], providing an empirical testing ground for models of human information seeking.
Crucially, three features of subject behavior at different Task Stages showed demonstrable biases in information seeking. Two of these biases, positive evidence approach and sampling the favorite, were elicited as a consequence of our manipulation of which item subjects approached across different conditions. A third bias, rejecting unsampled options, was demonstrated as an effect of rejecting an option on the preference of a subject for choosing that option. All three biases were a consequence of the item that subjects currently sought to approach. Although manifesting as suboptimal biases in our experiment, we contend that these behaviors are present because they are likely to be, and have been, adaptive ecologically [34]. In nature, foraging decisions (such as whether to stay or depart from a patch, or whether to engage with or reject an item of prey) are more common than those made between binary mutually exclusive options [35]. In such contexts, we hypothesize that an adaptive strategy is to engage with the most valuable alternative first and then accept or reject this alternative having acquired more information about its value. It would be intriguing to test whether approach-induced information sampling can produce optimal information sampling in more naturalistic foraging paradigms.
Further evidence supporting the claim that our biases are related to Pavlovian approach comes from their differential expression in two groups who varied in the degree of a Pavlovian approach-avoid parameter derived from a separate decision task. This provides a tentative suggestion of an underlying dopaminergic mechanism for control of Pavlovian approach on information seeking behaviors, given our recent demonstration that Pavlovian approach is boosted in subjects treated with L-DOPA [27]. We also note that polymorphisms in genes controlling dopamine function have recently been linked to individual differences in confirmation bias [36]. Moreover, recordings from midbrain dopaminergic neurons reveal that they signal information in a manner consistent with the animal’s preference for advance information in the same manner that they encode information about reward [17]. Future studies could easily exploit possibilities of data collection via smartphone to test this and related hypotheses via combined collection of genetic and behavioral data across large populations. It might also be possible to design future versions of our task with a larger number of trials/conditions per subject so as to elicit each of the three observed biases within-subject rather than depending upon examining amalgamated data across a population.
It should be noted, however, that this effect was relatively small. This may in part be due to the limited number of trials completed on both tasks, which provides significant challenges for characterization of individual subject behavior. This is particularly the case when multiple conditions/trial types need to be completed to obtain an effect. In the present study, there were only 22 trials per subject, and this is because we explicitly aimed to ensure that the average time to complete each game was less than 5 minutes, as shorter games yield the highest number of gameplays [25].
It is possible that subjects had miscalibrated beliefs about task structure. For example, they may not have realized that there was a card value 1, which is normally replaced by an “ace” in a regular deck of playing cards; or they may have believed that the average card value is 5, rather than 5.5. Such beliefs can straightforwardly be factored into the dynamic programming model, as can misunderstandings about the cost structure of the task, or additional opportunity costs for sampling further evidence. We found that such manipulations did indeed influence the relative preference of the model for guessing or sampling at different card values (S10 Fig). Crucially, however, none of these belief-based manipulations predict any of the three biases observed. “Positive evidence approach” and “sampling the favorite” depend upon comparisons of SMALL and BIG conditions: any normative model predicts that subjects’ information sampling should be identical between these conditions, and that they should simply flip their final choice. Similarly, “rejecting unsampled options” depends upon a comparison of final choice behavior in AA and AB trials in situations in which the subject has received identical information in both trial types.
It would also be possible to explore alternative versions of the current experiment that might examine the generality of our approach-avoidance account of information seeking biases. For instance, it would be intriguing to manipulate the affective valence of “points” such that they became aversive rather than rewarding. In such an experiment, we would predict that the approach-induced biases in information sampling would reverse. It would also be interesting to parametrically manipulate the costs involved in sampling different cards, as this would allow the experimenter to directly quantify the value of sampling information from different locations. It is also important to bear in mind that, even when information sampling is biased, posterior beliefs can remain unbiased if belief updating is performed normatively [37,38]. It would be informative in future experiments to formally dissociate subjects’ apparent biases in information sampling from their biases, if any, in their belief updating.
Our findings are closely linked to other evidence from recent studies that relates the value of stimuli to deployment of attention [23,39,40]. Both these studies and our own suggest that valuable items capture attention and, hence, cause more information to be sampled from the associated location. In contrast with these previous studies, however, we show that the influence of value on information sampling occurs rapidly, can be reshaped depending upon current task goals, and can manifest as several distinct behavioral biases that affect multiple stages of information sampling. Combined, this evidence argues that choice models in which attention and information sampling are determined purely stochastically [6] require revision. Whereas these models convincingly demonstrate an important role for the locus of attention on valuation, the present data imply that the converse is also true. In simple terms, the value subjects ascribe to a location influences how likely they are to sample from it.
Ethical approval for this study was obtained from University College London research ethics committee, application number 4354/001.
Researchers at the Wellcome Trust Centre for Neuroimaging at University College London worked with White Bat Games to develop The Great Brain Experiment [25], available as a free download on iOS and Android systems (see http://thegreatbrainexperiment.com). Ethical approval for this study was obtained from University College London research ethics committee, application number 4354/001. On downloading the app, participants filled out a short demographic questionnaire and provided informed consent before proceeding to the games. Each time a participant started a game, a counter recording the number of plays was incremented. At completion of a game, if internet connectivity was available, a dataset was submitted to the server containing fields defining the game's content and the responses given. The first time a participant completed any game the server assigned that device a unique ID number (UID). All further data submissions from that device, as well as the demographic information from the questionnaire, were linked to the UID. No personal identification of users was possible at any time. Users who responded that their age was less than 18 years during the demographic questionnaire (i.e., minors) were excluded from the study: the app allowed these users to play the games, but no data were submitted to the server (hence the minimum age category on Fig 7C is 18–25 years).
The information-seeking game was available by clicking on “Am I a risk-taker?”, which launched the game. On each gameplay, subjects were randomly assigned to play short blocks (11 trials each, as outlined in Fig 1A and main text) of two different conditions randomly selected from six possibilities (Fig 1B). In two of these, subjects had to select the row that they believe contained the largest sum (“ADD BIGGEST”) or largest product (“MULTIPLY BIGGEST”). In a further two conditions, subjects has to reverse their eventual choice and select the row containing the smallest sum (“ADD SMALLEST”) or product (“MULTIPLY SMALLEST”). The remaining two conditions required participants to select the row with the largest or smallest individual card (“FIND THE BIGGEST” and “FIND THE SMALLEST,” respectively). Full instructions for the task can be seen in S1 Text and S1 Movie.
The economic gambling task was available by clicking on “What makes me happy?" Subjects started the game with 500 points and made 30 choices in each play. In each trial, subjects chose between a certain option and a gamble. Chosen gambles, represented as spinners, were resolved after a brief delay. Subjects were presented with the question, “How happy are you at this moment?” after every two to three trials.
The probabilistic structure of the task means that it is straightforward to derive a normative solution of task performance that maximizes the expected average number of points to be obtained from a given set of moves. This is achieved by applying dynamic programming to the task [30]. At each step, dynamic programming calculates the expected value of every possible action (seeking more information in a particular location, or making a guess). To do so, it takes into account the full probability distribution of currently hidden cards and the possible gain in information that can be obtained from sampling further.
Each combination of presented cards is defined as a state s. The best possible action a that a subject can take in a given state is defined as:
Qs*=maxaQs,a
In a given state, the action value Q of making a particular guess in a particular state s can be calculated as:
Qs,guess=60*p(win)−50*p(lose)+totalcost
where p(win) is the current probability of winning by making that guess, p(lose) is the probability of losing, and totalcost is the incurred costs for sampling information thus far.
By contrast, sampling further information has a fixed probability (0.1) of transitioning into one of 10 possible subsequent states (10 different card values may be revealed). The value of sampling can then be calculated as the best action value in the subsequent state multiplied by the probability of transitioning:
Qs,sample=∑i=1100.1*Qsi*
where si is the state that the subject would transition into if card value i is revealed. To calculate the value of sampling, one works backwards from the terminal state (all four cards revealed, where Qs,guess = 15 (= 60–10–15–20)) to calculate
Qs*
in all previous states.
Full details of the approach-avoidance decision model are given in reference [27]. In brief, subjects’ expected utilities for choosing the safe option (Ucertain) and risky option (Ugamble) were fitted using an established parametric decision model based on Prospect Theory [41]. The probability of choosing to gamble was then modelled by modifying the softmax function:
Pgamble=1−β1+e−μ(Ugamble−Ucertain)+βifβ≥0
Pgamble=1+β1+e−μ(Ugamble−Ucertain)ifβ<0
This causes a value-independent change in the probability of gambling, with mapping choice probabilities to be bounded at (β,1) if β is greater than zero, and (0,β) if β is less than zero. β is fit separately for gain trials and loss trials, yielding two parameters, βgain and βloss.
Raw data, along with MATLAB scripts for reproducing all figures shown in the paper and code for the dynamic programming model, are freely available for download from http://dx.doi.org/10.5061/dryad.nb41c [31].
|
10.1371/journal.ppat.1002718 | Global Analysis of the Evolution and Mechanism of Echinocandin Resistance in Candida glabrata | The evolution of drug resistance has a profound impact on human health. Candida glabrata is a leading human fungal pathogen that can rapidly evolve resistance to echinocandins, which target cell wall biosynthesis and are front-line therapeutics for Candida infections. Here, we provide the first global analysis of mutations accompanying the evolution of fungal drug resistance in a human host utilizing a series of C. glabrata isolates that evolved echinocandin resistance in a patient treated with the echinocandin caspofungin for recurring bloodstream candidemia. Whole genome sequencing identified a mutation in the drug target, FKS2, accompanying a major resistance increase, and 8 additional non-synonymous mutations. The FKS2-T1987C mutation was sufficient for echinocandin resistance, and associated with a fitness cost that was mitigated with further evolution, observed in vitro and in a murine model of systemic candidemia. A CDC6-A511G(K171E) mutation acquired before FKS2-T1987C(S663P), conferred a small resistance increase. Elevated dosage of CDC55, which acquired a C463T(P155S) mutation after FKS2-T1987C(S663P), ameliorated fitness. To discover strategies to abrogate echinocandin resistance, we focused on the molecular chaperone Hsp90 and downstream effector calcineurin. Genetic or pharmacological compromise of Hsp90 or calcineurin function reduced basal tolerance and resistance. Hsp90 and calcineurin were required for caspofungin-dependent FKS2 induction, providing a mechanism governing echinocandin resistance. A mitochondrial respiration-defective petite mutant in the series revealed that the petite phenotype does not confer echinocandin resistance, but renders strains refractory to synergy between echinocandins and Hsp90 or calcineurin inhibitors. The kidneys of mice infected with the petite mutant were sterile, while those infected with the HSP90-repressible strain had reduced fungal burden. We provide the first global view of mutations accompanying the evolution of fungal drug resistance in a human host, implicate the premier compensatory mutation mitigating the cost of echinocandin resistance, and suggest a new mechanism of echinocandin resistance with broad therapeutic potential.
| The evolution of drug resistance poses a severe threat to human health. Candida glabrata is a leading cause of mortality due to fungal infections worldwide. It can rapidly evolve resistance to drugs such as echinocandins, which target the fungal cell wall and are front-line therapeutics for Candida infections. We harness whole genome sequencing to provide a global view of mutations that accumulate in C. glabrata during the evolution of echinocandin resistance in a human host. Nine non-synonymous mutations were identified, including one in the echinocandin target. A mutation in an additional gene conferred a small resistance increase and another was in a gene whose dosage mitigated the fitness cost of resistance. We further discovered that compromising function of the molecular chaperone Hsp90 abrogates drug resistance and reduces kidney fungal burden in a mouse model of infection. Hsp90 and its downstream effector calcineurin are required for induction of the drug target in response to drug. Thus, we reveal the first global portrait of antifungal resistance mutations that evolve in a human host, identify the first compensatory mutation that mitigates the cost of echinocandin resistance, and suggest a new mechanism of echinocandin resistance that can be exploited to treat life-threatening fungal infections.
| The emergence of drug resistance is an evolutionary process with a profound impact on human health. The widespread deployment of antimicrobial agents in medicine and agriculture exerts strong selection for organisms with enhanced capacity to survive and reproduce in the presence of drug, which has led to the rapid emergence of drug resistance in diverse pathogen populations [1]–[4]. The evolution of drug resistance compromises the efficacy of drugs that we depend on critically for a myriad of therapeutic interventions, and has striking economic consequences. The annual cost attributable to the evolution of drug resistance in the United States alone exceeds $33 billion to cover treatment of patients with drug-resistant infections, additional pesticides required to manage resistant pests, and loss of crops to resistant pests [5]. The emergence of drug resistance in fungal pathogens is of particular concern given the increasing incidence of invasive fungal infections, and the limited number of antifungal drugs. Fungi can cause life-threatening infectious disease in immunocompromised hosts, as well as in healthy humans, and the incidence of fungal bloodstream infections has increased by 207% in recent decades [6]–[8]. Fungi are eukaryotes and share close evolutionary relationships with their human hosts, which limits the number of drug targets that can be exploited to selectively kill fungal pathogens yet minimize host toxicity [1], [3]. Even with current treatment options, mortality rates due to invasive fungal infections can reach 50–90% depending on the pathogen and patient population [7], [8], demanding new strategies to prevent the evolution of drug resistance and enhance the efficacy of antifungal drugs.
The evolution of drug resistance is contingent on genetic variability, the ultimate source of which is mutation. One of the most fundamental questions of central importance to predicting and preventing the evolution of drug resistance is which mutations accompany the evolution of drug resistance in the human host. Developments in sequencing technology [9]–[11] now enable this question to be addressed on a genome-wide scale to reveal the identity of mutations that either confer drug resistance in a clinically relevant context or that modify the fitness consequences of resistance mutations. Whole genome sequencing has been applied to bacteria and has revealed principles underpinning the evolution and transmission of drug-resistant pathogens [12], risk factors for the evolution of drug resistance [13], and population dynamics during the evolution of drug resistance in vitro [14]. In fungi, changes in genome-wide gene expression and chromosomal alterations that accompany the evolution of drug resistance have been monitored in experimental populations that evolved resistance in vitro [15], [16], and targeted sequence and expression analysis of specific genes has been implemented to identify mechanisms of resistance that evolve in the human host [1], [17]. However, a global approach to mapping mutations that underpin the evolution of fungal drug resistance has yet to be achieved.
Candida glabrata is a leading fungal pathogens of humans and provides a particularly powerful system for studying the evolution of drug resistance in a human host. Candida species are the fourth most common cause of hospital acquired blood-stream infections and are the most prevalent cause of invasive fungal infection worldwide, with mortality rates approaching 50% [7], [18]. C. glabrata is now second to C. albicans as the most prevalent Candida species isolated from clinical specimens [7], [8], [19]. This is due in part to both intrinsic and rapidly acquired resistance of C. glabrata to the azoles, which are the most widely used class of antifungal drugs and inhibit the biosynthesis of the key sterol in fungal cell membranes, ergosterol [20]. As a consequence, the echinocandins are the front-line therapeutic agent for C. glabrata infections [19]. C. glabrata is closely related to the model yeast Saccharomyces cerevisiae and is placed within the Saccharomyces clade rather than the Candida clade to which the leading cause of candidiasis, C. albicans, belongs [21]–[23]. It is thought that C. glabrata emerged as a human pathogen independently from other Candida species. Notably, gene families associated with pathogenicity in C. albicans including iron acquisition and host cell adhesion and invasion are absent from C. glabrata [24]. C. glabrata is an obligate haploid and mating has never been reported, although two mating types, mating type switching, and other mating and meiotic machinery have been described [25]–[27]. To increase genetic diversity C. glabrata undergoes chromosomal translocations and variation in gene copy number [28], [29], mechanisms that also contribute to C. albicans resistance to azoles [16], [30], [31]. As a haploid, analysis of C. glabrata genome sequence is more facile than in diploids such as C. albicans, where mitotic recombination and gene conversion can inflate the number of polymorphisms that accrue and obscure the signal of those functionally associated with drug resistance or adaptation to the host.
Compared to the most widely used antifungal drugs in clinical use for the treatment of systemic infections, the azoles, mechanisms of resistance remain more limited for echinocandins. The echinocandins are the only novel class of antifungal drugs approved for clinical use in decades and target the biosynthesis of the key fungal cell wall component, 1,3-β-D-glucan [3], [20], [32]. The 1,3-β-D-glucan synthases are encoded by FKS1, FKS2, and FKS3 in S. cerevisiae, C. glabrata, and C. albicans, and require a regulatory subunit encoded by RHO1 for activity [3], [20], [32], [33]. It is thought that the echinocandins bind to and inhibit the Fks protein; however, the exact mechanism of inhibition remains unknown [33]. Although the echinocandins have been in clinical use since only 2001, there have been numerous reports of C. glabrata echinocandin resistance in patients [34]–[37]. Thus far, the only echinocandin resistance mechanism described is mutation in the drug target, Fks, particularly in highly conserved hot spots regions [33], [36]–[39]. Such mutations can reduce echinocandin susceptibility of 1,3-β-D-glucan synthase by 2 to 3 log orders relative to the wild-type enzyme [36]. Additional resistance mechanisms may remain to be described given that Fks mutations have not been identified in some echinocandin-resistant isolates [20], [37], [40], and that isolates with identical Fks mutations exhibit different resistance phenotypes with distinct responses to cellular perturbations [41]. Even with azoles, for which resistance mechanisms have been studied for decades, new resistance mechanisms and modulators of resistance continue to be discovered, expanding the repertoire of strategies employed by fungi to survive drug exposure to include mutation in the drug target, overexpression of multidrug-efflux transporters, and metabolic alterations that minimize drug toxicity [3], [20], [32]. In addition to these canonical resistance mechanisms where mutations in relevant genes confer resistance, there is also an emergent paradigm in which regulators of cellular stress responses are crucial for enabling the evolution and maintenance of drug resistance [3], [20], [32]; while mutations in these regulators have not been identified as a cause of resistance, stress response regulators are key resistance modulators critical for enabling the phenotypic effects of resistance acquired by diverse mechanisms.
Beyond mapping mutations that confer resistance, there is pressing clinical need to elucidate strategies to block the evolution of drug resistance and abrogate resistance once it has evolved. One of the most well studied examples of a protein that governs the emergence and maintenance of fungal drug resistance is the molecular chaperone Hsp90. Hsp90 regulates the folding and function of diverse client proteins, including many signal transducers [42], [43]. In C. albicans, compromise of Hsp90 function reduces basal tolerance and resistance of clinical isolates to both the azoles and the echinocandins [41], [44], [45]. Hsp90 enables crucial responses to drug-induced stress by orchestrating signaling through the protein phosphatase calcineurin and the protein kinase C (PKC) cell wall integrity signaling cascade [41], [46]. Hsp90 stabilizes the catalytic subunit of calcineurin and the terminal mitogen-activated protein kinase (MAPK) in the Pkc1 cell wall integrity pathway. Genetic or pharmacological compromise of Hsp90 function can enhance the efficacy of antifungals against C. albicans in multiple metazoan models of infection [41], [45]. Notably, Hsp90's role in governing cellular responses to azoles in C. albicans is conserved in S. cerevisiae [44], [47]. In contrast, Hsp90 and calcineurin play a key role in crucial cellular responses to echinocandins in C. albicans, but not in S. cerevisiae [41]. Whether Hsp90 influences drug resistance in C. glabrata remains entirely unknown. In C. glabrata, both calcineurin and PKC signaling have been implicated in basal tolerance to echinocandins [48], [49], though the role of Hsp90 remains unknown as does the impact of any of these regulators on bona fide echinocandin resistance.
Here, we provide the first global analysis of mutations accompanying the evolution of fungal drug resistance in a human host. We report on a series of C. glabrata isolates that evolved echinocandin resistance in a patient undergoing treatment with the echinocandin caspofungin for recurring C. glabrata candidemia over a 10-month period. Whole genome sequencing revealed that a mutation occurred in the gene encoding the drug target, FKS2, accompanying a major increase in resistance, as well as 8 other non-synonymous mutations in genes not previously implicated in echinocandin resistance, including CDC6 and CDC55. The FKS2-T1987C(S663P) mutation was sufficient to confer echinocandin resistance in a susceptible laboratory strain; however, the mutant allele also imparted a growth defect in clinically relevant conditions using RPMI medium at 37°C. The fitness cost of resistance was mitigated with further evolution, and this trend was also observed in a murine model of disseminated infection. A CDC6-A511G (K171E) mutation acquired prior to the FKS2-T1987C(S663P) mutation was sufficient to confer a small increase in resistance. Elevated dosage of CDC55, which acquired a C463T(P155S) mutation after FKS2-T1987C(S663P), ameliorated the fitness cost imparted by the FKS2 mutation. To uncover mechanisms that abrogate echinocandin resistance, we turned to Hsp90 and found that genetic or pharmacological compromise of C. glabrata Hsp90 function reduced basal tolerance and resistance of clinical isolates. Compromising calcineurin function pharmacologically or genetically phenocopied compromising Hsp90 function. Caspofungin induced FKS2 expression in a manner that depended upon Hsp90 and calcineurin, providing a molecular mechanism by which Hsp90 and calcineurin regulate echinocandin resistance. Furthermore, one of the clinical isolates in the series is a petite mutant based on morphology and inability to respire; although the petite phenotype was not intrinsically involved in echinocandin resistance, it imparted resistance to the combination of echinocandins and inhibitors of Hsp90 or calcineurin. In a mouse model of candidemia, the petite mutant was rapidly cleared and completely avirulent while infection with the HSP90-repressible strain yielded a reduced fungal burden compared to wild type. Thus, our results provide the first global view of mutations that accompany the evolution of fungal drug resistance in a human host, implicate the first compensatory mutation that mitigates the cost of echinocandin resistance, and suggest a new molecular mechanism regulating echinocandin resistance, with broad therapeutic potential.
While numerous echinocandin-resistant isolates have been recovered from patients [34]–[37], in most cases, there has not been adequate sampling over the course of drug treatment to identify related fungal lineages with which to study the evolution of drug resistance in the human host. The most detailed sampling includes a clinical isolate pre-caspofungin treatment and two isolates taken serially post-caspofungin treatment, however, the mechanism by which resistance evolved was not investigated [50]. In contrast, series of clinical isolates recovered over time from patients undergoing treatment with azoles have proven instrumental for dissecting mechanisms of azole resistance in C. albicans [44], [51]–[55]. The more detailed sampling and analysis of the evolution of resistance to azoles likely reflect the fact that azoles have been used clinically for a longer duration.
We report here on a series of C. glabrata isolates recovered over a 10-month period from a 45-year old female patient with Crohn's disease who suffered from recurrent C. glabrata candidemia and underwent multiple rounds of caspofungin treatment before ultimately succumbing to the infection. A case report on the details of the patient history and therapeutic interventions is provided in the Materials and Methods. Given that multiple blood cultures were negative prior to stopping treatment at any interval and that susceptibility testing was not routinely performed at the time of treatment, caspofungin remained the main therapeutic intervention for candidemia. The 7 blood culture isolates analyzed are labeled alphabetically in the order in which they were recovered, such that isolate A was recovered prior to caspofungin treatment and isolate G was recovered 10 months after recurrent infection and several rounds of caspofungin treatment. The isolates were determined to be related based on molecular typing analysis including pulsed-field gel electrophoresis (PFGE) - karyotype analysis (Figure S1), as well as restriction enzyme PFGE using SfiI (data not shown). Antifungal susceptibility of the 7 isolates in the series was determined for caspofungin, as well as for numerous azoles (fluconazole, ketoconazole, itraconazole, and voriconazole) and for amphotericin B, which binds to ergosterol and disrupts membrane integrity [20], using broth microdilution with RPMI 1640 and following the standard CLSI M27-A3 protocol [56] (Table S1). The major trend observed was an increase in caspofungin resistance over the course of treatment. There were only minimal changes in susceptibility to the other antifungal drugs tested, with the exception of an increase in resistance to azoles that peaked at isolate F and returned to intermediate levels at isolate G. Because the patient was not treated with azoles, the changes in azole susceptibility may be due to mutations that arose in the lineage due to genetic drift rather than selection, or may reflect additional phenotypic consequences of mutations associated with echinocandin resistance, mutations associated with adaptation to the bloodstream, or mutations that were not directly selected for but simply hitch-hiked along with mutations under selection in this predominantly clonal system.
The genomic spectrum of resistance mutations that emerge during the evolution of a fungal pathogen in its human host can be addressed for the first time using next-generation sequencing technology and the series of C. glabrata clinical isolates we describe here. As a haploid, genome analysis of C. glabrata is simplified relative to diploids such as C. albicans, where mitotic recombination and gene conversion can amplify the number of polymorphisms observed between early and late isolates (changes from heterozygous to homozygous states are also classified as nucleotide changes) and thereby hinder functional analysis of mutations conferring resistance. Further, C. glabrata shares a recent common ancestor with the model yeast S. cerevisiae with a large number of orthologs between the two species, facilitating bioinformatic analysis [21], [22]. Given the clinical importance of echinocandin resistance, and that to date resistance has only been attributed to mutations in the echinocandin target [33], [36]–[39], with evidence that additional resistance determinants and modulators may remain to be discovered [20], [37], [40], we sought to identify the mutations that accompany the evolution of echinocandin resistance in the human host on a genome-wide scale.
We performed whole genome sequencing of the C. glabrata isolate recovered prior to caspofungin treatment (isolate A) and the last isolate recovered after multiple rounds of treatment (isolate G) using the Illumina Genome Analyzer II platform. We obtained 5.1 and 3.8 million 76 base pair single-end reads for isolate A and isolate G, respectively, resulting in 22 to 30× genome coverage. Reads were aligned against the reference genome sequence of CBS138 [57]. Single nucleotide variants were identified using a machine learning approach. A total of 45,797 single nucleotide variants were identified between late clinical isolate G and CBS138. Of these single nucleotide variants, 39,146 had sufficient sequencing depth in isolate A to be reliably assigned. Overall, 26 single nucleotide variants were uncovered between isolate A and G, with only 17 of these within open reading frames and only 9 resulting in non-synonymous changes (Table 1). Although the 9 mutations that were not within open reading frames (Table 2) could include mutations that affect regulation of genes important for drug resistance, we focused our analysis on mutations within open reading frames given that silent mutations can more readily be distinguished from those with functional consequences in coding sequences and given their potential impact on gene product function. All 9 non-synonymous changes were verified and then mapped across isolates B to F using Sanger sequencing to determine when each mutation arose in the series (Figure 1). Genes are named based on homology to S. cerevisiae genes [57].
Mutations in the MOH1, GPH1, CDC6, and TCB1/2 genes accompanied the first modest increase in resistance in the series at isolate C (Figure 1). The function of MOH1 in S. cerevisiae is largely unknown except that it is required for survival in stationary phase [58], [59], and it was found to genetically interact with Hsp90 in a genome-wide chemical-genetic screen [60]. Expression of C. albicans MOH1 is induced by alpha pheromone in filament-inducing Spider medium, by weak acid stress via Mnl1, and in biofilm conditions [61]–[63]. GPH1 encodes a glycogen phosphorylase regulated by the high osmolarity glycerol (HOG) mitogen-activated protein (MAP) kinase pathway in S. cerevisiae [64], [65], and is induced upon fluconazole treatment in C. albicans [66]. The CDC6 gene product is involved in DNA replication initiation by forming and maintaining the pre-replicative complex and serving as a loading factor for the Mcm2–7 proteins onto chromatin [67], [68]. TCB1 and TCB2 encode proteins containing both calcium and lipid binding domains and appear to be involved in membrane trafficking in S. cerevisiae [69], [70]. None of these genes or proteins have been previously implicated in echinocandin resistance.
Additional mutations emerged accompanying the subsequent major increase in echinocandin resistance, and in the last isolate of the series. Mutations in FKS2, DOT6, MRPL11, and SUI2 accompanied the largest increase in echinocandin resistance at isolate D (Figure 1). Of these genes, only FKS2 has been implicated in echinocandin resistance. FKS2 encodes the catalytic subunit of 1,3-β-D-glucan synthase, the target of the echinocandins, and the Fks2 S663P mutation identified is in mutational hot spot 1 and has been found in other echinocandin-resistant C. glabrata clinical isolates [36], [37], [71]. Dot6 is a subunit of the RPD3L histone deacetylase complex in S. cerevisiae and is involved in both pseudohyphal morphogenesis as well as silencing at telomeres [72]–[74]. Mrpl11 is a mitochondrial protein and part of the large ribosomal subunit [75], [76]. SUI2 plays a role in translation initiation and encodes the alpha subunit of the translation initiation factor eIF2 in S. cerevisiae [77]. Finally, a non-synonymous mutation occurred in CDC55 in the latest clinical isolate G, despite no further increase in echinocandin resistance (Figure 1). CDC55 encodes the regulatory B subunit of protein phosphatase 2A and has a number of functions, including roles in spindle assembly during meiosis, mitotic exit, pseudohyphal morphogenesis, and chromosome disjunction [78]–[81].
Thus, in addition to a mutation in the known echinocandin target, whole genome sequencing revealed the acquisition of 8 additional non-synonymous mutations in 8 genes not previously implicated in echinocandin resistance or adaptation to host conditions during the evolution of echinocandin resistance in a human host. The genome sequence analysis further confirms clonality of the lineage given the very limited number of single nucleotide variants genome-wide compared to large number observed between the late clinical isolate G and the reference genome CBS138. Further, that each of the mutations identified persisted throughout the lineage once it emerged suggests that there may have been strong selective sweeps in the population such that polymorphisms rapidly reached near fixation.
To determine which of the mutations identified by whole genome sequencing contributes to echinocandin resistance we first turned to the most likely candidate resistance gene, FKS2. The FKS2 mutation that emerged in isolate D and was maintained throughout the rest of the series (T1987C) results in substitution of a serine to proline at amino acid 663 in the target of the echinocandins. This FKS2 T1987C (S663P) mutation has been previously associated with high levels of caspofungin resistance in C. glabrata clinical isolates [36], [37]. In vitro biochemical studies established that this mutant Fks2 enzyme displays reduced sensitivity to inhibition by echinocandins, as indicated by a higher kinetic inhibition parameter, as well as decreased catalytic capacity, and reduced enzyme velocity, compared to the wild-type enzyme; notably the binding affinity of the mutant enzyme for echinocandins remains unchanged [36]. While there is an association of FKS mutations with resistance, and biochemical data support the resistance mechanism, it has only been conclusively demonstrated that such mutations are sufficient to confer echinocandin resistance in S. cerevisiae [82], [83]. To test whether Fks2 S663P is sufficient to confer echinocandin resistance, we introduced the T1987C mutation into the sensitive laboratory strain BG2 using a strategy involving single-stranded DNA containing the mutation and a silent marker, followed by selection of transformants on medium containing caspofungin. No echinocandin-resistant colonies were obtained following control transformations with single-stranded DNA containing the equivalent wild-type sequence or with a water control, while many resistant colonies were obtained with the sequence containing the T1987C mutation. We assessed resistance of four sequence-verified transformants that harboured both the T1987C and silent mutation via minimum inhibitory concentration (MIC) assays. The transformants displayed resistance similar to clinical isolate G (Figure 2A). Thus, Fks2 S663P is sufficient to confer echinocandin resistance in a susceptible laboratory strain of C. glabrata.
Given that specific Fks amino acid substitutions that decrease sensitivity of the 1,3-β-D-glucan synthase enzyme to echinocandins also reduce the enzyme catalytic capacity, the Fks2 S663P mutation may confer a fitness cost in terms of reduced growth or viability in the absence of the drug [36], [39]. To determine if the FKS2 mutations compromises fitness, we monitored growth kinetics of the lab strain BG2 and the progeny harbouring the Fks2 S663P substitution. In clinically relevant conditions, RPMI at 37°C, we found that strains harbouring the Fks2 S663P substitution displayed significantly reduced fitness relative to the parental wild-type laboratory strain BG2 (P<0.01, area under the curve followed by ANOVA, Bonferroni's Multiple Comparison Test, Figure 2B, left panel). Furthermore, we observed a significant reduction in fitness between isolate A and isolate D (P<0.001), which is rescued to some extent in isolate G (P<0.01, Figure 2B, right panel). Thus, the FKS2 mutation that emerged in clinical isolate D is sufficient to confer a high level of caspofungin resistance equivalent to that observed in the late isolate G, however, it is also associated with a cost in terms of reduced fitness in the absence of the drug.
Compensatory mutations that mitigate the cost of drug resistance have been well established in bacterial systems [84]–[87], but remain enigmatic in fungi. Given the fitness cost associated with echinocandin resistance due to mutation in the drug target observed in this study, and in others [88], one would anticipate selection to favour the emergence of compensatory mutations that mitigate the fitness cost of the resistance mutation. The series of C. glabrata isolates studied here provide the ideal opportunity to identify compensatory mutations given that a single non-synonymous mutation occurs between isolate D and G, CDC55 C463T (P155S), associated with an increase in fitness (Figure 1). The CDC55 C463T allele from late clinical isolate G was cloned into a plasmid under the control of its native promoter, as was the CDC55 allele from the early clinical isolate A, and introduced into the BG2 laboratory strain harbouring the FKS2 T1987C mutation. Monitoring growth kinetics of two independent transformants harbouring each of the CDC55 plasmids relative to two transformants with the empty vector control demonstrates that the additional copy of CDC55 ameliorates fitness (P<0.01, ANOVA, Bonferroni Multiple Comparison Test, Figure 3). This suggests that perhaps the CDC55 C463T mutation might be a gain-of-function mutation, such that increased fitness could be achieved either by mutation causing increased Cdc55 activity or by increased dosage of CDC55. These results identify the premier genetic alteration that mitigates the fitness cost of echinocandin resistance.
Despite the fact that the FKS2 T1987C mutation was sufficient to impart the full level of caspofungin resistance of isolate G on an otherwise susceptible laboratory strain (Figure 2A), there is evidence for additional mutations affecting resistance in the evolved lineage. The initial small increase in echinocandin resistance observed at isolate C (Figure 1) occurred prior to the FKS2 mutation and thus other mutations identified at this early transition may be responsible for this increase in resistance. Notably, isolate C showed a fitness defect in the absence of drug (Figure 2B), suggesting that the mutations imparting this small increase in resistance are also costly. Additional mutations that accumulated could mitigate the fitness cost of resistance mutations.
To prioritize mutations for functional analysis, we addressed the prevalence of mutations in any of the 8 genes found to harbour mutations in our whole genome sequence analysis in addition to FKS2 in other C. glabrata echinocandin-resistant mutants. To do so, we obtained 10 additional unrelated C. glabrata clinical isolates harbouring the Fks2 S663P mutation [71], and sequenced across the 8 genomic regions that were mutated in clinical isolate G via Sanger sequencing. We discovered non-synonymous changes in MOH1 in two out of the 10 clinical isolates sequenced and non-synonymous changes in CDC6 in 7 out of 10 of the clinical isolates (Figure 4). Given the prevalence of polymorphisms in CDC6 and MOH1 among these echinocandin-resistant clinical isolates, we prioritized these genes for functional analysis. We cloned plasmids with the gene sequences found in isolate A or in isolate G, and expressed these in a susceptible laboratory strain. While the MOH1-T13C(Y5H) allele did not confer any increase in echinocandin resistance (data not shown), the CDC6-A511G(K171E) allele did confer a small increase in resistance (Figure 5). Thus, we establish a novel mutation in CDC6 that contributes to reduced echinocandin susceptibility in C. glabrata clinical isolates.
Given the importance of echinocandins in the therapeutic front line against C. glabrata infections and the acute problem imposed by the evolution of echinocandin resistance, there is a pressing need to discover strategies to abrogate drug resistance. We focused on Hsp90 due to its critical role in enabling basal tolerance and acquired antifungal drug resistance in pathogenic fungi such as C. albicans and the most lethal mould Aspergillus fumigatus [41], [44], [45], [47]. In the context of the echinocandins, C. albicans Hsp90 orchestrates crucial cellular responses to survive echinocandin exposure by stabilizing the catalytic subunit of the protein phosphatase calcineurin, while in S. cerevisiae Hsp90 and calcineurin do not modulate echinocandin susceptibility under any of the standard conditions tested [41]. To date, no studies have examined the consequences of pharmacological or genetic compromise of Hsp90 function on cellular responses to echinocandins in C. glabrata, and thus whether this pathogen shows resistance circuitry more akin to the pathogen C. albicans or its closer relative S. cerevisiae remains unknown.
We first implemented a pharmacological approach to determine if inhibition of Hsp90 modulates echinocandin resistance of C. glabrata. We monitored growth across a gradient of the widely used echinocandin caspofungin relative to a drug-free growth control in the presence or absence of the two structurally unrelated Hsp90 inhibitors, geldanamycin and radicicol, that bind to the adenosine triphosphate (ATP) binding pocket of Hsp90 and thereby compromise ATP-dependent chaperone function [89], [90]. In the absence of geldanamycin, there was a small increase in caspofungin resistance between isolate B and isolate C and a large increase in resistance between isolate C and isolate D (Figures 1 and 6A). Pharmacological inhibition of Hsp90 with geldanamycin or radicicol decreased tolerance of the early clinical isolates A, B, and C, and reduced resistance of the late clinical isolates D, E, and G (Figure 6A). Synergy between caspofungin and geldanamycin was also observed in RPMI, a medium used for clinical susceptibility testing (Figure S2). Notably, isolate F was refractory to the synergy between caspofungin and Hsp90 inhibitors, as discussed in more detail below.
Next, we validated our pharmacological findings genetically. We engineered a strain of C. glabrata in which the only HSP90 allele was expressed under the control of the MET3 promoter, which is repressed in the presence of methionine and/or cysteine. We monitored the impact of a gradient of methionine concentrations on growth across a gradient of caspofungin concentrations for both isolate G and its derivative in which HSP90 expression is driven by the MET3 promoter. Methionine had no impact on caspofungin resistance of isolate G (Figure 6B). In contrast, methionine reduced echinocandin resistance of the MET3p-HSP90 derivative of isolate G in a dose-dependent manner, providing genetic validation of the importance of Hsp90 for echinocandin resistance (Figure 6B). The highest concentrations of methionine tested (≥1 µg/ml) blocked growth of the MET3p-HSP90 strain, consistent with Hsp90's essentiality in all eukaryotes tested. Thus, targeting Hsp90 provides the first and much-needed strategy to abrogate echinocandin resistance of C. glabrata.
Hsp90 enables resistance to both the azoles and echinocandins, in large part via the protein phosphatase calcineurin in C. albicans [3], [41], [44], [47]. Hsp90 stabilizes the catalytic subunit of calcineurin in both S. cerevisiae and C. albicans [41], [91], thereby enabling calcineurin-dependent responses to drug-induced cellular stress [3], [92]. While calcineurin has been implicated in basal tolerance to echinocandins in C. glabrata [49], whether calcineurin affects bona fide resistance remains unknown.
We used both pharmacological and genetic approaches to determine if calcineurin is a key mediator of Hsp90-dependent echinocandin resistance in C. glabrata. First, we took a pharmacological approach and assessed growth across a gradient of caspofungin concentrations in the presence or absence of two structurally unrelated calcineurin inhibitors, cyclosporin A and FK506. Cyclosporin A and FK506 inhibit calcineurin function in different ways. Cyclosporin A binds to cyclophilin A, a peptidyl-prolyl cis-trans isomerase, and it is this drug-protein complex that inhibits calcineurin function [93]. FK506 binds a structurally unrelated peptidyl-prolyl cis-trans isomerase, FKBP12, and this distinct drug-protein complex also inhibits calcineurin function [93]. We used concentrations of each calcineurin inhibitor that abolished echinocandin resistance in C. albicans but did not inhibit growth on their own [41]. We found that inhibition of calcineurin with cyclosporin A or FK506 reduced caspofungin resistance of the late resistant C. glabrata clinical isolate G (Figure 7A). Next, we validated our pharmacological findings genetically by deletion of the regulatory subunit of calcineurin, encoded by CNB1, which is required for calcineurin function. Loss of calcineurin function reduced resistance of clinical isolate G in three independent mutants (Figure 7A). Thus, calcineurin is a key mediator of Hsp90-dependent echinocandin resistance in C. glabrata.
Since the clinical isolate G harbours multiple mutations relative to its echinocandin-susceptible counterpart, we next specifically assessed the dependence of Fks2-mediated echinocandin resistance on Hsp90 and calcineurin. Pharmacological inhibition of Hsp90 with geldanamycin, or calcineurin with cyclosporin A reduced echinocandin resistance of the laboratory strain harbouring the FKS2 T1987C allele (Figure 7B), confirming that both Hsp90 and calcineurin are required for echinocandin resistance acquired by mutation of the drug target and providing the first circuitry that governs echinocandin resistance in this pathogen.
The specific mechanism by which calcineurin governs echinocandin resistance remains unknown in any system. In S. cerevisiae, FKS2 expression is induced during high temperature growth via calcineurin, and deletion of both FKS1 and FKS2 is synthetically lethal [94]–[97]. Based on these findings, we propose a model in which calcineurin regulates echinocandin resistance in C. glabrata by controlling expression of the resistance determinant FKS2. Given the functional dependence of calcineurin on Hsp90 in other systems [41], [44], [91], one might expect Hsp90 to also influence expression of FKS2. According to our model, inhibition of calcineurin or Hsp90 would compromise expression of the echinocandin-resistant 1,3-β-D-glucan synthase and reduce resistance. Compromising calcineurin or Hsp90 function would also reduce basal tolerance of susceptible strains by reducing FKS2 expression, thereby decreasing the cellular pool of 1,3-β-D-glucan synthase and enhancing susceptibility to a given concentration of echinocandin.
To test this model, we used quantitative RT-PCR to measure transcript levels of FKS2, encoding the catalytic subunit of 1,3-β-D-glucan synthase, in the echinocandin-resistant clinical isolate G and derivatives in which we deleted the regulatory subunit of calcineurin required for its function, encoded by CNB1, or in which we reduced HSP90 levels. Transcript levels of CNB1, HSP90, and FKS2 were monitored after growth in rich medium for one hour with or without caspofungin treatment. We found that caspofungin induced expression of both CNB1 (P<0.001, ANOVA, Bonferroni's Multiple Comparison Test, Figure 8A) and FKS2 (P<0.001, Figure 8A). Importantly, deletion of CNB1 blocked caspofungin-induced upregulation of FKS2 (P<0.001, Figure 8A). To reduce HSP90 levels in the MET3p-HSP90 strain, we included methionine in the medium for both this strain and the wild-type control at a concentration that had minimal impact on growth (0.25 µg/ml). HSP90 levels were reduced in the MET3p-HSP90 strain relative the wild-type control (P<0.001, Figure 8B). Caspofungin-dependent induction of FKS2 was significantly reduced in the MET3p-HSP90 strain (P<0.001, Figure 8B), suggesting that Hsp90 enables FKS2 expression. Taken together, these results establish that FKS2 is induced upon echinocandin exposure, and that full induction of the resistance determinant FKS2 is dependent on both calcineurin and Hsp90. This provides a novel mechanism by which calcineurin and Hsp90 regulate echinocandin resistance in pathogenic fungi.
Respiratory deficient mutants with loss of mitochondrial function, referred to as petite mutants, are associated with azole resistance in S. cerevisiae, C. albicans, and C. glabrata [98]–[102]. To date, there have been no reports of the petite phenotype contributing to echinocandin resistance, although C. glabrata is able to produce petite mutants at high frequency in vitro as well as in vivo [98], [101], [103]. Isolate F in our C. glabrata series was isolated from the patient at the same time point as isolate E, however, it was morphologically distinct and thus archived separately (Text S1). When cultured on rich medium containing dextrose as the carbon source, this isolate had a reduced growth rate relative to the other isolates and produced smaller more transparent colonies (Figure 9A), despite possessing the same karyotype (Figure S1) and complement of polymorphisms in the nuclear genome as isolate E (Figure 1). When cultured on rich medium containing glycerol as the carbon source, which is non-fermentable, isolate F was unable to grow (Figure 9A). These results suggests that isolate F is a petite mutant as it behaves morphologically as a petite and is unable to respire based on failure to grow on glycerol. Notably, isolate F had a distinct echinocandin resistance phenotype from the other resistant isolates in the series in that its resistance was maintained even in the presence of the Hsp90 inhibitors geldanamycin or radicicol (Figure 6), or the calcineurin inhibitor cyclosporin A (Figure 9).
To determine if the petite phenotype is intrinsically involved in echinocandin resistance of C. glabrata clinical isolates, petite mutants were generated from clinical isolate A and their resistance profiles were tested via MIC assays (Figure 9B). If the petite phenotype is sufficient to impart echinocandin resistance, then petite mutants generated from isolate A should acquire resistance, however, they do not (Figure 9B). This suggests that the petite phenotype is not intrinsically involved in C. glabrata echinocandin resistance.
To determine if petite mutants of C. glabrata are intrinsically refractory to the synergy between echinocandins and inhibitors of Hsp90 or calcineurin, petite mutants were generated from late clinical isolate G and their resistance profiles were tested via MIC assays (Figure 9C). The petite mutants generated from isolate G are indeed resistant to the combination of caspofungin and geldanamycin or cyclosporin A, suggesting that they either no longer require calcineurin or Hsp90 for FKS2-mediated resistance, or that they are simply resistant to geldanamycin and cyclosporin A. Notably, petite mutants are known to up-regulate multidrug efflux transporters [98], which may confer resistance to these pharmacological inhibitors by their increased efflux from the cell. To distinguish between these two possibilities, petite mutants were generated from the isolate G derivative in which calcineurin function was genetically compromised due to deletion of CNB1. If petite mutants no longer require calcineurin for echinocandin resistance, then deletion of CNB1 should have no impact on resistance, however, this was not the case (Figure 9D). Petite mutants generated from isolate G cnb1Δ mutants were no longer resistant to caspofungin, suggesting that petite mutants are able to bypass the effects of cyclosporin A, and likely geldanamycin, potentially due to up-regulation of efflux pumps.
We next turned to a murine model of disseminated candidemia to evaluate fitness of the series of clinical isolates and the impact of reduction of Hsp90 levels in vivo. Mice were infected by tail vein injection, and fungal burden in the kidney and spleen was assessed 7 days post infection. Infection with the initial clinical isolate A led to kidney fungal burden greater than that observed with the reference strain CBS138 (P<0.001, Kruskal-Wallis Test, Dunn's Multiple Comparison, Figure 10A). There was a trend towards reduced fitness, as measured by kidney fungal burden, associated with the early acquisition of echinocandin resistance (Figure 10A), consistent with the trend observed in vitro (Figure 2B). Aside from the petite mutant, isolate C showed the greatest reduction in fitness of all the clinical isolates in the series relative to isolate A (P<0.05, Figure 10A). This fitness cost was mitigated with further evolution as shown by an increase in kidney fungal burden of isolate G compared to isolate C (P<0.01, Figure 10A). Kidneys recovered from mice infected with the petite mutant (isolate F) were completely sterile (Figure 10A), consistent with the fitness deficit observed in vitro (Figure 2B) and findings of reduced virulence of some C. glabrata petite mutants [102], [104]. Methionine levels in the mouse are sufficient to repress gene expression from the MET3 promoter in C. albicans leading to avirulence of conditional mutants [105], allowing us to test fungal burden of mice infected with the MET3p-HSP90 strain. One of the two MET3p-HSP90 derivatives of isolate G showed reduced kidney fungal burden relative to isolate G (P<0.05, Figure 10A), consistent with the importance of Hsp90 for growth in vitro and in vivo in a C. albicans murine model of systemic infection [106]. The second MET3p-HSP90 derivative also showed a trend towards reduced kidney fungal burden relative to isolate G (Figure 10A). There were no significant differences observed among any of the isolates tested in the spleen with the exception of isolate F, for which the recovered spleens were sterile (Figure 10B), suggesting that most of the mutations that accumulated in the lineage have negligible impact on fitness in this environment, and that methionine levels are likely not sufficient to achieve substantial reduction of Hsp90 levels or that Hsp90 has a less important role for fungal proliferation in the spleen compared to the kidney. Thus, analysis of kidney fungal burden in a mouse model of disseminated infection reveals a trend towards reduced fitness accompanying early stages of echinocandin resistance that is ameliorated with further evolution, avirulence associated with a petite mutant, and the importance of Hsp90 for fitness in the host.
Our results provide the first global view of mutations that accompany the evolution of fungal drug resistance in a human host, implicate the premier compensatory mutation that ameliorates the fitness cost of echinocandin resistance, and suggest a new molecular mechanism regulating echinocandin resistance, with broad therapeutic potential. We report on C. glabrata bloodstream isolates that evolved increased resistance to the echinocandin caspofungin over a 10-month period during which the patient underwent multiple rounds of caspofungin treatment for recurrent candidemia (Figure S1 and Table S1). This case demonstrates that echinocandin resistance can evolve during treatment, and that an undetected nidus of infection may contribute to fungal persistence and drug resistance despite apparent adequate therapy and documentation of negative blood cultures after treatment, emphasizing the importance of routine antifungal susceptibility testing. Whole genome sequencing of the susceptible isolate recovered prior to drug treatment and the last resistant isolate revealed that 9 non-synonymous mutations accumulated during evolution in the human host (Table 1 and Figure 1). A mutation in FKS2, encoding the drug target, accompanied the largest increase in echinocandin resistance in the lineage; this mutation was sufficient to confer echinocandin resistance in a susceptible C. glabrata strain, but was associated with a fitness cost with reduced growth rate in the absence of drug (Figure 2). The fitness cost of resistance was ameliorated with further evolution, based on observations in vitro and in a murine model of systemic infection (Figures 2 and 10). The 8 additional mutations in genes not previously implicated in echinocandin resistance (Table 1) provide candidates for novel resistance determinants as well as mutations that mitigate the cost of resistance. Consistent with these possibilities, increased dosage of CDC55, which acquired a mutation accompanying the increase in fitness, mitigated the fitness cost of the FKS2 mutation (Figure 3), while a mutation in CDC6 that arose prior to the FKS2 mutation was sufficient to confer a small increase in echinocandin resistance (Figure 5). Further, we establish that Hsp90 governs both basal tolerance to the echinocandins and bona fide resistance of clinical isolates (Figure 6), and that Hsp90 is important for C. glabrata proliferation in the mouse kidney (Figure 10). We found that calcineurin is a key mediator of Hsp90-dependent resistance (Figures 7). Hsp90 and calcineurin regulate echinocandin resistance by controlling expression of the resistance determinant FKS2 (Figure 8), providing a novel mechanism via which Hsp90 and calcineurin govern echinocandin resistance in pathogenic fungi.
The whole genome sequence analysis yields powerful insights into the evolutionary dynamics of adaptation in the host, as well as novel mutations associated with resistance or with ameliorating the fitness cost of resistance. To date, mechanisms of echinocandin resistance remained restricted to mutations in the drug target. In the C. glabrata series, mutations in 4 genes not previously associated with echinocandin resistance (MOH1, GPH1, CDC6, and TCB1/2) accompanied an early and small increase in echinocandin resistance (Figure 1). Mutations in these genes could confer the small increase in resistance, or could create a genetic background in which the FKS2 mutation is less detrimental; in the latter case, the mutation would have to confer a fitness benefit on its own in order to be selected for in advance of the FKS2 mutation, or it could be selectively neutral and fixed by genetic drift. Given that a reduction in fitness accompanied the acquisition of these 4 mutations (Figure 2), it is likely that they contribute to resistance or were fixed by genetic drift. Consistent with the former possibility, polymorphisms in one of the genes that acquired a mutation associated with the small increase in resistance, CDC6, were common in an unrelated set of echinocandin-resistant C. glabrata isolates with Fks2 S663P (Figure 4), and the CDC6-A511G(K171E) mutation identified by whole genome sequencing was sufficient to confer a small increase in resistance (Figure 5). Mutations in 3 additional genes not previously implicated in echinocandin resistance (DOT6, MRPL11, and SUI2) coincided with the FKS2 mutation, and a last mutation in CDC55 arose in the last isolate, without any associated change in resistance (Figure 1). Given that the FKS2 mutation is sufficient for the full resistance phenotype of the late clinical isolate (Figure 7), it is likely that these other mutations are unrelated to echinocandin resistance or that they mitigate the fitness cost of the FKS2 mutation. It is notable that multiple mutations accumulated at the two major transitions in resistance (Figure 1). This is consistent with strong selection favouring the rapid accumulation of mutations in the lineage. That each of the mutations identified persisted throughout the lineage is consistent with the occurrence of selective sweeps, such that each mutation rose to near fixation. Selective sweeps in response to drug selection are also observed in experimental populations of C. albicans during the evolution of azole resistance in vitro [107].
The fate of drug-resistant mutants in nature depends on their fitness relative to drug-susceptible counterparts. While resistance mutations are expected to confer a fitness benefit in the presence of the drug, they may also confer a cost in terms of reduced fitness in the absence of the drug. This model is consistent with the impact of the FKS2 T1987C mutation observed in the C. glabrata lineage, and reported for target-mediated echinocandin resistance in C. albicans [88]. This mutation confers a major increase in growth in the presence of echinocandin (Figure 2A), but also confers reduced growth in the absence of the drug (Figure 2B). The deleterious impact on fitness is likely due to the reduced catalytic capacity commonly observed among 1,3-β-D-glucan synthase enzymes that acquire amino acid substitutions that reduce their sensitivity to echinocandins [36]. C. glabrata may upregulate FKS2 expression to compensate for its decreased catalytic capacity [36], or may acquire additional mutations that mitigate the cost of the resistance mutation. The fitness effects of antibiotic resistance mutations have been studied extensively in bacteria, where most resistance mechanisms are associated with a fitness cost that manifests in reduced growth rate [84]. In the vast majority of cases, the fitness cost is mitigated by the acquisition of compensatory mutations [84]–[87]. Consistent with these patterns, any cost of resistance in experimental populations of C. albicans that evolved azole resistance in vitro was mitigated was further evolution [108], as many changes in gene expression observed in the less fit, resistant population were restored to the ancestral state [15]. In the C. glabrata lineage studied here, a mutation in CDC55 accompanied an increase in fitness of isolate G relative to isolate D (Figures 1 and 2). Increased dosage of CDC55 ameliorated fitness of an independent FKS2 T1987C mutant, independent of whether it was the C463T(P155S) allele (Figure 3). This suggests that the C463T(P155S) mutation may confer increased Cdc55 function and that fitness can be ameliorated by either elevated activity or dosage of Cdc55. This C463T(P155S) mutation was not identified in other C. glabrata echinocandin-resistant FKS2 T1987C mutants (Figure 4), suggesting that distinct compensatory mutations might be favoured in vivo or that they may harbor duplications of CDC55. The mechanisms by which alterations in CDC55 ameliorates fitness of the FKS2 mutant and the scope of beneficial effects in other backgrounds remains to be determined.
Morphological variants that emerge in an evolutionary lineage can reveal important features of mechanisms of drug resistance or drug synergy. Isolate F is a petite mutant based on morphology and inability to grow on a non-fermentable carbon source (Figure 9A). Such respiratory deficient mutants with loss of mitochondrial function are associated with azole resistance in S. cerevisiae, C. albicans, and C. glabrata [98]–[102]. The azole resistance of petites is attributed to increased expression of multidrug transporters of the ATP binding cassette family [98], [99], [101]. The petite phenotype has not been linked to echinocandin resistance to date, consistent with the limited evidence that multidrug transporters are involved in resistance to this drug class [109]–[112]. Indeed, induction of petite mutants in the early C. glabrata clinical isolate A does not confer echinocandin resistance, confirming that petite mutants are not intrinsically resistant to echinocandins (Figure 9B). In S. cerevisiae, several mitochondrial proteins have been identified as required for echinocandin tolerance [113], consistent with the slight reduction in tolerance we observe in C. glabrata petite mutant derivatives of isolate A (Figure 9). A striking feature of the isolate F is that its echinocandin resistance phenotype is recalcitrant to the impact of the Hsp90 inhibitor geldanamycin or calcineurin inhibitor cyclosporin A, unlike that of all other isolates in the series (Figure 9C). Induction of petite mutants in late clinical isolate G confirms that the petite phenotype is intrinsically recalcitrant to the impact of geldanamycin or cyclosporin A (Figure 9C). Genetic compromise of calcineurin function in petite mutants abrogates echinocandin resistance, suggesting that petites are simply resistant to cyclosporin A (Figure 9D), and likely geldanamycin; this may be attributable to overexpression of multidrug transporters in petites that remove these inhibitors from the cell. Whether the original isolate F petite mutant arose during evolution in the human host or during sampling remains unknown, although our finding that kidneys and spleens from mice infected with the petite mutant were completely sterile (Figure 10), suggests that petite mutant arose shortly before or during its isolation. In contrast, one C. glabrata petite mutant was reported to have enhanced virulence relative to an isolate recovered from the same patient at an earlier time point, however, an isogenic control was lacking [102]. Consistent with our findings, most C. glabrata petite mutants have been reported to have attenuated virulence relative to isogenic controls [104].
Our findings of reduced kidney fungal burden in mice infected with the late clinical isolate G with conditional expression of HSP90 driven by the MET3 promoter support a role for Hsp90 in proliferation in the host (Figure 10). That there was only a modest reduction in fungal burden is surprising in light of Hsp90's essentiality in vitro, as observed by methionine-mediated transcriptional repression (Figure 6 and 8). It is likely that methionine levels in the mouse were not sufficient to repress the C. glabrata MET3 promoter and deplete Hsp90, and that methionine levels were even lower in the spleen, leading to no significant differences in fungal burden (Figure 10). Notably, methionine levels in the mouse were sufficient to cause avirulence of an ino1Δ/ino1Δ itr1Δ/MET3p-ITR1 conditional mutant of C. albicans [105], but only partial attenuation of virulence of a C. albicans conditional mutant of the essential gene FBR1 (fbr1Δ/MET3p-FBR1) [114]. The impact on virulence using the MET3 system to deplete essential genes may depend on the level of depletion required to observe phenotypic effects. Using a tetracycline-repressible promoter system and delivery of tetracycline to the mice in a systemic model of infection, Hsp90 has been shown to be required for proliferation of C. albicans [106], suggesting that this promoter system may be more suitable for in vivo studies given that doses of tetracycline can be titrated.
Our results further establish Hsp90 and calcineurin as the first regulators of bona fide echinocandin resistance in C. glabrata, and reveal that resistance circuitry has been rewired over evolutionary time. The molecular chaperone Hsp90 and its client protein calcineurin govern basal tolerance and resistance to both the azoles and the echinocandins in C. albicans [41], [44], [47]. Pharmacological inhibition of Hsp90 can enhance the efficacy of azoles against C. glabrata [115], and calcineurin plays an important role in both azole and echinocandin tolerance [49]. The roles of Hsp90 and calcineurin in azole resistance are conserved in S. cerevisiae [44], [47]. However, compromise of Hsp90 or calcineurin function does not alter echinocandin susceptibility in S. cerevisiae under any of the standard conditions tested where echinocandin susceptibility of C. albicans is affected [41]. Here, we find that Hsp90 and calcineurin are required for basal tolerance to echinocandins in C. glabrata as well as for resistance that evolves in a human host (Figures 4 and 5), suggesting that despite the closer evolutionary relationship of C. glabrata to S. cerevisiae, the C. glabrata cellular circuitry governing resistance to drugs that target the cell wall shares more similarity to that of its more distant pathogenic relative, C. albicans. While conditions may exist in which Hsp90 and calcineurin influence echinocandin susceptibility in S. cerevisiae, it is clear that the C. glabrata phenotypic response more closely resembles that of C. albicans than S. cerevisiae. Notably, signaling pathways governing cell wall integrity have been rewired between C. albicans and S. cerevisiae [116]. The cell wall is essential for fungal viability and is an elaborate structure, components of which are recognized by vigilant immune cells in the human host [117]. As commensals and opportunistic pathogens, C. glabrata and C. albicans are likely to harbour circuitry governing cell wall architecture that is subject to strong selection in response to host immune system challenge.
This work establishes that targeting Hsp90 or calcineurin has broad therapeutic potential for infections caused by one of the leading fungal pathogens of humans, and complements the expanding repertoire of therapeutic applications for inhibitors of Hsp90 and calcineurin in the treatment of infectious disease. Inhibition of Hsp90 or calcineurin transforms echinocandins from ineffective to highly efficacious against echinocandin-resistant C. glabrata (Figures 6 and 7). Notably, a human recombinant antibody against Hsp90 also has synergistic activity with echinocandins against C. glabrata in a mouse model [118], though the mechanism by which this antibody works remains entirely unknown as it is unlikely to enter intact fungal cells to influence function of the cytosolic Hsp90 chaperone. Genetic compromise of Hsp90 function enhances the efficacy of azoles and echinocandins in a mouse model of systemic C. albicans infection [41], [45]. Genetic or pharmacological compromise of Hsp90 also transforms fluconazole from ineffective to highly efficacious against C. albicans biofilms in a mammalian model of biofilm infection [119]. Beyond Candida species, inhibition of Hsp90 also enhances antifungal efficacy against the most lethal mould, A. fumigatus, in biofilms and in a metazoan model of infection [45], [119]. Consistent with the functional relationship between Hsp90 and calcineurin, calcineurin inhibitors also have therapeutic potential and are synergistic with azoles against C. albicans endocarditis, keratitis, and biofilms in mammalian models [120]–[122]. Beyond their utility in the treatment of fungal infections, Hsp90 and calcineurin are promising targets for treating infections caused by protozoan parasites including Plasmodium falciparum, Trypanosoma evansi, and Leishmania major [123]–[126]. Supporting their clinical relevance, Hsp90 inhibitors have advanced in clinical trials for the treatment of cancer and other diseases [9], [127], [128], and calcineurin inhibitors are widely used in the clinic as immunosuppressants [92]. Given the potential for toxicity upon inhibition of key cellular regulators in the host during infection [45], the challenge for further development of Hsp90 and calcineurin as therapeutic targets for infectious disease lies in developing pathogen-selective inhibitors or drugs that target pathogen-specific components of the cellular circuitry governing drug resistance and virulence.
Dr. Susan M. Poutanen discussed this study and specifically highlighted the inclusion of the case history and the case isolates in this study with Nushrat Sultana, Research Ethics Board Coordinator for the Research Ethics Board at Mount Sinai Hospital in Toronto, Canada, the hospital in which the case patient had been hospitalized. Nushrat Sultana confirmed that completing and publishing a case study that involves only a single case does not require review by the Research Ethics Board. Written confirmation has been provided by Dr. Ronald Heslegrave, Chair, Research Ethics Board, Mount Sinai Hospital. Oral and written consent was provided by the case patient's mother, as the patient is deceased. Animals studies conducted in the Division of Laboratory Animal Resources (DLAR) facilities at Duke University Medical Center (DUMC) were handled with good practice as defined by the United States Animal Welfare Act and in full compliance with the guidelines of the DUMC Institutional Animal Care and Use Committee (IACUC). The murine systemic infection model was reviewed and approved by the DUMC IACUC under protocol number A238-09-08.
A 43 year-old female had four episodes of Candida glabrata candidemia between April 2004 and October 2005. Her past medical history was significant for severe fistulizing Crohn's disease diagnosed at the age of 9 years. She had been on total parenteral nutrition (TPN) since 2003 for short gut syndrome due to numerous small bowel resections over the 30-year course of her disease.
Archives of C. glabrata strains were maintained at −80°C in 25% glycerol. Strains were grown in either YPD (1% yeast extract, 2% bactopeptone, 2% gucose), YPG (1% yeast extract, 2% bactopeptone, 2% glycerol), synthetic defined medium (0.67% yeast nitrogen base, 2% glucose) supplemented with required amino acids, or RPMI medium 1640 (Gibco, 3.5% MOPS, 2% glucose, pH 7.0). 2% agar was added for solid media. Strains used in this study are listed in Table S2. Strain construction is described in Text S1.
Recombinant DNA procedures were performed according to standard protocols. Plasmids used in this study are listed in Table S3. Plasmid construction is described in the Text S1. Plasmids were sequenced to verify the absence of any nonsynonymous mutations. Primers used in this study are listed in Table S4.
PFGE-karyotyping and restriction enzyme-PFGE using SfiI were performed following previously reported methods [129].
Genomic DNA was extracted from clinical isolate A and G, and sequencing libraries were prepared using the Illumina genomic DNA library preparation kit according to the manufacturers recommendations (Illumina, CA) with several modifications. In brief, DNA was sheared by sonication to an average fragment length of 200 base pairs. Illumina adapters were blunt-end ligated and libraries were amplified by PCR and purified using Ampure (Agencourt) beads at a DNA:bead ratio of 1:0.9. Each sample was sequenced together in a single lane on an Illumina Genome Analyzer II platform, yielding 5.1 and 3.8 million 76 base pair single-end reads for isolate A and isolate G, respectively, resulting in 22 to 30× genome coverage. Reads were aligned using SOAP2 (PMID: 19497933) against the reference genome sequence of CBS138 [57]. Single nucleotide variants were identified using a machine learning approach as described previously [130]. All non-synonymous mutations identified were validated independently using Sanger sequencing. Raw data is available for download from the Short Read Archive under accession SRA047280.2.
The 9 non-synonymous mutations identified by whole genome sequencing were mapped across clinical isolates B, C, D, E, and F using Sanger sequencing. CgFKS2 was amplified using oLC1344/1345, CgDOT6 with oLC1559/1560, CgMOH1 with oLC1561/1562, CgGPH1 with oLC1563/1564, CgMRPL11 with oLC1565/1566, CgCDC6 with oLC1567/1568, CgCDC55 with oLC1569/1570, CgSUI2 with oLC1571/1572, and CgTCB1/2 with oLC1573/1574. CgFKS2 was sequenced with oLC1344, CgDOT6 with oLC1559, CgMOH1 with oLC1561, CgGPH1 with oLC1563, CgMRPL11 with oLC1565, CgCDC6 with oLC1567, CgCDC55 with oLC1569, CgSUI2 with oLC1571, and CgTCB1/2 with oLC1573.
Antifungal susceptibility was determined in flat bottom, 96-well microtiter plates (Sarstedt) using a modified broth microdilution protocol, as described [44]. Minimum inhibitor concentration (MIC) tests were set up in a total volume of 0.2 ml/well with 2-fold dilutions of caspofungin (CF, generously provided by Rochelle Bagatell). Echinocandin gradients were from 16 µg/ml down to 0 with the following concentration steps in µg/ml: 16, 8, 4, 2, 1, 0.5, 0.25, 0.125, 0.0625, 0.03125, 0.015625, and 0. Cell densities of overnight cultures were determined and dilutions were prepared such that ∼103 cells were inoculated into each well. Geldanamycin (GdA, A.G. Scientific, Inc.) and radicicol (RAD, A.G. Scientific, Inc.) were used to inhibit Hsp90 at the indicated concentrations, and cyclosporin A (CsA, CalBiochem) and FK506 (A.G. Scientific, Inc.) were used to inhibit calcineurin at the indicated concentrations. Dimethyl sulfoxide (DMSO, Sigma Aldrich Co.) was the vehicle for GdA, RAD, CsA, and FK506. Sterile water was the vehicle for CF. Plates were incubated in the dark at 30°C for the time period indicated, at which point plates were sealed and re-suspended by agitation. Absorbance was determined at 600 nm using a spectrophotometer (Molecular Devices) and was corrected for background from the corresponding medium. Each strain was tested in duplicate on at least two occasions. MIC data was quantitatively displayed with colour using the program Java TreeView 1.1.3 (http://jtreeview.sourceforge.net).
Clinical antifungal MICs were determined using broth microdilution with RPMI 1640 broth for amphotericin, fluconazole, ketoconazole, itraconazole, voriconazole, and caspofungin following Clinical and Laboratory Standards Institute document M27-A3 [56]. Visual MIC endpoints were read after 24 hours of incubation at 35°C for caspofungin and after 48 hours of incubation for all other drugs. Complete inhibition was used to determine amphotericin endpoints; 50% inhibition (compared to growth control) was used for caspofungin and 80% inhibition was used for the other drugs.
To measure gene expression changes in response to caspofungin treatment in C. glabrata, cells were grown overnight in YPD at 30°C. Cells were diluted to OD600 of 0.2 in SD and grown for 2 hours in duplicate for each strain at 25°C. After 2 hours of growth 120 ng/ml CF was added to one of the two duplicate cultures and left to grow for one additional hour at 25°C. Cells were centrifuged and pellets were frozen at −80°C immediately. RNA was isolated using the QIAGEN RNeasy kit and RNAse-free DNase (QIAGEN), and cDNA synthesis was performed using the AffinityScript cDNA synthesis kit (Stratagene). PCR was performed using SYBR Green JumpStart Taq ReadyMix (Sigma-Aldrich Co) with the following cycling conditions: 94°C for 2 minutes, 94°C for 15 seconds, 60°C for 1 minute, 72°C for 1 minute, for 40 cycles. All reactions were performed in triplicate, using primers for the following genes: CgACT1 (oLC1500/1501), CgCNB1 (oLC1502/1503), and CgFKS2 (oLC1498/1499). Data were analyzed using iQ5 Optical System Software Version 2.0 (Bio-Rad Laboratories, Inc). Statistical significance was evaluated using GraphPad Prism 5.0.
To monitor gene expression changes in response to reduction of HSP90 levels, strains CgLC751 and CgLC2121 were grown overnight at 30°C in synthetic defined medium supplemented with 0.25 µg/ml methionine. Stationary phase cultures were diluted to an OD600 of 0.2 and grown for 2 hours at 30°C in synthetic defined medium with methionine. Following incubation, cultures were split and 120 ng/ml caspofungin added to one set. Cells were harvested after one hour and RNA was isolated using the QIAGEN RNeasy kit and cDNA synthesis was performed using the AffinityScript cDNA synthesis kit (Stratagene). PCR was carried out using the SYBR Green Fast Mix (Applied Biosystems) with the following cycle conditions: 95°C for 20 seconds, and 95°C for 3 seconds, 60°C for 30 seconds, for 40 cycles. All reactions were done in triplicate using the following primer pairs: CgACT1 (oLC1500/1501), CgHSP90 (oLC2155/2156), CgFKS2 (oLC1498/1499). Data were analyzed in the StepOne analysis software (Applied Biosystems).
Growth kinetics were measured in C. glabrata strains by inoculating cells from an overnight culture grown in YPD at 30°C to an OD600 of 0.0625 in 100 µl of RPMI with 2% glucose in flat bottom, 96-well microtiter plates (Sarstedt). Cells were grown in a Tecan GENios microplate reader (Tecan Systems Inc., San Jose, USA) at 37°C with orbital shaking. Optical density measurements (OD600) were taken every 15 minutes for 48 hours. Statistical significance was evaluated using GraphPad Prism 4.0. For assays involving plasmids, selection was maintained with 150 µg/ml nourseothricin (Werner BioAgents).
C. glabrata strains were inoculated from solid YPD medium to liquid YPD medium containing 10 µg/ml ethidium bromide (EtBr). The culture was grown overnight, shaking at 30°C in the dark. Approximately 100 cells were plated on YPD agar to isolate single colonies. After 2 days of incubation at 30°C, single colonies were tested for growth on YPD agar and YP-glycerol agar. Colonies able to grow on glucose as the sole carbon source but not on glycerol as the sole carbon source were selected.
Four- to five-week-old male CD1 mice from The Jackson Laboratory (n = 10 for each group) were utilized in this study. C. glabrata strains were grown in 10 ml liquid methionine-free minimal medium (6.7 g yeast nitrogen base without amino acids and 20 g glucose in 1 liter) overnight at 30°C. Cultures were washed twice with 10 ml of phosphate buffered saline (PBS), and the cells were then resuspended in 2 ml of PBS. Cells were counted with a hemocytometer and resuspended in an appropriate amount of PBS to obtain an infection inocula concentration of 1×108 cells/ml. Two hundred microliters (2×107 cells) were used to infect mice by lateral tail vein injection. Appropriate dilutions of the cells were plated onto solid methionine-free minimal medium and incubated at room temperature for 48–96 hours to confirm cell viability. C. glabrata infected mice were sacrificed and dissected on day 7 post-infection. The kidney and spleen tissues were removed, weighed, transferred to a 15 ml Falcon tube filled with 5 ml PBS, and homogenized for 10 seconds at 17,500 rpm (Power Gen 500, Fisher Scientific). Tissue homogenates were serially diluted, and 100 µl was plated onto YPD solid medium (except CgLC2121 and CgLC2122 strains which were plated onto methionine-free minimal medium containing 100 µg/ml chloramphenicol). The plates were incubated at room temperature for 48–96 hours to determine CFUs per gram of organ. We confirmed that organ-recovered CgLC2121 and CgLC2122 cells only grew on methionine-free minimal medium but not on medium containing methionine. All experimental procedures were carried out according to NIH guidelines and Duke IACUC protocols for the ethical treatment of animals.
C. glabrata: ACT1 (2890423); CDC55 (2890752); CDC6 (2890231); CNB1 (2890566); DOT6 (2886442); FKS1 (2888318); FKS2 (2890040); FKS3 (2891236); GPH1 (2887861); HSP90 (2891108); MET3 (2886500); MOH1 (2887590); MRPL11 (2889720); RHO1 (2888920); SUI2 (2886561); TCB1/2 (2889714).
S. cerevisiae: CDC55 (852685); DOT6 (856822); FKS1 (851055); FKS2 (852920); FKS3 (855353); GPH1 (856289); MOH1 (852231); MRPL11 (851325); RHO1 (856294); SUI2 (853463); TCB1 (854253); TCB2 (855637).
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.